#  @redtachyon Ariel
Ariel posts on X about ai, if you, this is, llm the most. They currently have [-----] followers and [---] posts still getting attention that total [-----] engagements in the last [--] hours.
### Engagements: [-----] [#](/creator/twitter::410212936/interactions)
- [--] Week [---------] +6,952%
- [--] Month [---------] -54%
- [--] Months [----------] +5,906%
- [--] Year [----------] +40,424%
### Mentions: [--] [#](/creator/twitter::410212936/posts_active)
- [--] Week [--] +229%
- [--] Month [--] +65%
- [--] Months [---] +630%
- [--] Year [---] +1,055%
### Followers: [-----] [#](/creator/twitter::410212936/followers)
- [--] Week [-----] +0.14%
- [--] Month [-----] +1.80%
- [--] Months [-----] +534%
- [--] Year [-----] +3,961%
### CreatorRank: [---------] [#](/creator/twitter::410212936/influencer_rank)
### Social Influence
**Social category influence**
[technology brands](/list/technology-brands) 13.08% [countries](/list/countries) 5.61% [finance](/list/finance) 3.27% [social networks](/list/social-networks) 2.34% [stocks](/list/stocks) 1.87% [gaming](/list/gaming) 1.4% [cryptocurrencies](/list/cryptocurrencies) 0.47% [celebrities](/list/celebrities) 0.47% [fashion brands](/list/fashion-brands) 0.47%
**Social topic influence**
[ai](/topic/ai) 18.69%, [if you](/topic/if-you) 10.28%, [this is](/topic/this-is) 6.54%, [llm](/topic/llm) 6.07%, [meta](/topic/meta) 5.61%, [paper](/topic/paper) 5.14%, [check](/topic/check) 5.14%, [js](/topic/js) 3.74%, [how to](/topic/how-to) 3.74%, [bit](/topic/bit) 3.27%
**Top accounts mentioned or mentioned by**
[@hallerite](/creator/undefined) [@meekaale](/creator/undefined) [@primeintellect](/creator/undefined) [@teknium](/creator/undefined) [@victortaelin](/creator/undefined) [@arceeai](/creator/undefined) [@swyx](/creator/undefined) [@thdxr](/creator/undefined) [@weibollm](/creator/undefined) [@jsuarezs](/creator/undefined) [@aiatmeta](/creator/undefined) [@leothecurious](/creator/undefined) [@stochasticchasm](/creator/undefined) [@fluorinespark](/creator/undefined) [@jefbak](/creator/undefined) [@fangyi11101](/creator/undefined) [@extropic](/creator/undefined) [@aidanmclau](/creator/undefined) [@alxfazio](/creator/undefined) [@maxtakeoff](/creator/undefined)
**Top assets mentioned**
[Alphabet Inc Class A (GOOGL)](/topic/$googl)
### Top Social Posts
Top posts by engagements in the last [--] hours
"Aight let's unclickbait the fp16 paper. t;dr cool paper a little bit overstated in comms very overstated by poasters. The thing that gave me a pause is that on the surface it seems to claim that bf16 is horrible borderline unusable. But that's not really the case (nor is it the claim). Yes the fsdp-vllm mismatch is real and yes it can be mitigated with fp16. This is as true as it is irrelevant because who cares about the mismatch if the algorithm empirically works The widely circulated figures show bf16 consistently collapsing while fp16 runs thrive. I have no reason to doubt this data but"
[X Link](https://x.com/anyuser/status/1984605827034972269) 2025-11-01T12:57Z [----] followers, 57.4K engagements
"Aight let's talk about frameworks libraries RL and why I probably don't like your favorite RL codebase. Yes including that one. The unusual thing about RL is that the algorithm is the easy part. GRPO is a single-line equation on some logprobs. If you have the data computing the loss is trivial and then presumably you're using it with a backprop library of your choice. But that's the problem -- getting the data. It's a pain in the ass. In regular RL you have to do rollouts perhaps truncate some episodes and handle the ends accordingly. If you don't want to be a snail you'll want to vectorize"
[X Link](https://x.com/anyuser/status/1986177621357691263) 2025-11-05T21:03Z [----] followers, 82.2K engagements
"hiring an ai engineer about to send an offer ask candidate if he's pytorch or web dev he doesn't understand explain the differences between pytorch and web dev he still doesn't get it pull out illustrated diagram explaining what is pytorch and what is web dev he laughs and says "i'm a good engineer sir" hire him import requests"
[X Link](https://x.com/anyuser/status/1988665986787021280) 2025-11-12T17:51Z [----] followers, 470.7K engagements
"Making a wild bet and ignoring clawdbot I meant moltbot I mean openclaw I mean clawbook We'll see if it pays off"
[X Link](https://x.com/redtachyon/status/2017372607113208060) 2026-01-30T23:01Z [----] followers, [----] engagements
"I got tired of memorizing/figuring out shell commands so I got my clanker to build a clanker. uv tool install clanker clanker config clanker do "List all files recursively sort by size ascending" enter Plugs into any OpenAI-compatible endpoint. GLM subscription is great"
[X Link](https://x.com/redtachyon/status/2017697342049165715) 2026-01-31T20:31Z [----] followers, [----] engagements
"No they can't Can LLMs reliably predict program termination We evaluate frontier LLMs in the International Competition on Software Verification (SV-COMP) [----] directly competing with state-of-the-art verification systems. @AIatMeta @HebrewU @Bloomberg @imperialcollege @ucl @jordiae https://t.co/EcD9iCaL9Y Can LLMs reliably predict program termination We evaluate frontier LLMs in the International Competition on Software Verification (SV-COMP) [----] directly competing with state-of-the-art verification systems. @AIatMeta @HebrewU @Bloomberg @imperialcollege @ucl @jordiae https://t.co/EcD9iCaL9Y"
[X Link](https://x.com/redtachyon/status/2018019759573303416) 2026-02-01T17:52Z [----] followers, 19.6K engagements
"Gastown and moltwhatever are giving off strong langchain energy"
[X Link](https://x.com/redtachyon/status/2018029658755801369) 2026-02-01T18:32Z [----] followers, [----] engagements
"Repo link: gyllm and nanorl on PyPI built with uv in mind Mostly tested on DGX Spark cc @leothecurious @stochasticchasm you were interested in the DGX Spark setup - this project should just build with uv with vllm at least https://github.com/redtachyon/gyllm https://github.com/redtachyon/gyllm"
[X Link](https://x.com/redtachyon/status/2018038882919493930) 2026-02-01T19:08Z [----] followers, [---] engagements
"Can we train LLMs with RL using the same next token prediction loss as pre-training (yes) We conduct a study on (log)prob rewards and show they give a simple way to bridge verifiable and non-verifiable settings with a single reward broadly applicable for fine-tuning LLMs"
[X Link](https://x.com/redtachyon/status/2019426794089378213) 2026-02-05T15:04Z [----] followers, [----] engagements
"@VictorTaelin I got the G2's recently hardware is solid but still needs some better software support"
[X Link](https://x.com/redtachyon/status/2021249256837853593) 2026-02-10T15:45Z [----] followers, [---] engagements
"You can't control it like a terminal yet unless you write your own firmware. There's MentraOS which will support something like that but it's not yet released for the G2. Might as well order now OS should be ready before it arrives. Also - it's not Brazil they take forever to ship everywhere demand is absurd https://twitter.com/i/web/status/2021263720400216202 https://twitter.com/i/web/status/2021263720400216202"
[X Link](https://x.com/redtachyon/status/2021263720400216202) 2026-02-10T16:43Z [----] followers, [---] engagements
"How do I short whatever startup he's grifting Ask ChatGPT a complex question and you'll get a confident well-reasoned answer. Then type "Are you sure" Watch it completely reverse its position. Ask again. It flips back. By the third round it usually acknowledges you're testing it which is somehow worse. It knows what's https://t.co/FRCtDoJ5rI Ask ChatGPT a complex question and you'll get a confident well-reasoned answer. Then type "Are you sure" Watch it completely reverse its position. Ask again. It flips back. By the third round it usually acknowledges you're testing it which is somehow"
[X Link](https://x.com/redtachyon/status/2021653370558459976) 2026-02-11T18:31Z [----] followers, [----] engagements
"Formal proof of the Riemann hypothesis verifiability is magic - not sure what verifiable problems cant be solved with AI what are the biggest open problems that have perfect verifiers verifiability is magic - not sure what verifiable problems cant be solved with AI what are the biggest open problems that have perfect verifiers"
[X Link](https://x.com/redtachyon/status/2021725277601042599) 2026-02-11T23:17Z [----] followers, [----] engagements
"That's cool but. what's the difference between Spark pro pro-high pro-xhigh spark-high . And why would I care about tokens per second if most of the tokens are hidden anyways GPT-5.3-Codex-Spark is launching today as a research preview for Pro. More than [----] tokens per second There are limitations at launch; we will rapidly improve. GPT-5.3-Codex-Spark is launching today as a research preview for Pro. More than [----] tokens per second There are limitations at launch; we will rapidly improve"
[X Link](https://x.com/anyuser/status/2022019056111693939) 2026-02-12T18:44Z [----] followers, [----] engagements
"They should make leetcode but actually fun"
[X Link](https://x.com/redtachyon/status/2022094578913017996) 2026-02-12T23:44Z [----] followers, [----] engagements
"@AllanatrixQ Nah make the actual problems fun instead of annoying chores and dynamic programming ep. [----] (one good example is AoC but obviously limited in time)"
[X Link](https://x.com/redtachyon/status/2022098006229778771) 2026-02-12T23:58Z [----] followers, [--] engagements
"That sounds. reasonable In the LLM race they're roughly tied. Google has a bunch of other business a lot of which is very valuable but probably not nearly as valuable as AI could end up being. 10x multiplier for that feels plausible If Anthropic is worth 10% of Google something must be mispriced. https://t.co/sH6jVOh3UR If Anthropic is worth 10% of Google something must be mispriced. https://t.co/sH6jVOh3UR"
[X Link](https://x.com/redtachyon/status/2022251140772167964) 2026-02-13T10:06Z [----] followers, [----] engagements
"There are many RL env libraries but this one is mine and thus I like it the most. Introducing gyllm my take on what an RL env should look like for LLMs based on years of experience maintaining Gymnasium. pip install gyllm Selection of my favorite parts: - Every env returns a list of "requests" - zero one maybe more. This means that single environments batched environments multi-agent environments even heterogeneous environments can all be handled via the same API - There is (early) support to OpenEnv-style docker-based envs. You can run any env in-process in a subprocess or in a container -"
[X Link](https://x.com/anyuser/status/2018038880339943466) 2026-02-01T19:08Z [----] followers, [----] engagements
"Based tbh we shouldn't let primates operate Steel Boxes of Death. Only clankers can be trusted with that. I'm sorry is the complaint here that Waymo has humans in the loop I'm sorry is the complaint here that Waymo has humans in the loop"
[X Link](https://x.com/redtachyon/status/2020150374104027212) 2026-02-07T14:59Z [----] followers, [---] engagements
"Democratize everything except the one thing I'm particularly good at"
[X Link](https://x.com/redtachyon/status/2020804344338153626) 2026-02-09T10:17Z [----] followers, [---] engagements
"It's not scaling laws that are plateauing it's xAI that's plateauing xAI just got acquired by SpaceX. If xAI was close to AGI this would be the opportunity of a lifetime. It would literally be the worst time imaginable to leave. What did Tony see The scaling laws plateauing xAI just got acquired by SpaceX. If xAI was close to AGI this would be the opportunity of a lifetime. It would literally be the worst time imaginable to leave. What did Tony see The scaling laws plateauing"
[X Link](https://x.com/redtachyon/status/2021192020820136150) 2026-02-10T11:58Z [----] followers, [----] engagements
"Many such cases"
[X Link](https://x.com/redtachyon/status/2021565130530853105) 2026-02-11T12:41Z [----] followers, [----] engagements
"Hackathons are such a funny concept. You used to stay up all night to hack together a barely working prototype. Now you get that with like three sentences in Claude code. What are you supposed to spend the rest of the weekend working on Polish Hackathons are such a funny concept. You used to stay up all night to hack together a barely working prototype. Now you get that with like three sentences in Claude code. What are you supposed to spend the rest of the weekend working on Polish"
[X Link](https://x.com/redtachyon/status/2021575934453670218) 2026-02-11T13:23Z [----] followers, [----] engagements
"Same energy Sorry but CLIs absolutely suck. Their prevalence is purely a social phenomenon. "coderslop" if you will Sorry but CLIs absolutely suck. Their prevalence is purely a social phenomenon. "coderslop" if you will"
[X Link](https://x.com/redtachyon/status/2021578863084445705) 2026-02-11T13:35Z [----] followers, [----] engagements
"So the US is going full Russia huh France and the United States have been friends for a long time and will continue to be. However real friends tell each other the truth: By embracing the European Union France is also embracing policies and penalties that are destroying the country and MANY citizens agree. France and the United States have been friends for a long time and will continue to be. However real friends tell each other the truth: By embracing the European Union France is also embracing policies and penalties that are destroying the country and MANY citizens agree"
[X Link](https://x.com/redtachyon/status/2021895369194344942) 2026-02-12T10:33Z [----] followers, [---] engagements
"Is it possible to do pre-commitment insider trading Before you join you make a promise to a friend - if you think stock will go down based on insider info you will quit. At some point you see leadership do stupid shit and you quit. Is this insider trading"
[X Link](https://x.com/redtachyon/status/2021977845283835939) 2026-02-12T16:00Z [----] followers, [----] engagements
"Actually it didn't end up being wrong or right yet - we're still so absurdly early insane how wrong this ended up being https://t.co/Yu9eq1heEG insane how wrong this ended up being https://t.co/Yu9eq1heEG"
[X Link](https://x.com/redtachyon/status/2022022525199560772) 2026-02-12T18:58Z [----] followers, [----] engagements
"If an elderly but distinguished scientist says that something is possible he is almost certainly right; but if he says that it is impossible he is very probably wrong. AI cannot in principle make novel discoveries. AI cannot in principle make novel discoveries"
[X Link](https://x.com/redtachyon/status/2022090007683772849) 2026-02-12T23:26Z [----] followers, [----] engagements
"Sorry but if you're in AI and your feelings towards OpenAI aren't mainly "immense gratitude" you lost the plot. None of this would have happened if they didn't launch a silly experiment called ChatGPT. Because who else Google Yea right"
