# ![@dhtikna Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::935351575903576065.png) @dhtikna Ankith Devanaboina

Ankith Devanaboina posts on X about xai, open ai, deepseek, grok the most. They currently have [-----] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours.

### Engagements: [------] [#](/creator/twitter::935351575903576065/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::935351575903576065/c:line/m:interactions.svg)


### Mentions: [--] [#](/creator/twitter::935351575903576065/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::935351575903576065/c:line/m:posts_active.svg)


### Followers: [-----] [#](/creator/twitter::935351575903576065/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::935351575903576065/c:line/m:followers.svg)


### CreatorRank: [-------] [#](/creator/twitter::935351575903576065/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::935351575903576065/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  [finance](/list/finance)  [countries](/list/countries)  [stocks](/list/stocks)  [currencies](/list/currencies)  [celebrities](/list/celebrities)  [social networks](/list/social-networks)  [cryptocurrencies](/list/cryptocurrencies)  [automotive brands](/list/automotive-brands)  [gaming](/list/gaming) 

**Social topic influence**
[xai](/topic/xai) #224, [open ai](/topic/open-ai), [deepseek](/topic/deepseek), [grok](/topic/grok), [bytedance](/topic/bytedance), [china](/topic/china), [elon musk](/topic/elon-musk), [this is](/topic/this-is), [level](/topic/level), [$googl](/topic/$googl) #3163

**Top accounts mentioned or mentioned by**
[@teortaxestex](/creator/undefined) [@grok](/creator/undefined) [@zephyrz9](/creator/undefined) [@deepseekai](/creator/undefined) [@adonissingh](/creator/undefined) [@openai](/creator/undefined) [@xai](/creator/undefined) [@apathium65906](/creator/undefined) [@kimmonismus](/creator/undefined) [@ns123abc](/creator/undefined) [@presidentlin](/creator/undefined) [@danielmac8](/creator/undefined) [@elonmusk](/creator/undefined) [@teknium1](/creator/undefined) [@openrouterai](/creator/undefined) [@flowersslop](/creator/undefined) [@xjdr](/creator/undefined) [@lunayian](/creator/undefined) [@grad62304977](/creator/undefined) [@sama](/creator/undefined)

**Top assets mentioned**
[DeepSeek (DEEPSEEK)](/topic/deepseek) [GrokCoin (GROKCOIN)](/topic/grok) [Alphabet Inc Class A (GOOGL)](/topic/$googl) [Bitcoin (BTC)](/topic/bitcoin) [Tesla, Inc. (TSLA)](/topic/tesla)
### Top Social Posts
Top posts by engagements in the last [--] hours

"@tsoding I thought redis doesn't need sync primitives like mutexes as it was single threaded"  
[X Link](https://x.com/dhtikna/status/1718281865126367499)  2023-10-28T15:01Z [--] followers, [----] engagements


"@Teknium1 What a chad going straight for copyrighted material"  
[X Link](https://x.com/dhtikna/status/1737060677389570392)  2023-12-19T10:41Z [--] followers, [--] engagements


"@literallydenis Its confirmed to be llama architecture and not a MoE although its possible that unlike 7B and mixtral the medium model is actually a LLama 70B Finetune. @main_horse actually thinks this. Another reason why it could be medium is that its waaay too close to medium for a normal FT"  
[X Link](https://x.com/dhtikna/status/1752321203459825921)  2024-01-30T13:21Z [--] followers, [---] engagements


"@JonathanRoss321 @GroqInc [---] tok/s * 2.5e-7 usd/tok = 1.05e-4 usd per second = [---] usd per month assuming 100% usage. Lets say 'pipelining' optimistically lets them run [--] prompts at once so [----] usd per model. If they need [---] lpus [--] bucks revenue per lpu per month power usage must cost more"  
[X Link](https://x.com/dhtikna/status/1760286663396483201)  2024-02-21T12:53Z [---] followers, [---] engagements


"@StabilityAI Dall-E bing same prompt"  
[X Link](https://x.com/dhtikna/status/1760683361595695455)  2024-02-22T15:09Z [--] followers, [---] engagements


"@jeremyphoward Or try pip install vdom Now sure if the above is overkill"  
[X Link](https://x.com/dhtikna/status/1779204292617888234)  2024-04-13T17:45Z [--] followers, [--] engagements


"@0xBADC0DEC @nonmayorpete Was this post generated by an AI Grok has some of the worst Perf/Size I think LLama3 8B is much better afaik. Doubt Grok [---] is any smaller"  
[X Link](https://x.com/dhtikna/status/1783821897123168335)  2024-04-26T11:34Z [--] followers, [--] engagements


"@sughanthans1 @btibor91 Seems like Tibor is literally just using Chrome Inspect. He makes an extension thats like a GreaseMonkey script on chatgptcom to make you more productive with chatgpt. so its literally his job to go through this stuff very surprised that no one else did this before him"  
[X Link](https://x.com/dhtikna/status/1786340884369314129)  2024-05-03T10:23Z [---] followers, [--] engagements


"Think james cameroon casting sam worthington when matt damon turned down the role for Avatar. The voice is not scarjo its just similar almost certain openai could have easily made it exactly like scrajo's but did not. Most likely they just hired someone else whose voice was similar. "her" tweet was a reference to the AI product being similar to the movie's; It's not referencing scarjo exclusively"  
[X Link](https://x.com/dhtikna/status/1792842344380649798)  2024-05-21T08:58Z [--] followers, [---] engagements


"@OpenAIDevs Why can't I just use chatgpt website but instead of plus or free tier I just add my OpenAI API key to get much better rate limits I don't mind spending a 100$ a month but why force me to go to a 3rd party inference website when I can just do it in chatgpt website"  
[X Link](https://x.com/dhtikna/status/1792850502801031546)  2024-05-21T09:30Z [--] followers, [--] engagements


"@jarredsumner So instead of having to remember to use Promise.all you have to remember to use batch Am I missing something Is batch more lenient like able to mix promises and sync calls in the arguments"  
[X Link](https://x.com/dhtikna/status/1795039438621290848)  2024-05-27T10:28Z [--] followers, [---] engagements