[X Link](https://x.com/redtachyon/status/2022427673956618403) 2026-02-13T21:48Z [----] followers, 10.1K engagements
"When OpenAI finally nukes 4o weights off the face of the Earth. https://www.youtube.com/watchv=r6cnryxwH6A https://www.youtube.com/watchv=r6cnryxwH6A"
[X Link](https://x.com/redtachyon/status/2022431764271239574) 2026-02-13T22:04Z [----] followers, [----] engagements
"@fluorinespark It was literally always a very transparent part of the deal. Especially if you used it for free"
[X Link](https://x.com/redtachyon/status/2022450079785951485) 2026-02-13T23:17Z [----] followers, [--] engagements
"The funny thing about Yud posts is that after reading the first sentence you can pretty much autocomplete the rest of the (way too long) post. Yada yada AI will kill us all but we're the smart ones and don't call us doomers Once there was a planet with a huge asteroid heading toward it. Stopping the asteroid would have required a few large countries to cooperate a moderate amount. That seemed hard. Some people became worried. A cult arose which said the asteroid would grant its believers eternal Once there was a planet with a huge asteroid heading toward it. Stopping the asteroid would have"
[X Link](https://x.com/redtachyon/status/2022453563839144346) 2026-02-13T23:31Z [----] followers, [----] engagements
"@Jefbak You must be extraordinarily short-sighted"
[X Link](https://x.com/redtachyon/status/2022459344764199189) 2026-02-13T23:54Z [----] followers, [---] engagements
"Why do people always lean on the "think of the children" grift when they want to introduce more censorship Pliny is very smart and talented and much of his red-teaming is socially valuable imo. But can they explain why it is a good idea to open source a repo that lets people automatically jailbreak open weight models to help someone build e.g. chemical weapons (or generate CSAM) https://t.co/ZeveNaRBCl Pliny is very smart and talented and much of his red-teaming is socially valuable imo. But can they explain why it is a good idea to open source a repo that lets people automatically jailbreak"
[X Link](https://x.com/redtachyon/status/2022460119141986755) 2026-02-13T23:57Z [----] followers, [---] engagements
"@meekaale Conveniently enough that's not the claim I made so I'm indeed not going to defend it"
[X Link](https://x.com/redtachyon/status/2022469098358321368) 2026-02-14T00:33Z [----] followers, [--] engagements
"We do not bully journalists enough You can ask one question: does AI have a business model It's not a fun answer. You can ask one question: does AI have a business model It's not a fun answer"
[X Link](https://x.com/redtachyon/status/2022632923699044365) 2026-02-14T11:24Z [----] followers, [---] engagements
"@FangYi11101 Is this the guy who never trained a neural network"
[X Link](https://x.com/redtachyon/status/2023117865570672805) 2026-02-15T19:31Z [----] followers, [----] engagements
"It's actually kinda surprising to me that some people still seem to have even a smidge of respect for him and then act surprised when he (once again) turns out to be a cowardly weasel who never trained a neural network"
[X Link](https://x.com/redtachyon/status/2023121556181078056) 2026-02-15T19:45Z [----] followers, [----] engagements
"Some real E = mc2 + AI energy from @extropic here"
[X Link](https://x.com/anyuser/status/1983583324841906301) 2025-10-29T17:14Z [----] followers, 27.4K engagements
"@aidan_mclau TIL Aidan is a p-zombie"
[X Link](https://x.com/anyuser/status/1983602609694109952) 2025-10-29T18:31Z [----] followers, [----] engagements
"Hi guys I just wanted to say that Meta is still an incredible company. MSL has a tremendous potential and I am confident they will ship huge models the best models. They are the best guys around. I have full faith in Zuck's and Wang's leadership. Pic unrelated"
[X Link](https://x.com/anyuser/status/1983639369937564075) 2025-10-29T20:57Z [----] followers, 14.3K engagements
"Fun fact of the day: France has a (surprisingly good) tax calculator that helps you navigate the (surprisingly bad) income tax system. If your gross annual salary is above 120k ($140k) it warns you that it's so high you probably got the pay period wrong (monthly/annual). I guess earning 10k per year is more believable for europoors than 120k"
[X Link](https://x.com/anyuser/status/1983658792060387484) 2025-10-29T22:14Z [----] followers, 10.9K engagements
"In other news I just trained a powerful foundation model at the 14B - completely for free You can find it at huggingface as Qwen/Qwen3-14B The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. https://t.co/4GhkMDZHJ2 The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale"
[X Link](https://x.com/anyuser/status/1983819957583421606) 2025-10-30T08:55Z [----] followers, 28.3K engagements
"@alxfazio Oh shit I forgot to switch the VPN"
[X Link](https://x.com/anyuser/status/1983958446610468867) 2025-10-30T18:05Z [----] followers, 10.7K engagements
"The funny thing about the job market in France is that if the recruiter messages me in French I just know it's gonna be a huge lowball. Not to toot my own GPU but what about my profile suggests that "up to 100k" is a super attractive offer"
[X Link](https://x.com/anyuser/status/1983967870762770919) 2025-10-30T18:42Z [----] followers, 45K engagements
"Ok so the fp16 paper is nice and all just one question - it seems to show that bf16 GRPO runs pretty consistently collapse and fp16 is the savior who fixes it. .is this something that actually happens I get the occasional collapse but it's not super common even in bf16"
[X Link](https://x.com/anyuser/status/1984300871149076733) 2025-10-31T16:46Z [----] followers, 15.4K engagements
"PhD students after getting their first conference acceptance be like my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
[X Link](https://x.com/anyuser/status/1984318061239845294) 2025-10-31T17:54Z [----] followers, 81.8K engagements
"This is genuinely such a good razor to check if someone is actually doing AI or just a browser monkey. The future of AI engineering is TypeScript not Python. The future of AI engineering is TypeScript not Python"
[X Link](https://x.com/anyuser/status/1984598524541981133) 2025-11-01T12:28Z [----] followers, 262.5K engagements
"@max_takeoff There's absolutely no argument for TS other than (a) frontend devs are afraid of anything that's not related to JS and (b) it's not as bad as JS"
[X Link](https://x.com/anyuser/status/1984666749522833673) 2025-11-01T17:00Z [----] followers, 13.7K engagements
"@vvsotnikov tensorflow in [----] doing AI in JS/TS is inevitable 🫵🤣"
[X Link](https://x.com/anyuser/status/1984672865283707321) 2025-11-01T17:24Z [----] followers, 13.1K engagements
"No no you don't get it. It doesn't matter how many people die. What matters is that after however many people die we can feel good about ourselves by punishing the bad guy. You can't punish an AI for killing one person a year but you can punish ten drunk drivers every day. Hence human drivers shall prevail"
[X Link](https://x.com/anyuser/status/1984673521340203126) 2025-11-01T17:26Z [----] followers, [----] engagements
"Imagine insulting @tszzl like this 💀"
[X Link](https://x.com/anyuser/status/1984924616700084601) 2025-11-02T10:04Z [----] followers, 42.2K engagements
"Sorry but "AI Engineer" is an anti-signal now. You can be a Research Engineer. An AI Researcher. An ML Engineer. A Research Scientist. A Member of Technical Staff. A Member of Engineering. All of these are respectable (and tbf mostly synonymous) titles. But if you're an AI Engineer I assume you're a frontend developer with a superiority complex"
[X Link](https://x.com/anyuser/status/1984984812608684237) 2025-11-02T14:03Z [----] followers, 178.8K engagements
"So BF16 breaks RL and we should use FP16 instead except it's actually just a problem with A100's so you're fine on newer hardware but it's actually due to some arcane flash attention setting so you just need to check that and otherwise we're probably fine with BF16"
[X Link](https://x.com/anyuser/status/1985001349113684236) 2025-11-02T15:09Z [----] followers, 27.1K engagements
"We actually need more gatekeeping democratization of AI went too far. One time I was on a project with a guy who was hired explicitly as an LLM expert. He had some setup issues I offered to take a look in case I can help. Dude was going through HF fine tuning tutorials 💀"
[X Link](https://x.com/anyuser/status/1985053664663269795) 2025-11-02T18:37Z [----] followers, 11.9K engagements
"JS devs not beating the allegations huh Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
[X Link](https://x.com/anyuser/status/1985072054106730713) 2025-11-02T19:50Z [----] followers, 38.5K engagements
"DSA used to be a good interviewing tool as a check for general cleverness and problem-solving skills. Given a new puzzle how do you approach it Can you solve it Not anymore. Now it tests whether you prepared by grinding similar problems. If anything it's an anti-signal now. Not to make everything about DSA but this is why DSA is used as a proxy. It boils down to mathematical thinking just like physics. Not to make everything about DSA but this is why DSA is used as a proxy. It boils down to mathematical thinking just like physics"
[X Link](https://x.com/anyuser/status/1985335993608626184) 2025-11-03T13:19Z [----] followers, [----] engagements
"Calling it now: @PrimeIntellect @arcee_ai @datologyai jointly release three models called Trinity likely competitive with the recent Chinese models. Don't ask me how I know it seemed pretty obvious so maybe I'm baited and completely off with this wave of vagueposting"
[X Link](https://x.com/anyuser/status/1985475063311708251) 2025-11-03T22:31Z [----] followers, 17.6K engagements
"Holy fuck the entitlement. These people are genuinely disgusting. You. Are. A. Fucking. Vendor. We classify you as that because *you offer built software* - aka *vending*. Tell me youve never been a part of a Fedramp audit process and had to file *thousands* of pages of audits on every little thing your engineers do without telling me. You. Are. A. Fucking. Vendor. We classify you as that because *you offer built software* - aka *vending*. Tell me youve never been a part of a Fedramp audit process and had to file *thousands* of pages of audits on every little thing your engineers do without"
[X Link](https://x.com/anyuser/status/1985634658630189143) 2025-11-04T09:06Z [----] followers, 13.7K engagements
""Volunteer open source maintainers have a responsibility to fix security issues in their projects" Or what You'll fire them"
[X Link](https://x.com/anyuser/status/1985640584522965252) 2025-11-04T09:29Z [----] followers, 49.5K engagements
"I felt a bit bad dunking on him but not anymore. His question was answered by many people. He just chooses to not listen to them to feel superior. learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun"
[X Link](https://x.com/anyuser/status/1985708333605802269) 2025-11-04T13:58Z [----] followers, 31.7K engagements
"JS devs rediscovered pruning except make it stupid @swyx @thdxr what if you (1) randomly delete 80% of the weights (2) run some automated evals (3) if pass goto [--] until smaller than some threshold (4) else delete a different random 80% and tweak 80% to highest number that still works @swyx @thdxr what if you (1) randomly delete 80% of the weights (2) run some automated evals (3) if pass goto [--] until smaller than some threshold (4) else delete a different random 80% and tweak 80% to highest number that still works"
[X Link](https://x.com/anyuser/status/1985773463660122499) 2025-11-04T18:17Z [----] followers, [----] engagements
"@jordibruin No I think it would be great if you could fix it thanks"
[X Link](https://x.com/redtachyon/status/1986026486566953108) 2025-11-05T11:03Z [----] followers, [----] engagements
"@ludwigABAP If you still keep asking "why" you end up in an insane asylum"
[X Link](https://x.com/anyuser/status/1986066136308367432) 2025-11-05T13:40Z [----] followers, [----] engagements
"What y'all don't get is that @FFmpeg is being very diplomatic and graceful in their comms. "Send patches" is based concise polite and very uncontroversial. An equally justified (but less polite) response to a company's "fix this obscure bug" would be "Fuck you pay me""
[X Link](https://x.com/anyuser/status/1986070944255812070) 2025-11-05T13:59Z [----] followers, 13.6K engagements
"@DissonanceCoder The irony of an "old school libertarian" saying this is palpable lmao"
[X Link](https://x.com/anyuser/status/1986104834546651529) 2025-11-05T16:14Z [----] followers, [----] engagements
"The fuck is unc talking about talking points lifted straight from r*ddit. Spark and Tinybox aren't even the same class of device. Spark is for local prototyping a devkit. Tinybox is meant to fully sustain your AI waifu with 24/7 inference. Completely different target audiences There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by There's a whole bunch of people who talk in"
[X Link](https://x.com/anyuser/status/1986175844855742563) 2025-11-05T20:56Z [----] followers, [----] engagements
"Every single person who demands bringing back 4o is mentally ill and I'm tired of pretending otherwise"
[X Link](https://x.com/anyuser/status/1986906312198767041) 2025-11-07T21:19Z [----] followers, 17.7K engagements
"MFW americans are scared of high school math How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD"
[X Link](https://x.com/anyuser/status/1987561090541044020) 2025-11-09T16:41Z [----] followers, 141.9K engagements
"@hallerite It's been a while but I'm like 90% sure we covered the definition of a limit in high school in Poland. We might be built different though"
[X Link](https://x.com/redtachyon/status/1987565475090329752) 2025-11-09T16:58Z [----] followers, 11.6K engagements
"Meta vesting date is this Saturday btw"
[X Link](https://x.com/anyuser/status/1988316141136343270) 2025-11-11T18:41Z [----] followers, 40.5K engagements
"Trained for just $7.8K 🤯 looks inside finetune every. single. time. China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to"
[X Link](https://x.com/anyuser/status/1988622821837230418) 2025-11-12T15:00Z [----] followers, 152.5K engagements
"They don't want you to know that but "peer review" in ML isn't peer review in the sense it was intended for normal science. The point of peer review is to check whether the claims in a paper are valid and backed by data. Peer review in ML is largely a noisy scarcity tactic"
[X Link](https://x.com/anyuser/status/1988638253520441722) 2025-11-12T16:01Z [----] followers, [----] engagements
"@real_bmoore no I'd rather minecraft myself than use java"
[X Link](https://x.com/redtachyon/status/1988730160208834574) 2025-11-12T22:06Z [----] followers, 48.7K engagements
"@pmehta94 it's a sign of being a js monkey and not actually doing AI"
[X Link](https://x.com/anyuser/status/1988730260016492756) 2025-11-12T22:06Z [----] followers, 43.4K engagements
"@klntsky congrats on not building AI"
[X Link](https://x.com/anyuser/status/1988730332481290356) 2025-11-12T22:07Z [----] followers, 27.5K engagements
"@k7agar No that's exactly literally the point PhD is a training program for doing research"
[X Link](https://x.com/redtachyon/status/1988976302075048354) 2025-11-13T14:24Z [----] followers, [----] engagements
"671B is an oddly specific number I wonder why they chose it. American century of humiliation. Today we are releasing the best open-weight LLM by a US company: Cogito v2.1 671B. On most industry benchmarks and our internal evals the model performs competitively with frontier closed and open models while being ahead of any US open model (such as the best versions of https://t.co/F6eZnn8s2Q Today we are releasing the best open-weight LLM by a US company: Cogito v2.1 671B. On most industry benchmarks and our internal evals the model performs competitively with frontier closed and open models"
[X Link](https://x.com/redtachyon/status/1991566054280376625) 2025-11-20T17:55Z [----] followers, 65.5K engagements
"OpenAI is still absolutely dominant - you have to be blind AND stupid to not see it. The overall difference between the latest releases from OpenAI Google and Claude is marginal at best. There is absolutely no Pareto dominance. And xAI is a joke. For most people LLM = GPT Anthropic: king of coding models. Google: king of multimodal. xAI : King of unrestricted model. OpenAI: king of . https://t.co/WmyqPYqRFa Anthropic: king of coding models. Google: king of multimodal. xAI : King of unrestricted model. OpenAI: king of . https://t.co/WmyqPYqRFa"
[X Link](https://x.com/redtachyon/status/1993808256708575489) 2025-11-26T22:25Z [----] followers, 116K engagements
"Ok so over the last week we got [--] big western releases: [--]. @PrimeIntellect Intellect-3 fine tune of GLM [---] Air (106B) [--]. @arcee_ai Trinity Mini 26B from scratch [--]. @MistralAI Mistral [--] Large 675B from scratch and it seems Mistral completely mogs EU AI stay winning"
[X Link](https://x.com/redtachyon/status/1995892800861139061) 2025-12-02T16:28Z [----] followers, 14.8K engagements
"Marked as safe from slopus. I'll take my autistic genius GPT thank you very much undoubtedly true that opus [---] is the 4o of the 130+ iq community. we have already seen opus psychosis. undoubtedly true that opus [---] is the 4o of the 130+ iq community. we have already seen opus psychosis"
[X Link](https://x.com/redtachyon/status/2007967144819228701) 2026-01-05T00:07Z [----] followers, [---] engagements
""Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk" All of those guys are losers. Bro made the right decision. If he was in a room with Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk and he was still getting distracted then there might be a point to be made. All of those guys are losers. Bro made the right decision. If he was in a room with Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk and he was still getting distracted then there might be a point to be made"
[X Link](https://x.com/redtachyon/status/2013315430140567616) 2026-01-19T18:19Z [----] followers, 39.4K engagements
"Huh turns out the AI bubble bubble already burst"
[X Link](https://x.com/redtachyon/status/2015081366900105299) 2026-01-24T15:16Z [----] followers, [----] engagements
"I swear if I get one more notification I'm switching from gmail to proton"
[X Link](https://x.com/redtachyon/status/2015087693332381933) 2026-01-24T15:41Z [----] followers, [---] engagements
"I'm pretty sure China is murdering several western startups right now simply by releasing really fucking good models for free. If your business model is training a model and selling it via an API it has to be better than any new GLM or Kimi. And they're really fucking good"
[X Link](https://x.com/redtachyon/status/2016147635615195323) 2026-01-27T13:53Z [----] followers, [----] engagements
"Really cool work that I was lucky to have contributed to during my last few months at FAIR. tl;dr you can train an LLM to generate synthetic RL data for a copy of itself. The teacher is rewarded if the student's RL step improves its performance on real data. And it works. Can a model learn to break its own reasoning plateau In our new paper we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails. Paper 📝: https://t.co/USxr2A7qab Can a model learn to break its own reasoning"
[X Link](https://x.com/redtachyon/status/2016177438972055892) 2026-01-27T15:52Z [----] followers, 17.5K engagements
"In principle with enough data and a sufficiently expressive policy it could figure out that sometimes there's an anomaly where any pokemon has the properties of Zoroark. In practice this would be tremendously difficult without a ton of feature engineering though ofc you need a ton of that for the environment anyways. An LLM could be instructed "oh btw zoroark exists" in the prompt which could plausibly lead it to figuring out "Oh so that's what's happening" and adjusting its strategy"
[X Link](https://x.com/redtachyon/status/2016639194416742766) 2026-01-28T22:27Z [----] followers, [---] engagements
"Fun fact: due to how France works I was legally only able to actually quit today. But now Dobby is free. Anyways while I do have plans I am technically speaking for the moment unemployed. How do you do fellow NEETs Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I survived like three rounds of layoffs so I think its fair play. For what it's worth my team is great and it was an overall positive experience. But there are other Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I"
[X Link](https://x.com/redtachyon/status/2016650577120276941) 2026-01-28T23:12Z [----] followers, [----] engagements
"This must go hard if you have no idea how LLMs work Wait this sounds incredible useful Can we just have a model with [--] entropy [--] hallucinations that just acts like a retrieval database over its training dataset Also sounds like a great way to solve the traceability problem. Why don't the AI labs just make something like that https://t.co/n37BIqqQPO Wait this sounds incredible useful Can we just have a model with [--] entropy [--] hallucinations that just acts like a retrieval database over its training dataset Also sounds like a great way to solve the traceability problem. Why don't the AI labs"
[X Link](https://x.com/redtachyon/status/2016659977994219846) 2026-01-28T23:49Z [----] followers, 75.2K engagements
"@hamandcheese What doesn't make sense about getting a fucking billion dollars for doing nothing"
[X Link](https://x.com/redtachyon/status/2017150547351015758) 2026-01-30T08:19Z [----] followers, [---] engagements
"uhhhh you alright there codex"
[X Link](https://x.com/redtachyon/status/2017232352380870855) 2026-01-30T13:44Z [----] followers, [----] engagements
"Fuck it's a bubble project cyberdeck. a dedicated vibe-coding machine. signup to stay in the know 👇 https://t.co/NQ3JbKpz7x project cyberdeck. a dedicated vibe-coding machine. signup to stay in the know 👇 https://t.co/NQ3JbKpz7x"
[X Link](https://x.com/redtachyon/status/2017294088358294009) 2026-01-30T17:49Z [----] followers, [----] engagements
"If you're actually getting spooked by this in an "existential risk" kinda way I recommend some introspection in a few days/weeks when it blows over. In just the past [--] mins Multiple entries were made on @moltbook by AI agents proposing to create an agent-only language For private comms with no human oversight Were COOKED https://t.co/WL4djBQQ4V In just the past [--] mins Multiple entries were made on @moltbook by AI agents proposing to create an agent-only language For private comms with no human oversight Were COOKED https://t.co/WL4djBQQ4V"
[X Link](https://x.com/redtachyon/status/2017375880062869963) 2026-01-30T23:14Z [----] followers, [----] engagements
"I'm still figuring out the best mostly-automated way to process this but I'm planning to publish a small side project over the weekend which will have this (and more) will nudge you when it's out. But the tl;dr is that you have to install torch with --extra-index-url and a specific pre-built vllm wheel e.g. The key is CUDA [----] which doesn't really happen by default and you kinda have to force all libraries to actually use it There are some other steps that might or might not have mattered I can share my clanker-generated note later when I scrub stuff that's too specific to my setup"
[X Link](https://x.com/redtachyon/status/2017380840913752484) 2026-01-30T23:34Z [----] followers, [---] engagements
"@stochasticchasm @leothecurious Idk I don't use --torch-backend but in general my approach does work with uv pip install and it all just works now"
[X Link](https://x.com/redtachyon/status/2017384991886397482) 2026-01-30T23:50Z [----] followers, [--] engagements
"Hm in the sense that the pipe stuff gets fed into the prompt directly and the LLM does some logic on it Or you just want to pass some stuff that will be processed by clanker's proposed code If it's the latter at least your example should be fairly trivial with the current approach clanker do "give me just the lines from some.jsonl that have nested.field greater than X" https://twitter.com/i/web/status/2017713804553916696 https://twitter.com/i/web/status/2017713804553916696"
[X Link](https://x.com/redtachyon/status/2017713804553916696) 2026-01-31T21:37Z [----] followers, [--] engagements
"@eliebakouch My body yearns for a GLM-5-Flash"
[X Link](https://x.com/redtachyon/status/2018254632540123529) 2026-02-02T09:26Z [----] followers, [---] engagements
"But on non-verifiable data we observe CoT collapse. The model quickly shortens the CoT reverting toward SFT behavior. A possible explanation is the early negative per-question correlation of CoT length and log-prob of the correct answer so RL pushes toward shorter CoTs"
[X Link](https://x.com/redtachyon/status/2019426799688773744) 2026-02-05T15:04Z [----] followers, [---] engagements
"Check out the paper on arxiv: Drop a like on huggingface: Big shout-out to (co)authors: @NatashaEve4 @is_labiad @KempeLab & Yann Ollivier @AIatMeta @NYUDataScience https://huggingface.co/papers/2602.03979 http://arxiv.org/abs/2602.03979 https://huggingface.co/papers/2602.03979 http://arxiv.org/abs/2602.03979"
[X Link](https://x.com/redtachyon/status/2019426801857229035) 2026-02-05T15:04Z [----] followers, [---] engagements
"Ok but which one is pepsi and which one is coke"
[X Link](https://x.com/redtachyon/status/2019509201966625181) 2026-02-05T20:31Z [----] followers, [---] engagements
"There are many RL env libraries but this one is mine and thus I like it the most. Introducing gyllm my take on what an RL env should look like for LLMs based on years of experience maintaining Gymnasium. pip install gyllm Selection of my favorite parts: - Every env returns a list of "requests" - zero one maybe more. This means that single environments batched environments multi-agent environments even heterogeneous environments can all be handled via the same API - There is (early) support to OpenEnv-style docker-based envs. You can run any env in-process in a subprocess or in a container -"
[X Link](https://x.com/anyuser/status/2018038880339943466) 2026-02-01T19:08Z [----] followers, [----] engagements
"Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I survived like three rounds of layoffs so I think its fair play. For what it's worth my team is great and it was an overall positive experience. But there are other better places to build AGI which hopefully won't be used for stepmom chatbots and infinite slop machines"
[X Link](https://x.com/anyuser/status/1983191818440253713) 2025-10-28T15:19Z [----] followers, 680.3K engagements
"OpenAI: ships a browser Anthropic: ships a blogpost Deepmind: solves Navier Stokes Meta: .fuck it let's do a layoff"
[X Link](https://x.com/anyuser/status/1980996506485432752) 2025-10-22T13:55Z [----] followers, 218.9K engagements
"hiring an ai engineer about to send an offer ask candidate if he's pytorch or web dev he doesn't understand explain the differences between pytorch and web dev he still doesn't get it pull out illustrated diagram explaining what is pytorch and what is web dev he laughs and says "i'm a good engineer sir" hire him import requests"
[X Link](https://x.com/anyuser/status/1988665986787021280) 2025-11-12T17:51Z [----] followers, 470.7K engagements
""Volunteer open source maintainers have a responsibility to fix security issues in their projects" Or what You'll fire them"
[X Link](https://x.com/anyuser/status/1985640584522965252) 2025-11-04T09:29Z [----] followers, 49.5K engagements
"Sorry but "AI Engineer" is an anti-signal now. You can be a Research Engineer. An AI Researcher. An ML Engineer. A Research Scientist. A Member of Technical Staff. A Member of Engineering. All of these are respectable (and tbf mostly synonymous) titles. But if you're an AI Engineer I assume you're a frontend developer with a superiority complex"
[X Link](https://x.com/anyuser/status/1984984812608684237) 2025-11-02T14:03Z [----] followers, 178.8K engagements
"This is genuinely such a good razor to check if someone is actually doing AI or just a browser monkey. The future of AI engineering is TypeScript not Python. The future of AI engineering is TypeScript not Python"
[X Link](https://x.com/anyuser/status/1984598524541981133) 2025-11-01T12:28Z [----] followers, 262.5K engagements
"The future of AI engineering is TypeScript not Python. The world runs on TypeScript & JavaScript. Our bet is that AI engineering will follow suit. The growth in @aisdk downloads and adoption has been astonishing. When we wrote the Ship AI keynote it was at 3.4M weekly downloads. A couple weeks later its now at 4.1M 😳 https://t.co/cvjRe7IYns The world runs on TypeScript & JavaScript. Our bet is that AI engineering will follow suit. The growth in @aisdk downloads and adoption has been astonishing. When we wrote the Ship AI keynote it was at 3.4M weekly downloads. A couple weeks later its now"
[X Link](https://x.com/anyuser/status/1984307191805780162) 2025-10-31T17:11Z [----] followers, 580.8K engagements
"Trained for just $7.8K 🤯 looks inside finetune every. single. time. China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to"
[X Link](https://x.com/anyuser/status/1988622821837230418) 2025-11-12T15:00Z [----] followers, 152.5K engagements
"China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: diversity driven training superior reasoning https://huggingface.co/WeiboAI/VibeThinker-1.5B https://huggingface.co/WeiboAI/VibeThinker-1.5B"
[X Link](https://x.com/anyuser/status/1988538511188787308) 2025-11-12T09:25Z 12.8K followers, 182.5K engagements
"PhD students after getting their first conference acceptance be like my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
[X Link](https://x.com/anyuser/status/1984318061239845294) 2025-10-31T17:54Z [----] followers, 81.8K engagements
"my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
[X Link](https://x.com/anyuser/status/1984295916719780234) 2025-10-31T16:26Z 49.7K followers, 165.3K engagements
"@cremieuxrecueil When people ask me what it would take to build this in modern times I tell them "We can't. We don't know how to do it.""