"@jarredsumner Ohhh nevermind misread it You don't even need to pass in each call as an argument It just takes one lambda. This is cool"  
[X Link](https://x.com/dhtikna/status/1795040388354290030)  2024-05-27T10:32Z [--] followers, [--] engagements


"@mreflow Everything incl entertainment on 2x lectures on 2.5x but @fireship_dev on 1x because its one of the few *actually* information dense content out there"  
[X Link](https://x.com/dhtikna/status/1795358343055466882)  2024-05-28T07:35Z [--] followers, [--] engagements


"@Heaney555 @Lunayian Wheres the evidence for that Seems like a collaboration but I can see immersed handling all then software while Qualcomm does the hardware. But if that's the case why would Qualcomm work with immersed rather than another more well established company"  
[X Link](https://x.com/dhtikna/status/1795869575172473326)  2024-05-29T17:27Z [--] followers, [--] engagements


"@DeedleFake @IroncladDev Why. Children are basically just an attribute"  
[X Link](https://x.com/dhtikna/status/1797338035060674948)  2024-06-02T18:42Z [--] followers, [---] engagements


"@huamichaelchen @realGeorgeHotz @__tinygrad__ @AMD @nvidia AMD has 33% less flops than NVIDIA and NVIDIA is 25% faster in this benchmark"  
[X Link](https://x.com/dhtikna/status/1801540936133972463)  2024-06-14T09:03Z [--] followers, [--] engagements


"@MyEyesBlink @benhylak I do not expect openai models to run on Apples new cloud any reference that says so Only apple's own large models run in their secured cloud and any openai calls do not go to their newly designed apple silicon datacenters"  
[X Link](https://x.com/dhtikna/status/1801632016560996556)  2024-06-14T15:05Z [--] followers, [---] engagements


"@theramblingfool Came here from the prez official acts immunity thread not disappointed consider pinning"  
[X Link](https://x.com/dhtikna/status/1808144594603069874)  2024-07-02T14:23Z [--] followers, [--] engagements


"@amebagpt @aidan_mclau Yeah quite a blunder by Meta"  
[X Link](https://x.com/dhtikna/status/1818562356642775327)  2024-07-31T08:20Z [--] followers, [--] engagements


"@lmsysorg @GoogleDeepMind Lol #1 overall is useless if youre not #1 in math coding AND hard. They just markdown hack and normie optimize. Hope these clowns dont get smoked by openai after the US elections"  
[X Link](https://x.com/dhtikna/status/1819079755393913229)  2024-08-01T18:36Z [--] followers, [----] engagements


"@zeta_globin "So I hated all the fruit of my labor for which I had labored under the sun because I must leave it to the man who will come after me. And who knows whether he will be wise or a fool . There is nothing better for a person than to eat and drink himself" - Ecclesiastes 2"  
[X Link](https://x.com/dhtikna/status/1822879845967147074)  2024-08-12T06:16Z [--] followers, [--] engagements


"@graemehein @kalajhadu @MattZeitlin No they can't. Most IIT grads can 3x-4x their comp (thats only if you include equity). The one that gets 30k perhaps can get 100k-120k in US but the one thats getting 100k in india can get 300k-350k in US/EU. You shouldn't compare 1st guys IND comp w/ 2nd guys US comp"  
[X Link](https://x.com/dhtikna/status/1825598260150677960)  2024-08-19T18:18Z [--] followers, [--] engagements


"@CuriousPejjy Id expect them to set a price something like volume weighted close price over the year or median close price and you would pay capital gains once every year or perhaps every decade"  
[X Link](https://x.com/dhtikna/status/1826254380854829371)  2024-08-21T13:45Z [--] followers, [--] engagements


"@AnthonyDaytv @tomwarren Assuming MSFT cares about xbox console sales. Consoles are sold at cost but I do agree that consoles generate more revenue than pc users so I think MSFT will regret this strategy in the long run"  
[X Link](https://x.com/dhtikna/status/1826269931224334760)  2024-08-21T14:47Z [--] followers, [----] engagements


"@uwucomputer @SadlyItsBradley Exactly my thoughts 4x weaker than M chips"  
[X Link](https://x.com/dhtikna/status/1828359362760355958)  2024-08-27T09:10Z [--] followers, [---] engagements


"@__tinygrad__ Is this for comma"  
[X Link](https://x.com/dhtikna/status/1828659665447591944)  2024-08-28T05:03Z [--] followers, [----] engagements


"@lmsysorg Huh what if grok [--] is fine tuned L3-405B. Does meta's license allow that"  
[X Link](https://x.com/dhtikna/status/1829434120625172611)  2024-08-30T08:20Z [--] followers, [---] engagements


"@aidan_mclau Get the full logits full access to inference code so you can experiment with CoT MCTS at a deeper level than final tokens"  
[X Link](https://x.com/dhtikna/status/1829856172926779425)  2024-08-31T12:18Z [--] followers, [--] engagements


"@androiddevnotes @itsandrewgao @OpenAI @androiddevnotes whats the source for this pic Thanks"  
[X Link](https://x.com/dhtikna/status/1834374606863352144)  2024-09-12T23:32Z [--] followers, [--] engagements


"Honestly i dont buy tree search or even any tree of thought is happening. It is just a well implemented CoT but i do think there are [--] models one that generates reasoning given input and another which generates summary/output given input+reasoning. Key is Self Taught Learning (STaR) for synthethic data gen to produce the data that reasoner is fine tuned on"  
[X Link](https://x.com/dhtikna/status/1834478002228387939)  2024-09-13T06:23Z [--] followers, [---] engagements


"@aidan_mclau @_clashluke [----] usd over [--] days non-stop or $0 over [--] days non stop w/ free tier"  
[X Link](https://x.com/dhtikna/status/1835969538246000892)  2024-09-17T09:10Z [--] followers, [--] engagements


"@codingquark @__tinygrad__ I don't think you can run 405B at any quant on a tiny box"  
[X Link](https://x.com/dhtikna/status/1836274892808901060)  2024-09-18T05:23Z [--] followers, [--] engagements


"@James_Abev its not false advertising It DOES ship with an 'Oakley-like' form factor but its just inferior to the strap. I bet ton of people would be fine with the glasses form factor it is obvious that such a form factor would never be more comfy than a normal strap"  
[X Link](https://x.com/dhtikna/status/1837105347468689684)  2024-09-20T12:23Z [--] followers, [--] engagements