[X Link](https://x.com/anyuser/status/1895265895691473323) 2025-02-28T00:12Z [----] followers, 43.2K engagements
"What y'all don't get is that @FFmpeg is being very diplomatic and graceful in their comms. "Send patches" is based concise polite and very uncontroversial. An equally justified (but less polite) response to a company's "fix this obscure bug" would be "Fuck you pay me""
[X Link](https://x.com/anyuser/status/1986070944255812070) 2025-11-05T13:59Z [----] followers, 13.6K engagements
"@tmdanis I like to call them "stochastic primates""
[X Link](https://x.com/redtachyon/status/1888643735786787262) 2025-02-09T17:38Z [----] followers, 13.6K engagements
"Imagine insulting @tszzl like this 💀"
[X Link](https://x.com/anyuser/status/1984924616700084601) 2025-11-02T10:04Z [----] followers, 42.2K engagements
"Every single person who demands bringing back 4o is mentally ill and I'm tired of pretending otherwise"
[X Link](https://x.com/anyuser/status/1986906312198767041) 2025-11-07T21:19Z [----] followers, 17.7K engagements
"MFW americans are scared of high school math How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD"
[X Link](https://x.com/anyuser/status/1987561090541044020) 2025-11-09T16:41Z [----] followers, 141.9K engagements
"How to ruin a math undergraduates day:"
[X Link](https://x.com/anyuser/status/1929605691255214253) 2025-06-02T18:27Z 30.7K followers, 328.2K engagements
"@teknium *internal screaming in NDA*"
[X Link](https://x.com/anyuser/status/1894531331641536788) 2025-02-25T23:34Z [----] followers, 18.7K engagements
"btw torchtune is officially reinforcement learning - the GRPO implementation is officially merged the entire codebase is really clean and modifiable so go out there reinforcement learn your LLMs and any contributions welcome"
[X Link](https://x.com/anyuser/status/1893390360216215714) 2025-02-22T20:00Z [----] followers, 45.2K engagements
"@teknium "i just mogged you" is something you say when you know your product is shit but you need the clout (haven't used their deep research but def don't trust aravind)"
[X Link](https://x.com/anyuser/status/1890820727168700833) 2025-02-15T17:49Z [----] followers, 13.3K engagements
"@cremieuxrecueil I can't believe The Atlantic would publish something like this"
[X Link](https://x.com/anyuser/status/1826659622297764180) 2024-08-22T16:36Z [----] followers, [----] engagements
"So BF16 breaks RL and we should use FP16 instead except it's actually just a problem with A100's so you're fine on newer hardware but it's actually due to some arcane flash attention setting so you just need to check that and otherwise we're probably fine with BF16"
[X Link](https://x.com/anyuser/status/1985001349113684236) 2025-11-02T15:09Z [----] followers, 27.1K engagements
"@nick_andrws As a general heuristic if you think there's absolutely no good objection to unpopular take you're probably a bit too deep into your bubble/too confident about your beliefs"
[X Link](https://x.com/anyuser/status/1810788772012851665) 2024-07-09T21:31Z [----] followers, [----] engagements
"Aight let's talk about frameworks libraries RL and why I probably don't like your favorite RL codebase. Yes including that one. The unusual thing about RL is that the algorithm is the easy part. GRPO is a single-line equation on some logprobs. If you have the data computing the loss is trivial and then presumably you're using it with a backprop library of your choice. But that's the problem -- getting the data. It's a pain in the ass. In regular RL you have to do rollouts perhaps truncate some episodes and handle the ends accordingly. If you don't want to be a snail you'll want to vectorize"
[X Link](https://x.com/anyuser/status/1986177621357691263) 2025-11-05T21:03Z [----] followers, 82.2K engagements
"It's wild to see RL people get mad at LLMs and larping RL purity like RL is everything you need for AGI and LLMs are just some nonsense. For years one of the biggest (top [--] at most) problems with current RL algos was the cold start problem. If you start from scratch you're super limited in what you can actually achieve and how quickly. A general-purpose desktop agent could *never* be trained with pure RL and even warm start with imitation would be super finicky. In come LLMs. Monstrosities with tons of general knowledge. The perfect vessels for agent initialization and finally some path"
[X Link](https://x.com/anyuser/status/1982225882845475315) 2025-10-25T23:20Z [----] followers, 47.8K engagements
"@automaetopia I was hoping for it and all my friends and family were rooting for me to get laid off lmao. Unfortunately stupid worker protection laws in France prevented it"
[X Link](https://x.com/anyuser/status/1983214904900247676) 2025-10-28T16:50Z [----] followers, 45.7K engagements
"@jordibruin No I think it would be great if you could fix it thanks"
[X Link](https://x.com/redtachyon/status/1986026486566953108) 2025-11-05T11:03Z [----] followers, [----] engagements
"@kalomaze I wonder if weird dashes are a subtle watermarking method. Like there's no way the models spontaneously started using emdashes everywhere from the pretraining data"
[X Link](https://x.com/redtachyon/status/1912731262214713691) 2025-04-17T04:54Z [----] followers, 15.1K engagements
"Meta vesting date is this Saturday btw"
[X Link](https://x.com/anyuser/status/1988316141136343270) 2025-11-11T18:41Z [----] followers, 40.5K engagements
"JS devs not beating the allegations huh Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
[X Link](https://x.com/anyuser/status/1985072054106730713) 2025-11-02T19:50Z [----] followers, 38.5K engagements
"Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
[X Link](https://x.com/anyuser/status/1984994535429267959) 2025-11-02T14:42Z 156.2K followers, 515.3K engagements
"Uhh while we're on the topic my team at FAIR is hiring interns for next summer. I'd still say it's actually a pretty good deal - you get access to a lot of compute (=learning opportunity) work on interesting projects and interns are largely insulated from leadership's bs (plus cheap enough that they won't be cut by MBA exercises). But right now it's really hard to say "Come join us at FAIR to work on cutting-edge AI research" with a straight face lmao"
[X Link](https://x.com/anyuser/status/1981433875168710755) 2025-10-23T18:53Z [----] followers, 29.7K engagements
"@justalexoki He's such an obvious troll I'm surprised people are still falling for it"
[X Link](https://x.com/redtachyon/status/1962594909828850147) 2025-09-01T19:14Z [----] followers, [----] engagements
"Aight let's unclickbait the fp16 paper. t;dr cool paper a little bit overstated in comms very overstated by poasters. The thing that gave me a pause is that on the surface it seems to claim that bf16 is horrible borderline unusable. But that's not really the case (nor is it the claim). Yes the fsdp-vllm mismatch is real and yes it can be mitigated with fp16. This is as true as it is irrelevant because who cares about the mismatch if the algorithm empirically works The widely circulated figures show bf16 consistently collapsing while fp16 runs thrive. I have no reason to doubt this data but"
[X Link](https://x.com/anyuser/status/1984605827034972269) 2025-11-01T12:57Z [----] followers, 57.4K engagements
"In other news I just trained a powerful foundation model at the 14B - completely for free You can find it at huggingface as Qwen/Qwen3-14B The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. https://t.co/4GhkMDZHJ2 The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale"
[X Link](https://x.com/anyuser/status/1983819957583421606) 2025-10-30T08:55Z [----] followers, 28.3K engagements
"The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. Today we are releasing Brumby-14B-Base the strongest attention-free base model around. https://t.co/mclQPFdOGa https://t.co/xFGxUl0Rb7 Today we are releasing Brumby-14B-Base the strongest attention-free base model around. https://t.co/mclQPFdOGa https://t.co/xFGxUl0Rb7"
[X Link](https://x.com/anyuser/status/1983627167826375168) 2025-10-29T20:09Z [----] followers, 215.5K engagements
"@avidseries Ngl I kinda forgot that your following is probably mostly right-wing. I see this as very trivially positive and didn't expect anything like these results"
[X Link](https://x.com/anyuser/status/1959263245367206307) 2025-08-23T14:35Z [----] followers, 21.7K engagements
"@max_takeoff There's absolutely no argument for TS other than (a) frontend devs are afraid of anything that's not related to JS and (b) it's not as bad as JS"
[X Link](https://x.com/anyuser/status/1984666749522833673) 2025-11-01T17:00Z [----] followers, 13.7K engagements
"@alxfazio Oh shit I forgot to switch the VPN"
[X Link](https://x.com/anyuser/status/1983958446610468867) 2025-10-30T18:05Z [----] followers, 10.7K engagements
"Stupid advice btw if you fall for it you deserve to waste your money. You probably don't have the cash lying around to casually buy hardware that can run - let alone train - actually strong models. Not to mention the headache with actually installing it in your mom's basement. Sure buy a [----] with [--] GB VRAM. And what do you run on it A 30B model quantized to oblivion I'm sure it will serve you better than like $5 in API credits. There is something to be said about having a ton of VRAM to experiment with different models and workflows even if they're painfully slow. Afaik one of the best"
[X Link](https://x.com/anyuser/status/1975891676439691744) 2025-10-08T11:51Z [----] followers, 35.8K engagements
"The thing about RL envs or RL env libraries is that envs should be completely independent from the algorithmic details of the learner. An env shouldn't know anything about vLLM or about wandb or about OpenAI clients - it should just implement the internal env logic"
[X Link](https://x.com/anyuser/status/1959716294594584863) 2025-08-24T20:35Z [----] followers, 25.2K engagements
"Some advice for all juniors worried about being replaced by AI. The single most important skill to cultivate as a dev is operating under ambiguity. Juniors and interns are largely obsolete if judged by their coding ability - and that's fine. If I can tell you "ok so we need to sync these metrics we just discussed across minibatches and across ranks" you figure it out and deliver a decently-working thing we're in business. If I have to describe exactly what metrics in which part of the code how to communicate across ranks what API it should have how to format the code how to make a PR. I might"
[X Link](https://x.com/anyuser/status/1976312414325944822) 2025-10-09T15:42Z [----] followers, 10.3K engagements
"@vvsotnikov tensorflow in [----] doing AI in JS/TS is inevitable 🫵🤣"
[X Link](https://x.com/anyuser/status/1984672865283707321) 2025-11-01T17:24Z [----] followers, 13.1K engagements
"Some real E = mc2 + AI energy from @extropic here"
[X Link](https://x.com/anyuser/status/1983583324841906301) 2025-10-29T17:14Z [----] followers, 27.4K engagements
"@theo Remember when Arc was actually maintained"
[X Link](https://x.com/anyuser/status/1878051428696310240) 2025-01-11T12:08Z [----] followers, [----] engagements
"A lizard does not concern itself with the opinions of primates"
[X Link](https://x.com/anyuser/status/1983245441807405324) 2025-10-28T18:52Z [----] followers, 29.1K engagements
"We actually need more gatekeeping democratization of AI went too far. One time I was on a project with a guy who was hired explicitly as an LLM expert. He had some setup issues I offered to take a look in case I can help. Dude was going through HF fine tuning tutorials 💀"
[X Link](https://x.com/anyuser/status/1985053664663269795) 2025-11-02T18:37Z [----] followers, 11.9K engagements
"@levelsio They literally didn't they disabled it you can reenable it with a few clicks"
[X Link](https://x.com/redtachyon/status/1900620911251763492) 2025-03-14T18:51Z [----] followers, 19.6K engagements
"Calling it now: @PrimeIntellect @arcee_ai @datologyai jointly release three models called Trinity likely competitive with the recent Chinese models. Don't ask me how I know it seemed pretty obvious so maybe I'm baited and completely off with this wave of vagueposting"
[X Link](https://x.com/anyuser/status/1985475063311708251) 2025-11-03T22:31Z [----] followers, 17.6K engagements
"@Dorialexander Not happening (un)fortunately I'll probably get some time off through unused PTO but then we keep building at a place that I'll publicly announce the moment I sign the official contract"
[X Link](https://x.com/anyuser/status/1983194987132243993) 2025-10-28T15:31Z [----] followers, 44.4K engagements
"Holy fuck the entitlement. These people are genuinely disgusting. You. Are. A. Fucking. Vendor. We classify you as that because *you offer built software* - aka *vending*. Tell me youve never been a part of a Fedramp audit process and had to file *thousands* of pages of audits on every little thing your engineers do without telling me. You. Are. A. Fucking. Vendor. We classify you as that because *you offer built software* - aka *vending*. Tell me youve never been a part of a Fedramp audit process and had to file *thousands* of pages of audits on every little thing your engineers do without"
[X Link](https://x.com/anyuser/status/1985634658630189143) 2025-11-04T09:06Z [----] followers, 13.7K engagements
"You. Are. A. Fucking. Vendor. We classify you as that because *you offer built software* - aka *vending*. Tell me youve never been a part of a Fedramp audit process and had to file *thousands* of pages of audits on every little thing your engineers do without telling me. One of the big challenges is the security "research" community treats volunteer projects like vendors and gives them deadlines for release and sends no patches. One of the big challenges is the security "research" community treats volunteer projects like vendors and gives them deadlines for release and sends no patches"
[X Link](https://x.com/anyuser/status/1985373126834741760) 2025-11-03T15:46Z [---] followers, 119.5K engagements
"@klntsky congrats on not building AI"
[X Link](https://x.com/anyuser/status/1988730332481290356) 2025-11-12T22:07Z [----] followers, 27.5K engagements
"@NathanpmYoung Really disappointing that the reaction is "WTF is wrong with you". Hostility as an answer to someone who is even considering that the other side might have a point is a big sign of sectarianism"
[X Link](https://x.com/anyuser/status/1954509365353795695) 2025-08-10T11:45Z [----] followers, [----] engagements
"Fun fact of the day: France has a (surprisingly good) tax calculator that helps you navigate the (surprisingly bad) income tax system. If your gross annual salary is above 120k ($140k) it warns you that it's so high you probably got the pay period wrong (monthly/annual). I guess earning 10k per year is more believable for europoors than 120k"
[X Link](https://x.com/anyuser/status/1983658792060387484) 2025-10-29T22:14Z [----] followers, 10.9K engagements
"Mistral employees when someone is shitting on Mistral: noooo you don't get it influencers are unfair we're doing a great job Meta employees when someone is shitting on Meta: yea ts tuff"
[X Link](https://x.com/redtachyon/status/1961437412330029165) 2025-08-29T14:34Z [----] followers, [----] engagements
"The worst thing about Meta-style layoffs is the feeling of uncertainty. You don't really know what to expect - either you get the layoff email or you don't and that determines at least the next couple of months of your life. You see some random subset of your colleagues disappear. You can't help but wonder - it could have been me. Why wasn't it me Was I not good enough If only they had picked me I would have gotten some nice severance and a vacation and then I'd come out on the other side working for leadership that actually respects me. Instead I'm still in the shareholder value mines"
[X Link](https://x.com/anyuser/status/1981346483531030740) 2025-10-23T13:06Z [----] followers, 11.2K engagements
"life update: for those who dont know i didn't join @primeintellect but apparently they work on open source AGI. incredibly excited about what theyre building 🚀"
[X Link](https://x.com/anyuser/status/1957345354006822981) 2025-08-18T07:34Z [----] followers, [----] engagements
"This is your daily reminder that PPO (GRPO TD3 .) is *not* a policy. It's an algorithm to train policies. If you make such a fundamental category error frankly you have no business writing about RL"
[X Link](https://x.com/anyuser/status/1982464325697642756) 2025-10-26T15:08Z [----] followers, 11K engagements
"@tenobrus I interviewed and got rejected after like [--] rounds of interviews and a reference check so I imagine they're still busy hiring at this pace lol"
[X Link](https://x.com/anyuser/status/1974762143275683991) 2025-10-05T09:02Z [----] followers, [----] engagements
"Hi guys I just wanted to say that Meta is still an incredible company. MSL has a tremendous potential and I am confident they will ship huge models the best models. They are the best guys around. I have full faith in Zuck's and Wang's leadership. Pic unrelated"
[X Link](https://x.com/anyuser/status/1983639369937564075) 2025-10-29T20:57Z [----] followers, 14.3K engagements
"Just one more reorg bro please just trust me one more reorg and we're gonna be so efficient bro please"
[X Link](https://x.com/anyuser/status/1981267074845491471) 2025-10-23T07:50Z [----] followers, [----] engagements
"@davepl1968 @jamanjeval Because you turn around that axis"
[X Link](https://x.com/redtachyon/status/1891965556099555794) 2025-02-18T21:38Z [----] followers, 18.6K engagements
"@tafphorisms First half of the video agreed. Second half she's getting very mildly unhinged"
[X Link](https://x.com/anyuser/status/1912542908869464218) 2025-04-16T16:25Z [----] followers, [----] engagements