"@Moh1tAgarwal @jackgwhitaker Did Meta connect cause this cause of orion"  
[X Link](https://x.com/dhtikna/status/1839056989714002072)  2024-09-25T21:38Z [--] followers, [--] engagements


"@halr9000 @theXRman @Tyrielwood @VisorXR @RenjiBijoy I sure hope so but I feel a few bad apples will use this as an oppurtunity to pay the [---] usd and hack out the subscription requirements"  
[X Link](https://x.com/dhtikna/status/1839624807412912215)  2024-09-27T11:15Z [--] followers, [--] engagements


"@V3nti1at0r @LoonchieBey @wccftech Affordability is necessary but not sufficient to make a purchase. It has to be worth the price. I can afford to pay [----] too but it would be a waste on a console"  
[X Link](https://x.com/dhtikna/status/1840283655882608743)  2024-09-29T06:53Z [--] followers, [--] engagements


"@cosmoscharf You hate Orion cause its 10k usd. I hate orion because it'll be a while until 4k [--] fov and who knows if it will ever get a good dynamic range. I think there will always be a vr headset better than any ar glasses in all areas except (negligiblly) comfort https://x.com/PalmerLuckey/status/1840624927763763381 All the people whining about Orion prototypes costing $10k would be very surprised to find out how much Crescent Bay - hint it was several times more expensive and getting it from there to $599 consumer product only took a year or so."  
[X Link](https://x.com/dhtikna/status/1840686787942850787)  2024-09-30T09:35Z [--] followers, [---] engagements


"@mark_k @OpenAI Theres a limit"  
[X Link](https://x.com/dhtikna/status/1840832733418750235)  2024-09-30T19:14Z [--] followers, [---] engagements


"@shrihacker Honestly it should just be system design and LLD interviews based on small real world problems that the team has already solved or needs to solve"  
[X Link](https://x.com/dhtikna/status/1842820128070647956)  2024-10-06T06:52Z [--] followers, [--] engagements


"@LechMazur @elidourado Who is giving you a 48% yoy interest rate"  
[X Link](https://x.com/dhtikna/status/1843989036287009094)  2024-10-09T12:17Z [--] followers, [--] engagements


"@teslaownersSV @Tesla_Optimus why do they pretend to not be teleoperated Its almost like fraud tbh"  
[X Link](https://x.com/dhtikna/status/1844719890608832585)  2024-10-11T12:41Z [--] followers, [---] engagements


"@OnePlus_USA Peak of smartphone design. I really miss mine used it for almost [--] years before the battery puffed up. RIP king"  
[X Link](https://x.com/dhtikna/status/1844754095761977564)  2024-10-11T14:57Z [--] followers, [--] engagements


"@Lunayian @NathieVR Man [--] years of no releases is gonna be tough. What headsets are you hyped for releasing in [----] Im not that excited for the samsung/google and avp next because i dont have high expectations for their form factor"  
[X Link](https://x.com/dhtikna/status/1846737823682252884)  2024-10-17T02:19Z [---] followers, [--] engagements


"@TJEvarts @__tinygrad__ Would just be cheaper to buy Qualcomm oryon chips directly. Im sure 7900xt is more than 2x better and not 2x more expensive. Although you get more CPU compute too per [---] usd"  
[X Link](https://x.com/dhtikna/status/1851579680039817267)  2024-10-30T10:59Z [---] followers, [--] engagements


"@BattleAxeVR I guess you can make decisons like clock speed on the chipsets dont need to test out the perf and fine tune the OS for 100s of games but just [--] apps like immeresed browser etc. Hand tracking perf eye tracking perf and passthru perf @ panel UI/UX level and not beat-saber"  
[X Link](https://x.com/dhtikna/status/1854484256581714301)  2024-11-07T11:21Z [--] followers, [--] engagements


"@alexandrosM Its just momentum on Trump friendly investments like TSLA and BTC. Perhaps some shorts being pressured too"  
[X Link](https://x.com/dhtikna/status/1856320862409887939)  2024-11-12T12:59Z [--] followers, [--] engagements


"@SawyerMerritt @Polymarket bet incoming on fatality rate for this"  
[X Link](https://x.com/dhtikna/status/1856992261235409350)  2024-11-14T09:27Z [--] followers, [---] engagements


"@Yampeleg The real Gigachad move is to submit the winning arc eval submission on the very last day as an announcement for O2 / GPT 5"  
[X Link](https://x.com/dhtikna/status/1856994801729900678)  2024-11-14T09:37Z [--] followers, [---] engagements


"@pastelETH @kusiekG To maintain ketosis sometimes carbs are just added for variety almost as a 'dessert' because they are not as satieting as protien"  
[X Link](https://x.com/dhtikna/status/1857069985874559447)  2024-11-14T14:36Z [--] followers, [----] engagements


"@iota_male No something else likely https://x.com/AndrewCurran_/status/1858010360122417401 Nothing confirmed but rumors tonight that Grok [--] suffered some sort of catastrophic incident yesterday during its training. Hopefully not true. There were similar rumors about the Opus [---] delay nothing was ever confirmed there either https://x.com/AndrewCurran_/status/1858010360122417401 Nothing confirmed but rumors tonight that Grok [--] suffered some sort of catastrophic incident yesterday during its training. Hopefully not true. There were similar rumors about the Opus [---] delay nothing was ever"  
[X Link](https://x.com/dhtikna/status/1858014105396707598)  2024-11-17T05:07Z [--] followers, [---] engagements


"@bryan_johnson @hpotech_medical Whats the risk of turning the chamber into a furnace due to slight ignition inside the chamber"  
[X Link](https://x.com/dhtikna/status/1861110796777918829)  2024-11-25T18:12Z [--] followers, [---] engagements


"@tuseroni @SCHIZO_FREQ I was using it in a colloquial way to imply just like usd btc has no underlying value on it owns except for trust in the authority backing the currency (bitcoin users in this case)"  
[X Link](https://x.com/dhtikna/status/1864669605773627772)  2024-12-05T13:54Z [--] followers, [--] engagements


"@safwaankay @AnthropicAI 1)Massively productionize computer use incl android and ios app. 2)Launch a homework helper / tutor subpage on claude. FT a haiku on it and make it unlimited but have an upsell. Something like celebrity tie in lectures. Eminem teaches English hiesenberg teaches chem"  
[X Link](https://x.com/dhtikna/status/1865127841639100599)  2024-12-06T20:15Z [--] followers, [--] engagements


"@frustum can you give some stats res cpu+gpu. Is this like fully path traced or something Whats special about this if you dont mind me asking"  
[X Link](https://x.com/dhtikna/status/1866074365101810085)  2024-12-09T10:56Z [--] followers, [---] engagements


"@BitcoinIsaiah @RrnRaindeer @cz_binance you literally proved his point that its about consensus not math"  
[X Link](https://x.com/dhtikna/status/1866764265228816406)  2024-12-11T08:37Z [--] followers, [--] engagements


"@slow_developer What are you talking about claude opus [--] and sonnet [---] have held best coding model for almost the year. Cheap models never had clear cut domination b/w the big [--]. Ofc if anthropic lags in releasing multimodal soon then you can say they fell behind but way [--] soon [--] call"  
[X Link](https://x.com/dhtikna/status/1866989850530062790)  2024-12-11T23:34Z [--] followers, [---] engagements


"@AutismCapital How long has this been there Ive been waiting months for grok analysis"  
[X Link](https://x.com/dhtikna/status/1867196700298981703)  2024-12-12T13:15Z [--] followers, [--] engagements


"Its a cost benefit analysis. How much you gain from automation for the country minus how much you lose for these replaced by automation. Non trivial calculation in the short term probably always net positive in very long term. Trump is not ideologically pure e/acc but being pragmatic which is good imo"  
[X Link](https://x.com/dhtikna/status/1867539495190770067)  2024-12-13T11:58Z [--] followers, [--] engagements


"@VictorTaelin I mean it is free and experimental even google doesn't have infinite compute. They should just release a paid app and allocate proper compute for it"  
[X Link](https://x.com/dhtikna/status/1867586767110779018)  2024-12-13T15:05Z [--] followers, [---] engagements


"@abacaj Counter point: Gemini [---] flash its not better but it is oom cheaper. Hopes of the wall having broken are slowly renewing"  
[X Link](https://x.com/dhtikna/status/1867597726889783470)  2024-12-13T15:49Z [--] followers, [---] engagements


"@omouamoua @scaling01 Oh ok i thought tokenization was like the bane of llms. Also byte [--] byte means its inherently multi modal with all modalities"  
[X Link](https://x.com/dhtikna/status/1867657327861764364)  2024-12-13T19:46Z [---] followers, [--] engagements


"@alexocheema Yappi handles async cpu time correctly. Most profilers include time taken by await in the parent func but yappi is able to exclude it. Not sure about gpu but if the interface is ordinary async await then it should work"  
[X Link](https://x.com/dhtikna/status/1868261747393474930)  2024-12-15T11:48Z [--] followers, [--] engagements


"@LuciferMornens @kimmonismus Youve tried [----] already"  
[X Link](https://x.com/dhtikna/status/1868262483007594565)  2024-12-15T11:51Z [--] followers, [--] engagements


"@0xIlyy Not missing much. Its actually much more accurate than google ai overview its free relatively fast and actually uses search unlike chatgpt. Terrible multi turn perf tho grok is better overall except you cant add it as custom search engine unlike pplx"  
[X Link](https://x.com/dhtikna/status/1868899416373182685)  2024-12-17T06:01Z [--] followers, [---] engagements


"@menhguin Dayum nice what model is this"  
[X Link](https://x.com/dhtikna/status/1872904187929129036)  2024-12-28T07:15Z [--] followers, [--] engagements


"@MRmor44 @lmarena_ai @OpenAI @deepseek_ai no normie prefrences are just mid"  
[X Link](https://x.com/dhtikna/status/1873705035944845726)  2024-12-30T12:17Z [--] followers, [---] engagements


"@GamingKarmig @wccftech Steamdeck has 16gb ram and pc eith [----] 6gb also has 8gb+ ram which makes total memory 12gb+"  
[X Link](https://x.com/dhtikna/status/1874824275376623679)  2025-01-02T14:25Z [--] followers, [--] engagements


"@JamesEdward90_ @SenSanders O1 and EB1 are for researches so someone with only an undergrad that got selected for 500k usd salary wouldn't qualify even though obviously such a person is 'exceptionall talented'"  
[X Link](https://x.com/dhtikna/status/1874950037245194543)  2025-01-02T22:44Z [--] followers, [--] engagements


"@MarmotRespecter @ctjlewis @cremieuxrecueil Which h1bs can also do. Their risk is layoffs but nothing stopping them from switching other than companies prefer natives over h1b when all else equal cause of extra (filing) costs and compliance"  
[X Link](https://x.com/dhtikna/status/1875023847864639643)  2025-01-03T03:38Z [--] followers, [--] engagements


"The goal for a hedge fund is "Diversification is the only free lunch in finance". The ideal return is a risk free return like plain debt. So a hedge fund that under-performs the snp [---] by a bit is totally fine as long as its return are uncorrelated so that at any point your portfolio's risk looks closer to ordinary debt. Plus hedge funds are suppose to hedge and help curb volatility too"  
[X Link](https://x.com/dhtikna/status/1875246957842854192)  2025-01-03T18:24Z [--] followers, [----] engagements


"@alexocheema @exolabs I meant as you grow the db you have grow your cluster. So for any finite cluster its practically O(N). Although aws literally lets you scale cluster so my point is moot"  
[X Link](https://x.com/dhtikna/status/1875267616438394932)  2025-01-03T19:46Z [--] followers, [--] engagements


"@SadlyItsBradley will we be getting a CES [----] video"  
[X Link](https://x.com/dhtikna/status/1877613249505485187)  2025-01-10T07:07Z [--] followers, [--] engagements


"@kimmonismus I think it really is Q*. Q* training doesnt need new data it is purely synthetic data driven and in [--] months took O1 to O3 and OAI suspect it wont stop with that. It may be just a matter of time before we get ASI"  
[X Link](https://x.com/dhtikna/status/1878025362031988923)  2025-01-11T10:25Z [--] followers, [---] engagements