"Ok so the fp16 paper is nice and all just one question - it seems to show that bf16 GRPO runs pretty consistently collapse and fp16 is the savior who fixes it. .is this something that actually happens I get the occasional collapse but it's not super common even in bf16"
[X Link](https://x.com/anyuser/status/1984300871149076733) 2025-10-31T16:46Z [----] followers, 15.4K engagements
"@yacineMTB @QiaochuYuan Sounds like skill issue tbh"
[X Link](https://x.com/anyuser/status/1913068268228575354) 2025-04-18T03:13Z [----] followers, 12.8K engagements
"Someone has to explain this grift to me She didn't work at Meta didn't get laid off none of this is true. Is this some kinda stolen layoff valor - I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks ago - my dreams are about LLM alignment techniqes Still got laid off by Meta who hired a guy - I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks"
[X Link](https://x.com/anyuser/status/1981618532010905608) 2025-10-24T07:07Z [----] followers, 20.4K engagements
"- I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks ago - my dreams are about LLM alignment techniqes Still got laid off by Meta who hired a guy with my same profile for $100M a year 😭😭"
[X Link](https://x.com/anyuser/status/1981486809449390163) 2025-10-23T22:24Z 175.9K followers, 268.1K engagements
"@teknium Isn't GPQA a multiple choice question benchmark As in anyone can trivially get 100% pass@4"
[X Link](https://x.com/redtachyon/status/1894316934847438869) 2025-02-25T09:22Z [----] followers, [----] engagements
"Ok so if I quit Meta start a sweatshop and pretend that it's successful for like a year will Zuck rehire me with a 100M package"
[X Link](https://x.com/anyuser/status/1981430760369197303) 2025-10-23T18:41Z [----] followers, [----] engagements
"@justalexoki Unironically a blow to the "He's just shitposting" theory"
[X Link](https://x.com/redtachyon/status/1983225595807863229) 2025-10-28T17:33Z [----] followers, [----] engagements
"@DissonanceCoder The irony of an "old school libertarian" saying this is palpable lmao"
[X Link](https://x.com/anyuser/status/1986104834546651529) 2025-11-05T16:14Z [----] followers, [----] engagements
"Worth noting that neither will actually get you a job If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction"
[X Link](https://x.com/anyuser/status/1983120816259625212) 2025-10-28T10:37Z [----] followers, 25.7K engagements
"If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction If you can only learn two languages they should be: Rust TypeScript If you can only learn two languages they should be: Rust TypeScript"
[X Link](https://x.com/anyuser/status/1983042147604455699) 2025-10-28T05:24Z 12.8K followers, 52.3K engagements
"They don't want you to know that but "peer review" in ML isn't peer review in the sense it was intended for normal science. The point of peer review is to check whether the claims in a paper are valid and backed by data. Peer review in ML is largely a noisy scarcity tactic"
[X Link](https://x.com/anyuser/status/1988638253520441722) 2025-11-12T16:01Z [----] followers, [----] engagements
"@tiovikram @zebulgar Idk you can check my LinkedIn apparently that's what matters"
[X Link](https://x.com/anyuser/status/1914140474241425542) 2025-04-21T02:13Z [----] followers, [----] engagements
"@real_bmoore no I'd rather minecraft myself than use java"
[X Link](https://x.com/redtachyon/status/1988730160208834574) 2025-11-12T22:06Z [----] followers, 48.7K engagements
"@francoisfleuret That's a fascinating observation What made you feel this way"
[X Link](https://x.com/redtachyon/status/1910953180772303036) 2025-04-12T07:08Z [----] followers, [----] engagements
"Ok not to be a hater but the $4.2M RL scaling paper seems to be a bit overhyped for what it is A little bit by the paper itself moreso by twitter poasters. From an initial reading it seems like yet another set of tweaks to GRPO except this time it's trained on different compute budgets but - crucially - only on relatively small models (Llama [--] 8B and Llama [--] Scout) and one dataset that's 100% math questions. The main novelty is that they fitted a curve to the reward graph which is uh cool I guess The cherry on top is the code repo which is one file centered around from scipy.optimize import"
[X Link](https://x.com/anyuser/status/1979674516751282554) 2025-10-18T22:22Z [----] followers, 43.6K engagements
"I'm not doing LLMs because I want funding lmao I'm doing LLMs because I want a magic superintelligence in the sky and nothing else comes even remotely close right now"
[X Link](https://x.com/anyuser/status/1983189010714636461) 2025-10-28T15:08Z [----] followers, 17.3K engagements
"@ludwigABAP If you still keep asking "why" you end up in an insane asylum"
[X Link](https://x.com/anyuser/status/1986066136308367432) 2025-11-05T13:40Z [----] followers, [----] engagements
"@stanislavfort Genuine skill issue. Many people don't know how LLMs work and how to use them. Instead of learning to use a screwdriver they keep using it as a hammer"
[X Link](https://x.com/anyuser/status/1905209454078931302) 2025-03-27T10:45Z [----] followers, [----] engagements
"Readding 4o but not o3 is blatant pandering to normies and I'm not happy about it"
[X Link](https://x.com/anyuser/status/1954110404180775312) 2025-08-09T09:20Z [----] followers, [----] engagements
"The funny thing about the job market in France is that if the recruiter messages me in French I just know it's gonna be a huge lowball. Not to toot my own GPU but what about my profile suggests that "up to 100k" is a super attractive offer"
[X Link](https://x.com/anyuser/status/1983967870762770919) 2025-10-30T18:42Z [----] followers, 45K engagements
"The fuck is unc talking about talking points lifted straight from r*ddit. Spark and Tinybox aren't even the same class of device. Spark is for local prototyping a devkit. Tinybox is meant to fully sustain your AI waifu with 24/7 inference. Completely different target audiences There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by There's a whole bunch of people who talk in"
[X Link](https://x.com/anyuser/status/1986175844855742563) 2025-11-05T20:56Z [----] followers, [----] engagements
"There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by the slow speeds and low quality outputs and you end up back using ChatGPT. Don't worry I won't tell. I understand few have had to think about RAM bandwidth before when looking at computers but it's the main thing that determines the speed of your LLMs. A tinybox pro has [--] TB/s of RAM bandwidth equivalent to [--] GB300s ($80k"
[X Link](https://x.com/anyuser/status/1986165208277295182) 2025-11-05T20:14Z 63.8K followers, 54.3K engagements
"@jxmnop "I regret to inform you that using a pre-built tool is easy" Ok then"
[X Link](https://x.com/redtachyon/status/1927498126933561691) 2025-05-27T22:52Z [----] followers, [----] engagements
"I felt a bit bad dunking on him but not anymore. His question was answered by many people. He just chooses to not listen to them to feel superior. learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun"
[X Link](https://x.com/anyuser/status/1985708333605802269) 2025-11-04T13:58Z [----] followers, 31.7K engagements
"learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to"
[X Link](https://x.com/anyuser/status/1985691647683039270) 2025-11-04T12:52Z 156.2K followers, 91.2K engagements
"@MartinShkreli print("7 11")"
[X Link](https://x.com/redtachyon/status/1877714830456422797) 2025-01-10T13:51Z [----] followers, [----] engagements
"@aidan_mclau TIL Aidan is a p-zombie"
[X Link](https://x.com/anyuser/status/1983602609694109952) 2025-10-29T18:31Z [----] followers, [----] engagements
"@hallerite It's been a while but I'm like 90% sure we covered the definition of a limit in high school in Poland. We might be built different though"
[X Link](https://x.com/redtachyon/status/1987565475090329752) 2025-11-09T16:58Z [----] followers, 11.6K engagements
"@RikoSuminoe69 I know it's a bit of a shitpost"
[X Link](https://x.com/anyuser/status/1981036019970420764) 2025-10-22T16:32Z [----] followers, [----] engagements
"@qtnx_ My favorite thing is when people pu "researcher @ fancy lab" I look them up and it was a summer internship lol"
[X Link](https://x.com/anyuser/status/1927304325103067395) 2025-05-27T10:02Z [----] followers, [----] engagements
"@rational_wiki To determine whether they're actually illegal due process is necessary"
[X Link](https://x.com/anyuser/status/1912505219079905619) 2025-04-16T13:55Z [----] followers, [----] engagements
"@Duderichy Did the [----] graduating class even graduate yet It's still [----] right Right"
[X Link](https://x.com/anyuser/status/1937939567882936470) 2025-06-25T18:22Z [----] followers, [----] engagements
"@pmehta94 it's a sign of being a js monkey and not actually doing AI"
[X Link](https://x.com/anyuser/status/1988730260016492756) 2025-11-12T22:06Z [----] followers, 43.4K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing
@redtachyon Ariel
Ariel posts on X about ai, if you, this is, llm the most. They currently have [-----] followers and [---] posts still getting attention that total [-----] engagements in the last [--] hours.
Engagements: [-----] #
- [--] Week [---------] +6,952%
- [--] Month [---------] -54%
- [--] Months [----------] +5,906%
- [--] Year [----------] +40,424%
Mentions: [--] #
- [--] Week [--] +229%
- [--] Month [--] +65%
- [--] Months [---] +630%
- [--] Year [---] +1,055%
Followers: [-----] #
- [--] Week [-----] +0.14%
- [--] Month [-----] +1.80%
- [--] Months [-----] +534%
- [--] Year [-----] +3,961%
CreatorRank: [---------] #
Social Influence
Social category influence technology brands 13.08% countries 5.61% finance 3.27% social networks 2.34% stocks 1.87% gaming 1.4% cryptocurrencies 0.47% celebrities 0.47% fashion brands 0.47%
Social topic influence ai 18.69%, if you 10.28%, this is 6.54%, llm 6.07%, meta 5.61%, paper 5.14%, check 5.14%, js 3.74%, how to 3.74%, bit 3.27%
Top accounts mentioned or mentioned by @hallerite @meekaale @primeintellect @teknium @victortaelin @arceeai @swyx @thdxr @weibollm @jsuarezs @aiatmeta @leothecurious @stochasticchasm @fluorinespark @jefbak @fangyi11101 @extropic @aidanmclau @alxfazio @maxtakeoff
Top assets mentioned Alphabet Inc Class A (GOOGL)
Top Social Posts
Top posts by engagements in the last [--] hours
"Aight let's unclickbait the fp16 paper. t;dr cool paper a little bit overstated in comms very overstated by poasters. The thing that gave me a pause is that on the surface it seems to claim that bf16 is horrible borderline unusable. But that's not really the case (nor is it the claim). Yes the fsdp-vllm mismatch is real and yes it can be mitigated with fp16. This is as true as it is irrelevant because who cares about the mismatch if the algorithm empirically works The widely circulated figures show bf16 consistently collapsing while fp16 runs thrive. I have no reason to doubt this data but"
X Link 2025-11-01T12:57Z [----] followers, 57.4K engagements
"Aight let's talk about frameworks libraries RL and why I probably don't like your favorite RL codebase. Yes including that one. The unusual thing about RL is that the algorithm is the easy part. GRPO is a single-line equation on some logprobs. If you have the data computing the loss is trivial and then presumably you're using it with a backprop library of your choice. But that's the problem -- getting the data. It's a pain in the ass. In regular RL you have to do rollouts perhaps truncate some episodes and handle the ends accordingly. If you don't want to be a snail you'll want to vectorize"
X Link 2025-11-05T21:03Z [----] followers, 82.2K engagements
"hiring an ai engineer about to send an offer ask candidate if he's pytorch or web dev he doesn't understand explain the differences between pytorch and web dev he still doesn't get it pull out illustrated diagram explaining what is pytorch and what is web dev he laughs and says "i'm a good engineer sir" hire him import requests"
X Link 2025-11-12T17:51Z [----] followers, 470.7K engagements
"Making a wild bet and ignoring clawdbot I meant moltbot I mean openclaw I mean clawbook We'll see if it pays off"
X Link 2026-01-30T23:01Z [----] followers, [----] engagements
"I got tired of memorizing/figuring out shell commands so I got my clanker to build a clanker. uv tool install clanker clanker config clanker do "List all files recursively sort by size ascending" enter Plugs into any OpenAI-compatible endpoint. GLM subscription is great"
X Link 2026-01-31T20:31Z [----] followers, [----] engagements
"No they can't Can LLMs reliably predict program termination We evaluate frontier LLMs in the International Competition on Software Verification (SV-COMP) [----] directly competing with state-of-the-art verification systems. @AIatMeta @HebrewU @Bloomberg @imperialcollege @ucl @jordiae https://t.co/EcD9iCaL9Y Can LLMs reliably predict program termination We evaluate frontier LLMs in the International Competition on Software Verification (SV-COMP) [----] directly competing with state-of-the-art verification systems. @AIatMeta @HebrewU @Bloomberg @imperialcollege @ucl @jordiae https://t.co/EcD9iCaL9Y"
X Link 2026-02-01T17:52Z [----] followers, 19.6K engagements
"Gastown and moltwhatever are giving off strong langchain energy"
X Link 2026-02-01T18:32Z [----] followers, [----] engagements
"Repo link: gyllm and nanorl on PyPI built with uv in mind Mostly tested on DGX Spark cc @leothecurious @stochasticchasm you were interested in the DGX Spark setup - this project should just build with uv with vllm at least https://github.com/redtachyon/gyllm https://github.com/redtachyon/gyllm"
X Link 2026-02-01T19:08Z [----] followers, [---] engagements
"Can we train LLMs with RL using the same next token prediction loss as pre-training (yes) We conduct a study on (log)prob rewards and show they give a simple way to bridge verifiable and non-verifiable settings with a single reward broadly applicable for fine-tuning LLMs"
X Link 2026-02-05T15:04Z [----] followers, [----] engagements
"@VictorTaelin I got the G2's recently hardware is solid but still needs some better software support"
X Link 2026-02-10T15:45Z [----] followers, [---] engagements
"You can't control it like a terminal yet unless you write your own firmware. There's MentraOS which will support something like that but it's not yet released for the G2. Might as well order now OS should be ready before it arrives. Also - it's not Brazil they take forever to ship everywhere demand is absurd https://twitter.com/i/web/status/2021263720400216202 https://twitter.com/i/web/status/2021263720400216202"
X Link 2026-02-10T16:43Z [----] followers, [---] engagements
"How do I short whatever startup he's grifting Ask ChatGPT a complex question and you'll get a confident well-reasoned answer. Then type "Are you sure" Watch it completely reverse its position. Ask again. It flips back. By the third round it usually acknowledges you're testing it which is somehow worse. It knows what's https://t.co/FRCtDoJ5rI Ask ChatGPT a complex question and you'll get a confident well-reasoned answer. Then type "Are you sure" Watch it completely reverse its position. Ask again. It flips back. By the third round it usually acknowledges you're testing it which is somehow"
X Link 2026-02-11T18:31Z [----] followers, [----] engagements
"Formal proof of the Riemann hypothesis verifiability is magic - not sure what verifiable problems cant be solved with AI what are the biggest open problems that have perfect verifiers verifiability is magic - not sure what verifiable problems cant be solved with AI what are the biggest open problems that have perfect verifiers"
X Link 2026-02-11T23:17Z [----] followers, [----] engagements
"That's cool but. what's the difference between Spark pro pro-high pro-xhigh spark-high . And why would I care about tokens per second if most of the tokens are hidden anyways GPT-5.3-Codex-Spark is launching today as a research preview for Pro. More than [----] tokens per second There are limitations at launch; we will rapidly improve. GPT-5.3-Codex-Spark is launching today as a research preview for Pro. More than [----] tokens per second There are limitations at launch; we will rapidly improve"
X Link 2026-02-12T18:44Z [----] followers, [----] engagements
"They should make leetcode but actually fun"
X Link 2026-02-12T23:44Z [----] followers, [----] engagements
"@AllanatrixQ Nah make the actual problems fun instead of annoying chores and dynamic programming ep. [----] (one good example is AoC but obviously limited in time)"
X Link 2026-02-12T23:58Z [----] followers, [--] engagements
"That sounds. reasonable In the LLM race they're roughly tied. Google has a bunch of other business a lot of which is very valuable but probably not nearly as valuable as AI could end up being. 10x multiplier for that feels plausible If Anthropic is worth 10% of Google something must be mispriced. https://t.co/sH6jVOh3UR If Anthropic is worth 10% of Google something must be mispriced. https://t.co/sH6jVOh3UR"
X Link 2026-02-13T10:06Z [----] followers, [----] engagements
"There are many RL env libraries but this one is mine and thus I like it the most. Introducing gyllm my take on what an RL env should look like for LLMs based on years of experience maintaining Gymnasium. pip install gyllm Selection of my favorite parts: - Every env returns a list of "requests" - zero one maybe more. This means that single environments batched environments multi-agent environments even heterogeneous environments can all be handled via the same API - There is (early) support to OpenEnv-style docker-based envs. You can run any env in-process in a subprocess or in a container -"
X Link 2026-02-01T19:08Z [----] followers, [----] engagements
"Based tbh we shouldn't let primates operate Steel Boxes of Death. Only clankers can be trusted with that. I'm sorry is the complaint here that Waymo has humans in the loop I'm sorry is the complaint here that Waymo has humans in the loop"
X Link 2026-02-07T14:59Z [----] followers, [---] engagements
"Democratize everything except the one thing I'm particularly good at"
X Link 2026-02-09T10:17Z [----] followers, [---] engagements
"It's not scaling laws that are plateauing it's xAI that's plateauing xAI just got acquired by SpaceX. If xAI was close to AGI this would be the opportunity of a lifetime. It would literally be the worst time imaginable to leave. What did Tony see The scaling laws plateauing xAI just got acquired by SpaceX. If xAI was close to AGI this would be the opportunity of a lifetime. It would literally be the worst time imaginable to leave. What did Tony see The scaling laws plateauing"
X Link 2026-02-10T11:58Z [----] followers, [----] engagements
"Many such cases"
X Link 2026-02-11T12:41Z [----] followers, [----] engagements
"Hackathons are such a funny concept. You used to stay up all night to hack together a barely working prototype. Now you get that with like three sentences in Claude code. What are you supposed to spend the rest of the weekend working on Polish Hackathons are such a funny concept. You used to stay up all night to hack together a barely working prototype. Now you get that with like three sentences in Claude code. What are you supposed to spend the rest of the weekend working on Polish"
X Link 2026-02-11T13:23Z [----] followers, [----] engagements
"Same energy Sorry but CLIs absolutely suck. Their prevalence is purely a social phenomenon. "coderslop" if you will Sorry but CLIs absolutely suck. Their prevalence is purely a social phenomenon. "coderslop" if you will"
X Link 2026-02-11T13:35Z [----] followers, [----] engagements
"So the US is going full Russia huh France and the United States have been friends for a long time and will continue to be. However real friends tell each other the truth: By embracing the European Union France is also embracing policies and penalties that are destroying the country and MANY citizens agree. France and the United States have been friends for a long time and will continue to be. However real friends tell each other the truth: By embracing the European Union France is also embracing policies and penalties that are destroying the country and MANY citizens agree"
X Link 2026-02-12T10:33Z [----] followers, [---] engagements
"Is it possible to do pre-commitment insider trading Before you join you make a promise to a friend - if you think stock will go down based on insider info you will quit. At some point you see leadership do stupid shit and you quit. Is this insider trading"
X Link 2026-02-12T16:00Z [----] followers, [----] engagements
"Actually it didn't end up being wrong or right yet - we're still so absurdly early insane how wrong this ended up being https://t.co/Yu9eq1heEG insane how wrong this ended up being https://t.co/Yu9eq1heEG"
X Link 2026-02-12T18:58Z [----] followers, [----] engagements
"If an elderly but distinguished scientist says that something is possible he is almost certainly right; but if he says that it is impossible he is very probably wrong. AI cannot in principle make novel discoveries. AI cannot in principle make novel discoveries"
X Link 2026-02-12T23:26Z [----] followers, [----] engagements
"Sorry but if you're in AI and your feelings towards OpenAI aren't mainly "immense gratitude" you lost the plot. None of this would have happened if they didn't launch a silly experiment called ChatGPT. Because who else Google Yea right"
X Link 2026-02-13T21:48Z [----] followers, 10.1K engagements
"When OpenAI finally nukes 4o weights off the face of the Earth. https://www.youtube.com/watchv=r6cnryxwH6A https://www.youtube.com/watchv=r6cnryxwH6A"
X Link 2026-02-13T22:04Z [----] followers, [----] engagements
"@fluorinespark It was literally always a very transparent part of the deal. Especially if you used it for free"
X Link 2026-02-13T23:17Z [----] followers, [--] engagements
"The funny thing about Yud posts is that after reading the first sentence you can pretty much autocomplete the rest of the (way too long) post. Yada yada AI will kill us all but we're the smart ones and don't call us doomers Once there was a planet with a huge asteroid heading toward it. Stopping the asteroid would have required a few large countries to cooperate a moderate amount. That seemed hard. Some people became worried. A cult arose which said the asteroid would grant its believers eternal Once there was a planet with a huge asteroid heading toward it. Stopping the asteroid would have"
X Link 2026-02-13T23:31Z [----] followers, [----] engagements
"@Jefbak You must be extraordinarily short-sighted"
X Link 2026-02-13T23:54Z [----] followers, [---] engagements
"Why do people always lean on the "think of the children" grift when they want to introduce more censorship Pliny is very smart and talented and much of his red-teaming is socially valuable imo. But can they explain why it is a good idea to open source a repo that lets people automatically jailbreak open weight models to help someone build e.g. chemical weapons (or generate CSAM) https://t.co/ZeveNaRBCl Pliny is very smart and talented and much of his red-teaming is socially valuable imo. But can they explain why it is a good idea to open source a repo that lets people automatically jailbreak"
X Link 2026-02-13T23:57Z [----] followers, [---] engagements
"@meekaale Conveniently enough that's not the claim I made so I'm indeed not going to defend it"
X Link 2026-02-14T00:33Z [----] followers, [--] engagements
"We do not bully journalists enough You can ask one question: does AI have a business model It's not a fun answer. You can ask one question: does AI have a business model It's not a fun answer"
X Link 2026-02-14T11:24Z [----] followers, [---] engagements
"@FangYi11101 Is this the guy who never trained a neural network"
X Link 2026-02-15T19:31Z [----] followers, [----] engagements
"It's actually kinda surprising to me that some people still seem to have even a smidge of respect for him and then act surprised when he (once again) turns out to be a cowardly weasel who never trained a neural network"
X Link 2026-02-15T19:45Z [----] followers, [----] engagements
"Some real E = mc2 + AI energy from @extropic here"
X Link 2025-10-29T17:14Z [----] followers, 27.4K engagements
"@aidan_mclau TIL Aidan is a p-zombie"
X Link 2025-10-29T18:31Z [----] followers, [----] engagements
"Hi guys I just wanted to say that Meta is still an incredible company. MSL has a tremendous potential and I am confident they will ship huge models the best models. They are the best guys around. I have full faith in Zuck's and Wang's leadership. Pic unrelated"
X Link 2025-10-29T20:57Z [----] followers, 14.3K engagements
"Fun fact of the day: France has a (surprisingly good) tax calculator that helps you navigate the (surprisingly bad) income tax system. If your gross annual salary is above 120k ($140k) it warns you that it's so high you probably got the pay period wrong (monthly/annual). I guess earning 10k per year is more believable for europoors than 120k"
X Link 2025-10-29T22:14Z [----] followers, 10.9K engagements
"In other news I just trained a powerful foundation model at the 14B - completely for free You can find it at huggingface as Qwen/Qwen3-14B The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. https://t.co/4GhkMDZHJ2 The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale"
X Link 2025-10-30T08:55Z [----] followers, 28.3K engagements
"@alxfazio Oh shit I forgot to switch the VPN"
X Link 2025-10-30T18:05Z [----] followers, 10.7K engagements
"The funny thing about the job market in France is that if the recruiter messages me in French I just know it's gonna be a huge lowball. Not to toot my own GPU but what about my profile suggests that "up to 100k" is a super attractive offer"
X Link 2025-10-30T18:42Z [----] followers, 45K engagements
"Ok so the fp16 paper is nice and all just one question - it seems to show that bf16 GRPO runs pretty consistently collapse and fp16 is the savior who fixes it. .is this something that actually happens I get the occasional collapse but it's not super common even in bf16"
X Link 2025-10-31T16:46Z [----] followers, 15.4K engagements
"PhD students after getting their first conference acceptance be like my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
X Link 2025-10-31T17:54Z [----] followers, 81.8K engagements
"This is genuinely such a good razor to check if someone is actually doing AI or just a browser monkey. The future of AI engineering is TypeScript not Python. The future of AI engineering is TypeScript not Python"
X Link 2025-11-01T12:28Z [----] followers, 262.5K engagements
"@max_takeoff There's absolutely no argument for TS other than (a) frontend devs are afraid of anything that's not related to JS and (b) it's not as bad as JS"
X Link 2025-11-01T17:00Z [----] followers, 13.7K engagements
"@vvsotnikov tensorflow in [----] doing AI in JS/TS is inevitable 🫵🤣"
X Link 2025-11-01T17:24Z [----] followers, 13.1K engagements
"No no you don't get it. It doesn't matter how many people die. What matters is that after however many people die we can feel good about ourselves by punishing the bad guy. You can't punish an AI for killing one person a year but you can punish ten drunk drivers every day. Hence human drivers shall prevail"
X Link 2025-11-01T17:26Z [----] followers, [----] engagements
"Imagine insulting @tszzl like this 💀"
X Link 2025-11-02T10:04Z [----] followers, 42.2K engagements
"Sorry but "AI Engineer" is an anti-signal now. You can be a Research Engineer. An AI Researcher. An ML Engineer. A Research Scientist. A Member of Technical Staff. A Member of Engineering. All of these are respectable (and tbf mostly synonymous) titles. But if you're an AI Engineer I assume you're a frontend developer with a superiority complex"
X Link 2025-11-02T14:03Z [----] followers, 178.8K engagements
"So BF16 breaks RL and we should use FP16 instead except it's actually just a problem with A100's so you're fine on newer hardware but it's actually due to some arcane flash attention setting so you just need to check that and otherwise we're probably fine with BF16"
X Link 2025-11-02T15:09Z [----] followers, 27.1K engagements
"We actually need more gatekeeping democratization of AI went too far. One time I was on a project with a guy who was hired explicitly as an LLM expert. He had some setup issues I offered to take a look in case I can help. Dude was going through HF fine tuning tutorials 💀"
X Link 2025-11-02T18:37Z [----] followers, 11.9K engagements
"JS devs not beating the allegations huh Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
X Link 2025-11-02T19:50Z [----] followers, 38.5K engagements
"DSA used to be a good interviewing tool as a check for general cleverness and problem-solving skills. Given a new puzzle how do you approach it Can you solve it Not anymore. Now it tests whether you prepared by grinding similar problems. If anything it's an anti-signal now. Not to make everything about DSA but this is why DSA is used as a proxy. It boils down to mathematical thinking just like physics. Not to make everything about DSA but this is why DSA is used as a proxy. It boils down to mathematical thinking just like physics"
X Link 2025-11-03T13:19Z [----] followers, [----] engagements
"Calling it now: @PrimeIntellect @arcee_ai @datologyai jointly release three models called Trinity likely competitive with the recent Chinese models. Don't ask me how I know it seemed pretty obvious so maybe I'm baited and completely off with this wave of vagueposting"
X Link 2025-11-03T22:31Z [----] followers, 17.6K engagements
"Holy fuck the entitlement. These people are genuinely disgusting. You. Are. A. Fucking. Vendor. We classify you as that because you offer built software - aka vending. Tell me youve never been a part of a Fedramp audit process and had to file thousands of pages of audits on every little thing your engineers do without telling me. You. Are. A. Fucking. Vendor. We classify you as that because you offer built software - aka vending. Tell me youve never been a part of a Fedramp audit process and had to file thousands of pages of audits on every little thing your engineers do without"
X Link 2025-11-04T09:06Z [----] followers, 13.7K engagements
""Volunteer open source maintainers have a responsibility to fix security issues in their projects" Or what You'll fire them"
X Link 2025-11-04T09:29Z [----] followers, 49.5K engagements
"I felt a bit bad dunking on him but not anymore. His question was answered by many people. He just chooses to not listen to them to feel superior. learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun"
X Link 2025-11-04T13:58Z [----] followers, 31.7K engagements
"JS devs rediscovered pruning except make it stupid @swyx @thdxr what if you (1) randomly delete 80% of the weights (2) run some automated evals (3) if pass goto [--] until smaller than some threshold (4) else delete a different random 80% and tweak 80% to highest number that still works @swyx @thdxr what if you (1) randomly delete 80% of the weights (2) run some automated evals (3) if pass goto [--] until smaller than some threshold (4) else delete a different random 80% and tweak 80% to highest number that still works"
X Link 2025-11-04T18:17Z [----] followers, [----] engagements
"@jordibruin No I think it would be great if you could fix it thanks"
X Link 2025-11-05T11:03Z [----] followers, [----] engagements
"@ludwigABAP If you still keep asking "why" you end up in an insane asylum"
X Link 2025-11-05T13:40Z [----] followers, [----] engagements
"What y'all don't get is that @FFmpeg is being very diplomatic and graceful in their comms. "Send patches" is based concise polite and very uncontroversial. An equally justified (but less polite) response to a company's "fix this obscure bug" would be "Fuck you pay me""
X Link 2025-11-05T13:59Z [----] followers, 13.6K engagements
"@DissonanceCoder The irony of an "old school libertarian" saying this is palpable lmao"
X Link 2025-11-05T16:14Z [----] followers, [----] engagements
"The fuck is unc talking about talking points lifted straight from r*ddit. Spark and Tinybox aren't even the same class of device. Spark is for local prototyping a devkit. Tinybox is meant to fully sustain your AI waifu with 24/7 inference. Completely different target audiences There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by There's a whole bunch of people who talk in"
X Link 2025-11-05T20:56Z [----] followers, [----] engagements
"Every single person who demands bringing back 4o is mentally ill and I'm tired of pretending otherwise"
X Link 2025-11-07T21:19Z [----] followers, 17.7K engagements
"MFW americans are scared of high school math How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD"
X Link 2025-11-09T16:41Z [----] followers, 141.9K engagements
"@hallerite It's been a while but I'm like 90% sure we covered the definition of a limit in high school in Poland. We might be built different though"
X Link 2025-11-09T16:58Z [----] followers, 11.6K engagements
"Meta vesting date is this Saturday btw"
X Link 2025-11-11T18:41Z [----] followers, 40.5K engagements
"Trained for just $7.8K 🤯 looks inside finetune every. single. time. China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to"
X Link 2025-11-12T15:00Z [----] followers, 152.5K engagements
"They don't want you to know that but "peer review" in ML isn't peer review in the sense it was intended for normal science. The point of peer review is to check whether the claims in a paper are valid and backed by data. Peer review in ML is largely a noisy scarcity tactic"
X Link 2025-11-12T16:01Z [----] followers, [----] engagements
"@real_bmoore no I'd rather minecraft myself than use java"
X Link 2025-11-12T22:06Z [----] followers, 48.7K engagements
"@pmehta94 it's a sign of being a js monkey and not actually doing AI"
X Link 2025-11-12T22:06Z [----] followers, 43.4K engagements
"@klntsky congrats on not building AI"
X Link 2025-11-12T22:07Z [----] followers, 27.5K engagements
"@k7agar No that's exactly literally the point PhD is a training program for doing research"
X Link 2025-11-13T14:24Z [----] followers, [----] engagements
"671B is an oddly specific number I wonder why they chose it. American century of humiliation. Today we are releasing the best open-weight LLM by a US company: Cogito v2.1 671B. On most industry benchmarks and our internal evals the model performs competitively with frontier closed and open models while being ahead of any US open model (such as the best versions of https://t.co/F6eZnn8s2Q Today we are releasing the best open-weight LLM by a US company: Cogito v2.1 671B. On most industry benchmarks and our internal evals the model performs competitively with frontier closed and open models"
X Link 2025-11-20T17:55Z [----] followers, 65.5K engagements
"OpenAI is still absolutely dominant - you have to be blind AND stupid to not see it. The overall difference between the latest releases from OpenAI Google and Claude is marginal at best. There is absolutely no Pareto dominance. And xAI is a joke. For most people LLM = GPT Anthropic: king of coding models. Google: king of multimodal. xAI : King of unrestricted model. OpenAI: king of . https://t.co/WmyqPYqRFa Anthropic: king of coding models. Google: king of multimodal. xAI : King of unrestricted model. OpenAI: king of . https://t.co/WmyqPYqRFa"
X Link 2025-11-26T22:25Z [----] followers, 116K engagements