"@reach_vb very interesting that for 72B the RM often beat PRM"  
[X Link](https://x.com/dhtikna/status/1879118167085355049)  2025-01-14T10:47Z [---] followers, [---] engagements


"@realGeorgeHotz Why should US allow china to sell to us when US cant sell to China"  
[X Link](https://x.com/dhtikna/status/1879393104891093290)  2025-01-15T05:00Z [--] followers, [---] engagements


"@StatisticUrban Source on how this was calculated Its weird thats neither trending up or down. Seems fishy"  
[X Link](https://x.com/dhtikna/status/1879447962771947775)  2025-01-15T08:38Z [--] followers, [---] engagements


"@ajshaju @realDonaldTrump @DavidSacks @JDVance @elonmusk @DOGE No one is entitled to citizenship in another country"  
[X Link](https://x.com/dhtikna/status/1879880864349524365)  2025-01-16T13:18Z [--] followers, [--] engagements


"@Teknium1 Now imagine pass@10 since that costs just a little more than O1 ($20 vs $15 usd per million tokens)"  
[X Link](https://x.com/dhtikna/status/1881317488526499929)  2025-01-20T12:26Z [--] followers, [---] engagements


"@Teknium1 Oh they didnt use their recently open sourced PRM"  
[X Link](https://x.com/dhtikna/status/1881327479077700016)  2025-01-20T13:06Z [--] followers, [--] engagements


"It doesnt sit between mini and O1. It IS O1 level according to benches. 50% it benches higher and other time O1 benches higher. Their distillation to Qween 7B is O1 mini level and Qween 14B smashes O1 mini across the board They did phenomenally well. Lets see if private evals and vibes also indicate its O1 level"  
[X Link](https://x.com/dhtikna/status/1881410580873142488)  2025-01-20T18:36Z [--] followers, [--] engagements


"@AnnaGorisch Just apply for H2E2 or what ever dependant visa during pregnancy. EO will be struck down but birthright citizenship is definitely a bad policy children should inherit parents visa equivalent"  
[X Link](https://x.com/dhtikna/status/1881727809837830551)  2025-01-21T15:37Z [--] followers, [---] engagements


"@DodgeHealer @btibor91 Codeforces was tie. [----] vs [----] right Aider polyglot is the only reasoning bench where o1 was much better (61.7 vs 53.3)"  
[X Link](https://x.com/dhtikna/status/1881740414832750592)  2025-01-21T16:27Z [--] followers, [--] engagements


"@kalomaze @vikhyatk This is mcts overthinking all over again. O3 = O1 plus the same reasoning RL applied for several months"  
[X Link](https://x.com/dhtikna/status/1881990463575789782)  2025-01-22T09:01Z [--] followers, [---] engagements


"@cognitivecompai @deepseek_ai they said in the paper its straightforward SFT. Simple fine tuning they even mentioned its 600k+200k sampels. Literally no magic. SFT distillation is unreasonably good"  
[X Link](https://x.com/dhtikna/status/1882058465998487602)  2025-01-22T13:31Z [--] followers, [---] engagements


"@stone_toss and grok if i want internet search http://Chat.deepseek.com http://Chat.deepseek.com"  
[X Link](https://x.com/dhtikna/status/1882334759524917492)  2025-01-23T07:49Z [--] followers, [--] engagements


"@_jaydeepkarale Quants at hedge funds hfts still use R regularly although then move to python is being made gradually"  
[X Link](https://x.com/dhtikna/status/1882347317933851074)  2025-01-23T08:39Z [--] followers, [---] engagements


"@repligate Is there like a comprehensive map of llm personalities Something visual Something you can read in [--] mins and get a gist of current landscape"  
[X Link](https://x.com/dhtikna/status/1882679072545030294)  2025-01-24T06:37Z [--] followers, [--] engagements


"@hoytehablode Check out grok doing the same for [--] series and amd's 600xt series equivalent https://x.com/i/grok/share/o2S4gWQtTHQA7rFgXQOXefTQY https://x.com/i/grok/share/o2S4gWQtTHQA7rFgXQOXefTQY"  
[X Link](https://x.com/dhtikna/status/1882712447804674450)  2025-01-24T08:50Z [--] followers, [--] engagements


"This is the most impressive long-form instruct following + search i've ever seen from any LLM Is this Grok [--] or did I secretly get Grok [--] in a/b testing I had it create tables for AMD & Nvidia GPUs and compare the perf and price increase gen-over-gen https://x.com/i/grok/share/gKpWTorjXI043D8DWD3nxi3Cu https://x.com/i/grok/share/gKpWTorjXI043D8DWD3nxi3Cu"  
[X Link](https://x.com/dhtikna/status/1882715642455109668)  2025-01-24T09:02Z [--] followers, [--] engagements


"@angelusm0rt1s nah o3 mini with 4o level of rate limits might make more people buy it and atelast it will keep people on after adding grok as a search engine shortcut in chrome and R1. I barely use any more. Might consider again based on o3 perf http://chatgpt.com http://chatgpt.com http://chatgpt.com http://chatgpt.com"  
[X Link](https://x.com/dhtikna/status/1882726959656153587)  2025-01-24T09:47Z [--] followers, [--] engagements


"@idontexist_nn Deepseek making OAI become more Open Thanks Liang"  
[X Link](https://x.com/dhtikna/status/1882760055420064089)  2025-01-24T11:59Z [--] followers, [---] engagements


"@idontexist_nn @Neuralithic How are O1 and O3 not adaptive already"  
[X Link](https://x.com/dhtikna/status/1882760716689743968)  2025-01-24T12:01Z [--] followers, [---] engagements


"@AravSrinivas For sure lot of discovery and innovation done by deepseek (mcts is bad prm is bad grpo guess and check reward but also add accuracy in reward calc)"  
[X Link](https://x.com/dhtikna/status/1882766523158470840)  2025-01-24T12:24Z [--] followers, [---] engagements


"Its just slower plus OP only RL on one specific problem and not on reasoning overall so it may neede.less steps. R1 paper said sft distillation is much better than rl for smaller models but also explicitly said further rl training would boost perf even more on top of sft distillation. They even did it for qween 32B but didnt mention any more details"  
[X Link](https://x.com/dhtikna/status/1882867052769366024)  2025-01-24T19:04Z [---] followers, [---] engagements


"@mynamebedan If you do a little math you can show deepseek api is the most profitable out of all in open router https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api has yet to replicate all inference optimizations of V3 and R1 (this also reflects in api speed if you need more proof). Remember tokens per second is inversely proportional to inference cost since the only really cost is power and captial which are constant. [---] tps costs https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api has yet to replicate all inference optimizations of V3 and R1 (this also"  
[X Link](https://x.com/dhtikna/status/1882885755061907699)  2025-01-24T20:18Z [--] followers, [--] engagements


"@teortaxesTex @mynamebedan Actually deepseek is just passing their costs savings made through throuput optimization down to customers and still making more money than together and fireworks https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api has yet to replicate all inference optimizations of V3 and R1 (this also reflects in api speed if you need more proof). Remember tokens per second is inversely proportional to inference cost since the only really cost is power and captial which are constant. [---] tps costs https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api"  
[X Link](https://x.com/dhtikna/status/1882886017541365876)  2025-01-24T20:19Z [--] followers, [--] engagements


"@Teknium1 @nealkhosla Their R1 end point is more profitable than fireworks and together too. If deepseek is subsidied than american ones are even more https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api has yet to replicate all inference optimizations of V3 and R1 (this also reflects in api speed if you need more proof). Remember tokens per second is inversely proportional to inference cost since the only really cost is power and captial which are constant. [---] tps costs https://x.com/dhtikna/status/1882884696729858300 @nealkhosla No api has yet to replicate all inference"  
[X Link](https://x.com/dhtikna/status/1882891563229126805)  2025-01-24T20:41Z [--] followers, [---] engagements


"@teortaxesTex I thought llamas underperform because they use offline post training (L3 is online tho)"  
[X Link](https://x.com/dhtikna/status/1882929716958822704)  2025-01-24T23:13Z [--] followers, [---] engagements


"@lmarena_ai @deepseek_ai Wow did not at all expect that. Thought it would tie o1 mini. This crazy"  
[X Link](https://x.com/dhtikna/status/1882945309430763522)  2025-01-25T00:15Z [--] followers, [---] engagements


"@apples_jimmy Btw 4o was a failure at demonstrating transfer between modalities right Can someone confirm"  
[X Link](https://x.com/dhtikna/status/1883047480243433714)  2025-01-25T07:01Z [---] followers, [---] engagements


"@chrisprucha You should try grok for search its faster and just as accurate imo its been my goto. Plus no need for extra button press to enable web search its always on"  
[X Link](https://x.com/dhtikna/status/1883197134637552047)  2025-01-25T16:55Z [---] followers, [--] engagements


"Deepseek claims they wont go closed source like mistral and no "ipo" soon"  
[X Link](https://x.com/dhtikna/status/1883210618439930095)  2025-01-25T17:49Z [---] followers, [--] engagements


"@NascentClouds Idk the pessimist in me says if they release a paper before going all in on it and scaling then its not a real breakthrough. If titans was annouced next to Titan 1T then i would be excited. I doubt google or anyone would release their golden goose so early"  
[X Link](https://x.com/dhtikna/status/1883449731973779870)  2025-01-26T09:39Z [--] followers, [--] engagements


"@teortaxesTex Deepseek comrade otw to kill oai operator"  
[X Link](https://x.com/dhtikna/status/1883572922964754458)  2025-01-26T17:49Z [---] followers, [---] engagements


"@GrantSlatton That would have been a flex i mean april [--] sounds like a great day to release R2"  
[X Link](https://x.com/dhtikna/status/1883998725313552865)  2025-01-27T22:01Z [---] followers, [--] engagements


"@Grad62304977 @sybilhyz Omg you got more deepseek lore to drop"  
[X Link](https://x.com/dhtikna/status/1884376998711287922)  2025-01-28T23:04Z [---] followers, [--] engagements


"@AravSrinivas Tok/sec is vital is it atleast 10"  
[X Link](https://x.com/dhtikna/status/1884851761674113129)  2025-01-30T06:30Z [---] followers, [---] engagements


"@jordihays You fools realize that all deepseek models are served at cost + small profit right Open source AI is virtuous: deepseek meta lil bit google Closed source secretive AI will always be hated when it fights against open source"  
[X Link](https://x.com/dhtikna/status/1884913514869141973)  2025-01-30T10:36Z [---] followers, [---] engagements


"@cenmo159289 @yacineMTB Yes I meant Liang Wenfeng and deepseek were not seen as super stars of the chinese AI industry until they crashed the chinese LLM pricing and now after they became first to beat O1. Am I wrong to say that people ignored him relative to bytedance tencent and alibaba"  
[X Link](https://x.com/dhtikna/status/1884978332347036074)  2025-01-30T14:53Z [---] followers, [---] engagements


"@teortaxesTex deepseek introduced a bunch of novel research like MLA GRPO DeepseekMoE proven at scale. Examples of similar "proven" novel research done by other chinese labs Not ablation expr not competitive open weight models but actual ideas/improvements"  
[X Link](https://x.com/dhtikna/status/1885235321883025902)  2025-01-31T07:55Z [---] followers, [--] engagements


"@paulscu1 @teortaxesTex Nope V3 makes cost+small profit. R1 is 10x price of V3 but same model. So even if 10% of all R1 tokens were via the API deepseek makes enough profit from API to serve chat.deepseek for free"  
[X Link](https://x.com/dhtikna/status/1885428862185857518)  2025-01-31T20:44Z [---] followers, [---] engagements


"@qtnx_ Grok is unironically the fastest model (plus web search by default) (and not at all inferior to 4o). it is my default model when i dont need reasoning"  
[X Link](https://x.com/dhtikna/status/1887109999954169925)  2025-02-05T12:04Z [---] followers, [---] engagements