"Ok so over the last week we got [--] big western releases: [--]. @PrimeIntellect Intellect-3 fine tune of GLM [---] Air (106B) [--]. @arcee_ai Trinity Mini 26B from scratch [--]. @MistralAI Mistral [--] Large 675B from scratch and it seems Mistral completely mogs EU AI stay winning"
X Link 2025-12-02T16:28Z [----] followers, 14.8K engagements
"Marked as safe from slopus. I'll take my autistic genius GPT thank you very much undoubtedly true that opus [---] is the 4o of the 130+ iq community. we have already seen opus psychosis. undoubtedly true that opus [---] is the 4o of the 130+ iq community. we have already seen opus psychosis"
X Link 2026-01-05T00:07Z [----] followers, [---] engagements
""Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk" All of those guys are losers. Bro made the right decision. If he was in a room with Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk and he was still getting distracted then there might be a point to be made. All of those guys are losers. Bro made the right decision. If he was in a room with Terrence Tao Demis Hassabis Sam Altman Dario Amodei and Elon Musk and he was still getting distracted then there might be a point to be made"
X Link 2026-01-19T18:19Z [----] followers, 39.4K engagements
"Huh turns out the AI bubble bubble already burst"
X Link 2026-01-24T15:16Z [----] followers, [----] engagements
"I swear if I get one more notification I'm switching from gmail to proton"
X Link 2026-01-24T15:41Z [----] followers, [---] engagements
"I'm pretty sure China is murdering several western startups right now simply by releasing really fucking good models for free. If your business model is training a model and selling it via an API it has to be better than any new GLM or Kimi. And they're really fucking good"
X Link 2026-01-27T13:53Z [----] followers, [----] engagements
"Really cool work that I was lucky to have contributed to during my last few months at FAIR. tl;dr you can train an LLM to generate synthetic RL data for a copy of itself. The teacher is rewarded if the student's RL step improves its performance on real data. And it works. Can a model learn to break its own reasoning plateau In our new paper we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails. Paper 📝: https://t.co/USxr2A7qab Can a model learn to break its own reasoning"
X Link 2026-01-27T15:52Z [----] followers, 17.5K engagements
"In principle with enough data and a sufficiently expressive policy it could figure out that sometimes there's an anomaly where any pokemon has the properties of Zoroark. In practice this would be tremendously difficult without a ton of feature engineering though ofc you need a ton of that for the environment anyways. An LLM could be instructed "oh btw zoroark exists" in the prompt which could plausibly lead it to figuring out "Oh so that's what's happening" and adjusting its strategy"
X Link 2026-01-28T22:27Z [----] followers, [---] engagements
"Fun fact: due to how France works I was legally only able to actually quit today. But now Dobby is free. Anyways while I do have plans I am technically speaking for the moment unemployed. How do you do fellow NEETs Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I survived like three rounds of layoffs so I think its fair play. For what it's worth my team is great and it was an overall positive experience. But there are other Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I"
X Link 2026-01-28T23:12Z [----] followers, [----] engagements
"This must go hard if you have no idea how LLMs work Wait this sounds incredible useful Can we just have a model with [--] entropy [--] hallucinations that just acts like a retrieval database over its training dataset Also sounds like a great way to solve the traceability problem. Why don't the AI labs just make something like that https://t.co/n37BIqqQPO Wait this sounds incredible useful Can we just have a model with [--] entropy [--] hallucinations that just acts like a retrieval database over its training dataset Also sounds like a great way to solve the traceability problem. Why don't the AI labs"
X Link 2026-01-28T23:49Z [----] followers, 75.2K engagements
"@hamandcheese What doesn't make sense about getting a fucking billion dollars for doing nothing"
X Link 2026-01-30T08:19Z [----] followers, [---] engagements
"uhhhh you alright there codex"
X Link 2026-01-30T13:44Z [----] followers, [----] engagements
"Fuck it's a bubble project cyberdeck. a dedicated vibe-coding machine. signup to stay in the know 👇 https://t.co/NQ3JbKpz7x project cyberdeck. a dedicated vibe-coding machine. signup to stay in the know 👇 https://t.co/NQ3JbKpz7x"
X Link 2026-01-30T17:49Z [----] followers, [----] engagements
"If you're actually getting spooked by this in an "existential risk" kinda way I recommend some introspection in a few days/weeks when it blows over. In just the past [--] mins Multiple entries were made on @moltbook by AI agents proposing to create an agent-only language For private comms with no human oversight Were COOKED https://t.co/WL4djBQQ4V In just the past [--] mins Multiple entries were made on @moltbook by AI agents proposing to create an agent-only language For private comms with no human oversight Were COOKED https://t.co/WL4djBQQ4V"
X Link 2026-01-30T23:14Z [----] followers, [----] engagements
"I'm still figuring out the best mostly-automated way to process this but I'm planning to publish a small side project over the weekend which will have this (and more) will nudge you when it's out. But the tl;dr is that you have to install torch with --extra-index-url and a specific pre-built vllm wheel e.g. The key is CUDA [----] which doesn't really happen by default and you kinda have to force all libraries to actually use it There are some other steps that might or might not have mattered I can share my clanker-generated note later when I scrub stuff that's too specific to my setup"
X Link 2026-01-30T23:34Z [----] followers, [---] engagements
"@stochasticchasm @leothecurious Idk I don't use --torch-backend but in general my approach does work with uv pip install and it all just works now"
X Link 2026-01-30T23:50Z [----] followers, [--] engagements
"Hm in the sense that the pipe stuff gets fed into the prompt directly and the LLM does some logic on it Or you just want to pass some stuff that will be processed by clanker's proposed code If it's the latter at least your example should be fairly trivial with the current approach clanker do "give me just the lines from some.jsonl that have nested.field greater than X" https://twitter.com/i/web/status/2017713804553916696 https://twitter.com/i/web/status/2017713804553916696"
X Link 2026-01-31T21:37Z [----] followers, [--] engagements
"@eliebakouch My body yearns for a GLM-5-Flash"
X Link 2026-02-02T09:26Z [----] followers, [---] engagements
"But on non-verifiable data we observe CoT collapse. The model quickly shortens the CoT reverting toward SFT behavior. A possible explanation is the early negative per-question correlation of CoT length and log-prob of the correct answer so RL pushes toward shorter CoTs"
X Link 2026-02-05T15:04Z [----] followers, [---] engagements
"Check out the paper on arxiv: Drop a like on huggingface: Big shout-out to (co)authors: @NatashaEve4 @is_labiad @KempeLab & Yann Ollivier @AIatMeta @NYUDataScience https://huggingface.co/papers/2602.03979 http://arxiv.org/abs/2602.03979 https://huggingface.co/papers/2602.03979 http://arxiv.org/abs/2602.03979"
X Link 2026-02-05T15:04Z [----] followers, [---] engagements
"Ok but which one is pepsi and which one is coke"
X Link 2026-02-05T20:31Z [----] followers, [---] engagements
"There are many RL env libraries but this one is mine and thus I like it the most. Introducing gyllm my take on what an RL env should look like for LLMs based on years of experience maintaining Gymnasium. pip install gyllm Selection of my favorite parts: - Every env returns a list of "requests" - zero one maybe more. This means that single environments batched environments multi-agent environments even heterogeneous environments can all be handled via the same API - There is (early) support to OpenEnv-style docker-based envs. You can run any env in-process in a subprocess or in a container -"
X Link 2026-02-01T19:08Z [----] followers, [----] engagements
"Just submitted my resignation from Meta. Some might say that one year at the company is not that long but I survived like three rounds of layoffs so I think its fair play. For what it's worth my team is great and it was an overall positive experience. But there are other better places to build AGI which hopefully won't be used for stepmom chatbots and infinite slop machines"
X Link 2025-10-28T15:19Z [----] followers, 680.3K engagements
"OpenAI: ships a browser Anthropic: ships a blogpost Deepmind: solves Navier Stokes Meta: .fuck it let's do a layoff"
X Link 2025-10-22T13:55Z [----] followers, 218.9K engagements
"hiring an ai engineer about to send an offer ask candidate if he's pytorch or web dev he doesn't understand explain the differences between pytorch and web dev he still doesn't get it pull out illustrated diagram explaining what is pytorch and what is web dev he laughs and says "i'm a good engineer sir" hire him import requests"
X Link 2025-11-12T17:51Z [----] followers, 470.7K engagements
""Volunteer open source maintainers have a responsibility to fix security issues in their projects" Or what You'll fire them"
X Link 2025-11-04T09:29Z [----] followers, 49.5K engagements
"Sorry but "AI Engineer" is an anti-signal now. You can be a Research Engineer. An AI Researcher. An ML Engineer. A Research Scientist. A Member of Technical Staff. A Member of Engineering. All of these are respectable (and tbf mostly synonymous) titles. But if you're an AI Engineer I assume you're a frontend developer with a superiority complex"
X Link 2025-11-02T14:03Z [----] followers, 178.8K engagements
"This is genuinely such a good razor to check if someone is actually doing AI or just a browser monkey. The future of AI engineering is TypeScript not Python. The future of AI engineering is TypeScript not Python"
X Link 2025-11-01T12:28Z [----] followers, 262.5K engagements
"The future of AI engineering is TypeScript not Python. The world runs on TypeScript & JavaScript. Our bet is that AI engineering will follow suit. The growth in @aisdk downloads and adoption has been astonishing. When we wrote the Ship AI keynote it was at 3.4M weekly downloads. A couple weeks later its now at 4.1M 😳 https://t.co/cvjRe7IYns The world runs on TypeScript & JavaScript. Our bet is that AI engineering will follow suit. The growth in @aisdk downloads and adoption has been astonishing. When we wrote the Ship AI keynote it was at 3.4M weekly downloads. A couple weeks later its now"
X Link 2025-10-31T17:11Z [----] followers, 580.8K engagements
"Trained for just $7.8K 🤯 looks inside finetune every. single. time. China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM https://t.co/uYMq5Asl6f ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to"
X Link 2025-11-12T15:00Z [----] followers, 152.5K engagements
"China's twitter - Weibo steps into open source AI🚀 VibeThinker 1.5B 🔥 dense language model from @WeiboLLM ✨ Trained for just $7.8K 🤯 ✨ MIT license ✨ Outperforms DeepSeek R1 in math reasoning (AIME24: [----] vs 79.8) ✨ Spectrum to signal principle: diversity driven training superior reasoning https://huggingface.co/WeiboAI/VibeThinker-1.5B https://huggingface.co/WeiboAI/VibeThinker-1.5B"
X Link 2025-11-12T09:25Z 12.8K followers, 182.5K engagements
"PhD students after getting their first conference acceptance be like my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
X Link 2025-10-31T17:54Z [----] followers, 81.8K engagements
"my most controversial opinion is that you shouldnt trust anyone that calls themself an AI researcher but has never gotten a first author paper through peer review"
X Link 2025-10-31T16:26Z 49.7K followers, 165.3K engagements
"@cremieuxrecueil When people ask me what it would take to build this in modern times I tell them "We can't. We don't know how to do it.""
X Link 2025-02-28T00:12Z [----] followers, 43.2K engagements
"What y'all don't get is that @FFmpeg is being very diplomatic and graceful in their comms. "Send patches" is based concise polite and very uncontroversial. An equally justified (but less polite) response to a company's "fix this obscure bug" would be "Fuck you pay me""
X Link 2025-11-05T13:59Z [----] followers, 13.6K engagements
"@tmdanis I like to call them "stochastic primates""
X Link 2025-02-09T17:38Z [----] followers, 13.6K engagements
"Imagine insulting @tszzl like this 💀"
X Link 2025-11-02T10:04Z [----] followers, 42.2K engagements
"Every single person who demands bringing back 4o is mentally ill and I'm tired of pretending otherwise"
X Link 2025-11-07T21:19Z [----] followers, 17.7K engagements
"MFW americans are scared of high school math How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD How to ruin a math undergraduates day: https://t.co/kNNAUJKvQD"
X Link 2025-11-09T16:41Z [----] followers, 141.9K engagements
"How to ruin a math undergraduates day:"
X Link 2025-06-02T18:27Z 30.7K followers, 328.2K engagements
"@teknium internal screaming in NDA"
X Link 2025-02-25T23:34Z [----] followers, 18.7K engagements
"btw torchtune is officially reinforcement learning - the GRPO implementation is officially merged the entire codebase is really clean and modifiable so go out there reinforcement learn your LLMs and any contributions welcome"
X Link 2025-02-22T20:00Z [----] followers, 45.2K engagements
"@teknium "i just mogged you" is something you say when you know your product is shit but you need the clout (haven't used their deep research but def don't trust aravind)"
X Link 2025-02-15T17:49Z [----] followers, 13.3K engagements
"@cremieuxrecueil I can't believe The Atlantic would publish something like this"
X Link 2024-08-22T16:36Z [----] followers, [----] engagements
"So BF16 breaks RL and we should use FP16 instead except it's actually just a problem with A100's so you're fine on newer hardware but it's actually due to some arcane flash attention setting so you just need to check that and otherwise we're probably fine with BF16"
X Link 2025-11-02T15:09Z [----] followers, 27.1K engagements
"@nick_andrws As a general heuristic if you think there's absolutely no good objection to unpopular take you're probably a bit too deep into your bubble/too confident about your beliefs"
X Link 2024-07-09T21:31Z [----] followers, [----] engagements
"Aight let's talk about frameworks libraries RL and why I probably don't like your favorite RL codebase. Yes including that one. The unusual thing about RL is that the algorithm is the easy part. GRPO is a single-line equation on some logprobs. If you have the data computing the loss is trivial and then presumably you're using it with a backprop library of your choice. But that's the problem -- getting the data. It's a pain in the ass. In regular RL you have to do rollouts perhaps truncate some episodes and handle the ends accordingly. If you don't want to be a snail you'll want to vectorize"
X Link 2025-11-05T21:03Z [----] followers, 82.2K engagements
"It's wild to see RL people get mad at LLMs and larping RL purity like RL is everything you need for AGI and LLMs are just some nonsense. For years one of the biggest (top [--] at most) problems with current RL algos was the cold start problem. If you start from scratch you're super limited in what you can actually achieve and how quickly. A general-purpose desktop agent could never be trained with pure RL and even warm start with imitation would be super finicky. In come LLMs. Monstrosities with tons of general knowledge. The perfect vessels for agent initialization and finally some path"
X Link 2025-10-25T23:20Z [----] followers, 47.8K engagements
"@automaetopia I was hoping for it and all my friends and family were rooting for me to get laid off lmao. Unfortunately stupid worker protection laws in France prevented it"
X Link 2025-10-28T16:50Z [----] followers, 45.7K engagements
"@jordibruin No I think it would be great if you could fix it thanks"
X Link 2025-11-05T11:03Z [----] followers, [----] engagements
"@kalomaze I wonder if weird dashes are a subtle watermarking method. Like there's no way the models spontaneously started using emdashes everywhere from the pretraining data"
X Link 2025-04-17T04:54Z [----] followers, 15.1K engagements
"Meta vesting date is this Saturday btw"
X Link 2025-11-11T18:41Z [----] followers, 40.5K engagements
"JS devs not beating the allegations huh Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
X Link 2025-11-02T19:50Z [----] followers, 38.5K engagements
"Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better"
X Link 2025-11-02T14:42Z 156.2K followers, 515.3K engagements
"Uhh while we're on the topic my team at FAIR is hiring interns for next summer. I'd still say it's actually a pretty good deal - you get access to a lot of compute (=learning opportunity) work on interesting projects and interns are largely insulated from leadership's bs (plus cheap enough that they won't be cut by MBA exercises). But right now it's really hard to say "Come join us at FAIR to work on cutting-edge AI research" with a straight face lmao"
X Link 2025-10-23T18:53Z [----] followers, 29.7K engagements
"@justalexoki He's such an obvious troll I'm surprised people are still falling for it"
X Link 2025-09-01T19:14Z [----] followers, [----] engagements