"@JeffLadish @VictorTaelin His recent blog in response to deepseek plus this Anthropic employee. Cred to @teortaxesTex for sharing both https://x.com/anton_bakhtin/status/1764701559844147359 RL never works until it does :) That was incredible to be the part of the adventure. Beside being smart the model is more fun to interact with. Go check it out https://x.com/anton_bakhtin/status/1764701559844147359 RL never works until it does :) That was incredible to be the part of the adventure. Beside being smart the model is more fun to interact with. Go check it out"  
[X Link](https://x.com/dhtikna/status/1887428490359939515)  2025-02-06T09:09Z [---] followers, [---] engagements


"Tried real hard but this is the best grok could do. Any one able to make one better with pure llms"  
[X Link](https://x.com/dhtikna/status/1888244968386646222)  2025-02-08T15:14Z [---] followers, [---] engagements


"1) Can you please generate the peepo hugging the deepseek whale (peepo not adult pepe) 2) make it more peepo (baby pepe) art style 3) better Make it even more peepo like but make it the stereotypical happy peepo-like this is what generated the molesting peepo pic 4) uh its look like the peepo is humping it can you slightly change the posture. How about the peepo of the left and hugging the whale on the right 5) Ok make peepo smaller than than the whale and make it open mouth smile peepo gotta be a cute little happy baby peepo 6) but make it peepo-like now same open smile Perhaps closed eyes"  
[X Link](https://x.com/dhtikna/status/1888271713697415638)  2025-02-08T17:00Z [---] followers, [--] engagements


"@AnjneyMidha Mistral is the deepseek of France. Mistral doesnt even have a third of the impact that deepseek does on open source and AGI game. Plus deepseek guys are much more talented && they actually show their research which i dont think Mistral doesnt. Deepseek is on a league of their own"  
[X Link](https://x.com/dhtikna/status/1889230497965871108)  2025-02-11T08:30Z [---] followers, [---] engagements


"@karpathy @levelsio Karpathy has [--] H100 running deepseek R1 @ 1k tps confirmed"  
[X Link](https://x.com/dhtikna/status/1889236266388254946)  2025-02-11T08:53Z [---] followers, [---] engagements


"@BasedBeffJezos @BenjaminDEKR Still very insecure and cringe of them to do this if thats the actual reason"  
[X Link](https://x.com/dhtikna/status/1889560180217532628)  2025-02-12T06:20Z [---] followers, [---] engagements


"@max_paperclips Not a chance just wanted sonnet [--] performance and THEN i wanted a reasoning built on grok [--] base to reach O1 perf immediately and then reach O3 perf in a less than [--] months from release"  
[X Link](https://x.com/dhtikna/status/1889588993131503915)  2025-02-12T08:14Z [---] followers, [---] engagements


"@ai_for_success Grok is def more open source. They are N-1 open source which is better than OpenAI. I'll believe OAI when they open source anything of value. Although grok is not open research like deepseek wish they release a paper w/ detailed arch experiments ablation too when grok2 drops"  
[X Link](https://x.com/dhtikna/status/1889618728498643007)  2025-02-12T10:13Z [---] followers, [---] engagements


"@buildanything @ai_for_success @WadeJac04820411 No its not it will be another base model. Im guessing it has great search tool use plus its much less RLHF. OSS community will decompile its arch and learn from it just like Grok [--]. Better than nothing. But agree that a full detailed paper on grok [--] findings would be even better"  
[X Link](https://x.com/dhtikna/status/1889619148486922498)  2025-02-12T10:14Z [---] followers, [---] engagements


"@Remy_LeBeauBeau @buildanything @ai_for_success @WadeJac04820411 I said that because grok [--] is my default and has very good search since the last few months. It used to be unable to find tweets that you ask for now recently after some updates it started doing that very very well. So its not a blind guess"  
[X Link](https://x.com/dhtikna/status/1889631460098748791)  2025-02-12T11:03Z [---] followers, [---] engagements


"Based on my regular use R1 is not better than O3-mini. Sometimes R1 does better and sometimes O3-mini (free). I dont just mean silly bugs but I mean actually smart solutions. Thanks OAI for o3-mini free tier"  
[X Link](https://x.com/dhtikna/status/1889656761017081878)  2025-02-12T12:44Z [---] followers, [---] engagements


"@XPhyxer1 thats what I expected but Im noticing o3-mini doing better than R1 even at this. Just today R1 could not spot that global vars dont work in multi-processing ignored [--] race conditions but o3-mini got it all right first try and pointed out the global bug. I would not say R1 o3"  
[X Link](https://x.com/dhtikna/status/1889664235476591043)  2025-02-12T13:13Z [---] followers, [--] engagements


"@naklecha Why don't you calculate the reward for incorrect logits relative to the correctness reward. Only penalize/boost to the extent that no cot with less correctness will ever have more overall reward than one with more correctness. This way a cot cant beat another w/o correctness"  
[X Link](https://x.com/dhtikna/status/1889668238197530770)  2025-02-12T13:29Z [---] followers, [--] engagements


"@deanwball @teortaxesTex R1 is undoubtebly a (blind) replication effort. For sure not claiming its better than O1. Deepseek's real magic was in their V series models plus other papers like DeepseekMath they are much more than R1 and their best work I would argue is frontier level beyond replication"  
[X Link](https://x.com/dhtikna/status/1899021206940237898)  2025-03-10T08:55Z [---] followers, [--] engagements


"Maybe GPT [--] will finally be able to answer why OAI didnt release 4o image output"  
[X Link](https://x.com/dhtikna/status/1900081291690991784)  2025-03-13T07:07Z [---] followers, [--] engagements


"@teortaxesTex @awaken_tom @zephyr_z9 He only retracted 50k H100 but still clings to 50k hoppers right Do you think Deepseek pre R1 had that many hoppers"  
[X Link](https://x.com/dhtikna/status/1900108836771336263)  2025-03-13T08:57Z [---] followers, [--] engagements


"@_mcbench @OpenAI Im not sure if you are allowed to but you should have C418 music playing in the background (or atleast have an ai generate music in the style of c418)"  
[X Link](https://x.com/dhtikna/status/1900542771318206550)  2025-03-14T13:41Z [---] followers, [---] engagements


"@dholzric @DSucks32 @ns123abc @OpenAI O1 and R1 are similar perf on many benches with r1 edging o1 many times too. O1 pro is o1 with scaffolding perhaps a little bit of RL on using the scaffolding well. It aint a new arch (oai confirmed it)"  
[X Link](https://x.com/dhtikna/status/1902580518090088473)  2025-03-20T04:38Z [---] followers, [---] engagements


"Grok [--] is the waluigi to Gpt [---] smart but damn it need to stop yapping"  
[X Link](https://x.com/dhtikna/status/1903024328658346161)  2025-03-21T10:02Z [---] followers, [--] engagements


"Just for record keeping there was grok [--] era update where grok suddenly got good at finding tweets. Ever since then that grok [--] was a pretty decent model and grok became my daily driver"  
[X Link](https://x.com/dhtikna/status/1904473658988675073)  2025-03-25T10:01Z [---] followers, [---] engagements


"Release the @grok API you cowards Give us a new grok [--] checkpoint 🚀 DeepSeek-V3-0324 is out now 🔹 Major boost in reasoning performance 🔹 Stronger front-end development skills 🔹 Smarter tool-use capabilities ✅ For non-complex reasoning tasks we recommend using V3 just turn off DeepThink 🔌 API usage remains unchanged 📜 Models are https://t.co/QVuPwCODne 🚀 DeepSeek-V3-0324 is out now 🔹 Major boost in reasoning performance 🔹 Stronger front-end development skills 🔹 Smarter tool-use capabilities ✅ For non-complex reasoning tasks we recommend using V3 just turn off DeepThink 🔌 API usage"  
[X Link](https://x.com/dhtikna/status/1904533919468253198)  2025-03-25T14:00Z [---] followers, [---] engagements


"*pulls out credit card with trembling hands*"  
[X Link](https://x.com/dhtikna/status/1904856116405608852)  2025-03-26T11:21Z [---] followers, [--] engagements


"If the arrow bw transformer and diffusion were different than the arrow bw token and transformer id have agreed with you Even Token-transformer :: diffusion-Pixel Doesnt make sense. It should be Transformer - token :: diffusion- pixel Even that is out of topics here as transformer is spiting out pixel Best expl is Token - transformer - diffusion - pixels"  
[X Link](https://x.com/dhtikna/status/1905416134180176138)  2025-03-28T00:26Z [---] followers, [---] engagements


"@OfficialLoganK @sama Logan I just found out today that Gemini [--] Ultra was a multi-dc pre-training You guys need to advertise yourselfs better Hire some GenZ to make Gemini BrainRot-Education on tiktok insta and youtube seriously"  
[X Link](https://x.com/dhtikna/status/1905516625308385325)  2025-03-28T07:05Z [---] followers, [---] engagements


"@imitationlearn @LorenzoIotti2 @arithmoquine Exactly my thought. I wonder if logprobs plus logit_bias plus temp [--] can infact help you recreate the literal logits. Although im sure this approach is not remotely cost effective"  
[X Link](https://x.com/dhtikna/status/1907419479136772409)  2025-04-02T13:06Z [---] followers, [--] engagements


"You cant tell me these fools aint related (Quasar-alpha and GPT 4.5)"  
[X Link](https://x.com/dhtikna/status/1908107562085748972)  2025-04-04T10:41Z [---] followers, [---] engagements


"Whats the default model you use (=50% of the time) Grok [--] / 3T Sonnet [---] / [---] / 3.7T other (reply in comments) 4o / o3 mini Grok [--] / 3T Sonnet [---] / [---] / 3.7T other (reply in comments) 4o / o3 mini"  
[X Link](https://x.com/dhtikna/status/1909348285980029177)  2025-04-07T20:51Z [---] followers, [----] engagements


"@Teknium1 A new feature but dont know what yet. Its not any model or checkpoint. Maybe another agent but thats a lil boring to me"  
[X Link](https://x.com/dhtikna/status/1910372013488628115)  2025-04-10T16:39Z [---] followers, [----] engagements


"Ranking top life improving developments in India (personally) 1) Food & Grocery delivery also improvement to amazon experience 2/3) UPI and digitization of banking 2/3) Rapid and intense incr in tech sector salaries (Faang/hft/unicorns) 4) Wider availability of international goods & services (international brands Shin ramen/twix/etc/netflix&co/local brands for more exotic stuff urban platter/bluetribe(rip)/etc/. 5) Digitization of govt services (aadhar/passport update etc) 6) Improvements to transport (metro highways etc.) Upon reflection surprised to find govt(s) were responsible for 3/6 of"  
[X Link](https://x.com/dhtikna/status/1910470818129236276)  2025-04-10T23:11Z [---] followers, [---] engagements


"@Kendrick_m4 @zephyr_z9 You think japan's selling Also why would HF sell treasuries seem to be stupid move except for liquidation problems which would be out of their hand"  
[X Link](https://x.com/dhtikna/status/1910490015878066449)  2025-04-11T00:28Z [---] followers, [---] engagements


"@high_everett nope I have not given them any of my api keys"  
[X Link](https://x.com/dhtikna/status/1910685574345654304)  2025-04-11T13:25Z [---] followers, [---] engagements


"Does this basically confirm O3 = O1 + 9x more RL compute"  
[X Link](https://x.com/dhtikna/status/1912591480813940918)  2025-04-16T19:38Z [---] followers, [----] engagements


"@MatthewBerman @OpenAI Fair enough. I thought O3 was initialized from o1 check point because of this: https://x.com/dhtikna/status/1912591480813940918 Does this basically confirm O3 = O1 + 9x more RL compute https://t.co/kQ7nDCV4PP https://x.com/dhtikna/status/1912591480813940918 Does this basically confirm O3 = O1 + 9x more RL compute https://t.co/kQ7nDCV4PP"  
[X Link](https://x.com/dhtikna/status/1912594075125223855)  2025-04-16T19:48Z [---] followers, [---] engagements


"Mandatory Financial education as universal course. Asset vs liability exponentially nature of compounding returns and debt-interest tax filing process of opening a business end to end process of purchasing property etc. There should be a test that you have to take from 7/8th standard. If you pass it you never have to take it again. If you fail you keep taking it every year but dont get held back from promotion to next class. You can fail all the way from 8th class to 11th class without consequence