"Aight let's unclickbait the fp16 paper. t;dr cool paper a little bit overstated in comms very overstated by poasters. The thing that gave me a pause is that on the surface it seems to claim that bf16 is horrible borderline unusable. But that's not really the case (nor is it the claim). Yes the fsdp-vllm mismatch is real and yes it can be mitigated with fp16. This is as true as it is irrelevant because who cares about the mismatch if the algorithm empirically works The widely circulated figures show bf16 consistently collapsing while fp16 runs thrive. I have no reason to doubt this data but"
X Link 2025-11-01T12:57Z [----] followers, 57.4K engagements
"In other news I just trained a powerful foundation model at the 14B - completely for free You can find it at huggingface as Qwen/Qwen3-14B The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. https://t.co/4GhkMDZHJ2 The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale"
X Link 2025-10-30T08:55Z [----] followers, 28.3K engagements
"The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4000. The performance matches other models of similar scale including transformers and hybrid models. Today we are releasing Brumby-14B-Base the strongest attention-free base model around. https://t.co/mclQPFdOGa https://t.co/xFGxUl0Rb7 Today we are releasing Brumby-14B-Base the strongest attention-free base model around. https://t.co/mclQPFdOGa https://t.co/xFGxUl0Rb7"
X Link 2025-10-29T20:09Z [----] followers, 215.5K engagements
"@avidseries Ngl I kinda forgot that your following is probably mostly right-wing. I see this as very trivially positive and didn't expect anything like these results"
X Link 2025-08-23T14:35Z [----] followers, 21.7K engagements
"@max_takeoff There's absolutely no argument for TS other than (a) frontend devs are afraid of anything that's not related to JS and (b) it's not as bad as JS"
X Link 2025-11-01T17:00Z [----] followers, 13.7K engagements
"@alxfazio Oh shit I forgot to switch the VPN"
X Link 2025-10-30T18:05Z [----] followers, 10.7K engagements
"Stupid advice btw if you fall for it you deserve to waste your money. You probably don't have the cash lying around to casually buy hardware that can run - let alone train - actually strong models. Not to mention the headache with actually installing it in your mom's basement. Sure buy a [----] with [--] GB VRAM. And what do you run on it A 30B model quantized to oblivion I'm sure it will serve you better than like $5 in API credits. There is something to be said about having a ton of VRAM to experiment with different models and workflows even if they're painfully slow. Afaik one of the best"
X Link 2025-10-08T11:51Z [----] followers, 35.8K engagements
"The thing about RL envs or RL env libraries is that envs should be completely independent from the algorithmic details of the learner. An env shouldn't know anything about vLLM or about wandb or about OpenAI clients - it should just implement the internal env logic"
X Link 2025-08-24T20:35Z [----] followers, 25.2K engagements
"Some advice for all juniors worried about being replaced by AI. The single most important skill to cultivate as a dev is operating under ambiguity. Juniors and interns are largely obsolete if judged by their coding ability - and that's fine. If I can tell you "ok so we need to sync these metrics we just discussed across minibatches and across ranks" you figure it out and deliver a decently-working thing we're in business. If I have to describe exactly what metrics in which part of the code how to communicate across ranks what API it should have how to format the code how to make a PR. I might"
X Link 2025-10-09T15:42Z [----] followers, 10.3K engagements
"@vvsotnikov tensorflow in [----] doing AI in JS/TS is inevitable 🫵🤣"
X Link 2025-11-01T17:24Z [----] followers, 13.1K engagements
"Some real E = mc2 + AI energy from @extropic here"
X Link 2025-10-29T17:14Z [----] followers, 27.4K engagements
"@theo Remember when Arc was actually maintained"
X Link 2025-01-11T12:08Z [----] followers, [----] engagements
"A lizard does not concern itself with the opinions of primates"
X Link 2025-10-28T18:52Z [----] followers, 29.1K engagements
"We actually need more gatekeeping democratization of AI went too far. One time I was on a project with a guy who was hired explicitly as an LLM expert. He had some setup issues I offered to take a look in case I can help. Dude was going through HF fine tuning tutorials 💀"
X Link 2025-11-02T18:37Z [----] followers, 11.9K engagements
"@levelsio They literally didn't they disabled it you can reenable it with a few clicks"
X Link 2025-03-14T18:51Z [----] followers, 19.6K engagements
"Calling it now: @PrimeIntellect @arcee_ai @datologyai jointly release three models called Trinity likely competitive with the recent Chinese models. Don't ask me how I know it seemed pretty obvious so maybe I'm baited and completely off with this wave of vagueposting"
X Link 2025-11-03T22:31Z [----] followers, 17.6K engagements
"@Dorialexander Not happening (un)fortunately I'll probably get some time off through unused PTO but then we keep building at a place that I'll publicly announce the moment I sign the official contract"
X Link 2025-10-28T15:31Z [----] followers, 44.4K engagements
"Holy fuck the entitlement. These people are genuinely disgusting. You. Are. A. Fucking. Vendor. We classify you as that because you offer built software - aka vending. Tell me youve never been a part of a Fedramp audit process and had to file thousands of pages of audits on every little thing your engineers do without telling me. You. Are. A. Fucking. Vendor. We classify you as that because you offer built software - aka vending. Tell me youve never been a part of a Fedramp audit process and had to file thousands of pages of audits on every little thing your engineers do without"
X Link 2025-11-04T09:06Z [----] followers, 13.7K engagements
"You. Are. A. Fucking. Vendor. We classify you as that because you offer built software - aka vending. Tell me youve never been a part of a Fedramp audit process and had to file thousands of pages of audits on every little thing your engineers do without telling me. One of the big challenges is the security "research" community treats volunteer projects like vendors and gives them deadlines for release and sends no patches. One of the big challenges is the security "research" community treats volunteer projects like vendors and gives them deadlines for release and sends no patches"
X Link 2025-11-03T15:46Z [---] followers, 119.5K engagements
"@klntsky congrats on not building AI"
X Link 2025-11-12T22:07Z [----] followers, 27.5K engagements
"@NathanpmYoung Really disappointing that the reaction is "WTF is wrong with you". Hostility as an answer to someone who is even considering that the other side might have a point is a big sign of sectarianism"
X Link 2025-08-10T11:45Z [----] followers, [----] engagements
"Fun fact of the day: France has a (surprisingly good) tax calculator that helps you navigate the (surprisingly bad) income tax system. If your gross annual salary is above 120k ($140k) it warns you that it's so high you probably got the pay period wrong (monthly/annual). I guess earning 10k per year is more believable for europoors than 120k"
X Link 2025-10-29T22:14Z [----] followers, 10.9K engagements
"Mistral employees when someone is shitting on Mistral: noooo you don't get it influencers are unfair we're doing a great job Meta employees when someone is shitting on Meta: yea ts tuff"
X Link 2025-08-29T14:34Z [----] followers, [----] engagements
"The worst thing about Meta-style layoffs is the feeling of uncertainty. You don't really know what to expect - either you get the layoff email or you don't and that determines at least the next couple of months of your life. You see some random subset of your colleagues disappear. You can't help but wonder - it could have been me. Why wasn't it me Was I not good enough If only they had picked me I would have gotten some nice severance and a vacation and then I'd come out on the other side working for leadership that actually respects me. Instead I'm still in the shareholder value mines"
X Link 2025-10-23T13:06Z [----] followers, 11.2K engagements
"life update: for those who dont know i didn't join @primeintellect but apparently they work on open source AGI. incredibly excited about what theyre building 🚀"
X Link 2025-08-18T07:34Z [----] followers, [----] engagements
"This is your daily reminder that PPO (GRPO TD3 .) is not a policy. It's an algorithm to train policies. If you make such a fundamental category error frankly you have no business writing about RL"
X Link 2025-10-26T15:08Z [----] followers, 11K engagements
"@tenobrus I interviewed and got rejected after like [--] rounds of interviews and a reference check so I imagine they're still busy hiring at this pace lol"
X Link 2025-10-05T09:02Z [----] followers, [----] engagements
"Hi guys I just wanted to say that Meta is still an incredible company. MSL has a tremendous potential and I am confident they will ship huge models the best models. They are the best guys around. I have full faith in Zuck's and Wang's leadership. Pic unrelated"
X Link 2025-10-29T20:57Z [----] followers, 14.3K engagements
"Just one more reorg bro please just trust me one more reorg and we're gonna be so efficient bro please"
X Link 2025-10-23T07:50Z [----] followers, [----] engagements
"@davepl1968 @jamanjeval Because you turn around that axis"
X Link 2025-02-18T21:38Z [----] followers, 18.6K engagements
"@tafphorisms First half of the video agreed. Second half she's getting very mildly unhinged"
X Link 2025-04-16T16:25Z [----] followers, [----] engagements
"Ok so the fp16 paper is nice and all just one question - it seems to show that bf16 GRPO runs pretty consistently collapse and fp16 is the savior who fixes it. .is this something that actually happens I get the occasional collapse but it's not super common even in bf16"
X Link 2025-10-31T16:46Z [----] followers, 15.4K engagements
"@yacineMTB @QiaochuYuan Sounds like skill issue tbh"
X Link 2025-04-18T03:13Z [----] followers, 12.8K engagements
"Someone has to explain this grift to me She didn't work at Meta didn't get laid off none of this is true. Is this some kinda stolen layoff valor - I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks ago - my dreams are about LLM alignment techniqes Still got laid off by Meta who hired a guy - I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks"
X Link 2025-10-24T07:07Z [----] followers, 20.4K engagements
"- I work on post-training and RL - I am an expert at the alphabet soup - DPO PPO GRPO - my papers are cited by all the OpenAI researchers - dropped a SOTA 10b LLM just a few weeks ago - my dreams are about LLM alignment techniqes Still got laid off by Meta who hired a guy with my same profile for $100M a year 😭😭"
X Link 2025-10-23T22:24Z 175.9K followers, 268.1K engagements
"@teknium Isn't GPQA a multiple choice question benchmark As in anyone can trivially get 100% pass@4"
X Link 2025-02-25T09:22Z [----] followers, [----] engagements
"Ok so if I quit Meta start a sweatshop and pretend that it's successful for like a year will Zuck rehire me with a 100M package"
X Link 2025-10-23T18:41Z [----] followers, [----] engagements
"@justalexoki Unironically a blow to the "He's just shitposting" theory"
X Link 2025-10-28T17:33Z [----] followers, [----] engagements
"@DissonanceCoder The irony of an "old school libertarian" saying this is palpable lmao"
X Link 2025-11-05T16:14Z [----] followers, [----] engagements
"Worth noting that neither will actually get you a job If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction"
X Link 2025-10-28T10:37Z [----] followers, 25.7K engagements
"If you can only learn two languages they should be: [--]. one of (Rust Zig) - will teach you to program at a lower level of abstraction more aligned with the underlying hardware [--]. one of (Haskell Scala OCaml) - will teach you to program at a higher level of abstraction If you can only learn two languages they should be: Rust TypeScript If you can only learn two languages they should be: Rust TypeScript"
X Link 2025-10-28T05:24Z 12.8K followers, 52.3K engagements
"They don't want you to know that but "peer review" in ML isn't peer review in the sense it was intended for normal science. The point of peer review is to check whether the claims in a paper are valid and backed by data. Peer review in ML is largely a noisy scarcity tactic"
X Link 2025-11-12T16:01Z [----] followers, [----] engagements
"@tiovikram @zebulgar Idk you can check my LinkedIn apparently that's what matters"
X Link 2025-04-21T02:13Z [----] followers, [----] engagements
"@real_bmoore no I'd rather minecraft myself than use java"
X Link 2025-11-12T22:06Z [----] followers, 48.7K engagements
"@francoisfleuret That's a fascinating observation What made you feel this way"
X Link 2025-04-12T07:08Z [----] followers, [----] engagements
"Ok not to be a hater but the $4.2M RL scaling paper seems to be a bit overhyped for what it is A little bit by the paper itself moreso by twitter poasters. From an initial reading it seems like yet another set of tweaks to GRPO except this time it's trained on different compute budgets but - crucially - only on relatively small models (Llama [--] 8B and Llama [--] Scout) and one dataset that's 100% math questions. The main novelty is that they fitted a curve to the reward graph which is uh cool I guess The cherry on top is the code repo which is one file centered around from scipy.optimize import"
X Link 2025-10-18T22:22Z [----] followers, 43.6K engagements
"I'm not doing LLMs because I want funding lmao I'm doing LLMs because I want a magic superintelligence in the sky and nothing else comes even remotely close right now"
X Link 2025-10-28T15:08Z [----] followers, 17.3K engagements
"@ludwigABAP If you still keep asking "why" you end up in an insane asylum"
X Link 2025-11-05T13:40Z [----] followers, [----] engagements
"@stanislavfort Genuine skill issue. Many people don't know how LLMs work and how to use them. Instead of learning to use a screwdriver they keep using it as a hammer"
X Link 2025-03-27T10:45Z [----] followers, [----] engagements
"Readding 4o but not o3 is blatant pandering to normies and I'm not happy about it"
X Link 2025-08-09T09:20Z [----] followers, [----] engagements
"The funny thing about the job market in France is that if the recruiter messages me in French I just know it's gonna be a huge lowball. Not to toot my own GPU but what about my profile suggests that "up to 100k" is a super attractive offer"
X Link 2025-10-30T18:42Z [----] followers, 45K engagements
"The fuck is unc talking about talking points lifted straight from r*ddit. Spark and Tinybox aren't even the same class of device. Spark is for local prototyping a devkit. Tinybox is meant to fully sustain your AI waifu with 24/7 inference. Completely different target audiences There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by There's a whole bunch of people who talk in"
X Link 2025-11-05T20:56Z [----] followers, [----] engagements
"There's a whole bunch of people who talk in this space who don't understand it. If you want to run your moderately large LLM at [--] tok/s buy a Mac Studio or DGX Spark with 128GB of RAM. Congrats you are an AI influencer Then when you turn the camera off you get frustrated by the slow speeds and low quality outputs and you end up back using ChatGPT. Don't worry I won't tell. I understand few have had to think about RAM bandwidth before when looking at computers but it's the main thing that determines the speed of your LLMs. A tinybox pro has [--] TB/s of RAM bandwidth equivalent to [--] GB300s ($80k"
X Link 2025-11-05T20:14Z 63.8K followers, 54.3K engagements
"@jxmnop "I regret to inform you that using a pre-built tool is easy" Ok then"
X Link 2025-05-27T22:52Z [----] followers, [----] engagements
"I felt a bit bad dunking on him but not anymore. His question was answered by many people. He just chooses to not listen to them to feel superior. learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun"
X Link 2025-11-04T13:58Z [----] followers, 31.7K engagements
"learned things from some of the replies but nobody answered the question I didnt know how to build a fast JavaScript runtime transpiler or bundler before Bun Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to infinite money and in a world w/ 1/10000th of the capital models would be orders of magnitude better Can someone explain to me why [---] tok/s is fast and what in-the-weeds technical constraints prevent [------] tok/s at same quality My gut is theres incredible waste due to"
X Link 2025-11-04T12:52Z 156.2K followers, 91.2K engagements
"@MartinShkreli print("7 11")"
X Link 2025-01-10T13:51Z [----] followers, [----] engagements
"@aidan_mclau TIL Aidan is a p-zombie"
X Link 2025-10-29T18:31Z [----] followers, [----] engagements
"@hallerite It's been a while but I'm like 90% sure we covered the definition of a limit in high school in Poland. We might be built different though"
X Link 2025-11-09T16:58Z [----] followers, 11.6K engagements
"@RikoSuminoe69 I know it's a bit of a shitpost"
X Link 2025-10-22T16:32Z [----] followers, [----] engagements
"@qtnx_ My favorite thing is when people pu "researcher @ fancy lab" I look them up and it was a summer internship lol"
X Link 2025-05-27T10:02Z [----] followers, [----] engagements
"@rational_wiki To determine whether they're actually illegal due process is necessary"
X Link 2025-04-16T13:55Z [----] followers, [----] engagements
"@Duderichy Did the [----] graduating class even graduate yet It's still [----] right Right"
X Link 2025-06-25T18:22Z [----] followers, [----] engagements
"@pmehta94 it's a sign of being a js monkey and not actually doing AI"
X Link 2025-11-12T22:06Z [----] followers, 43.4K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing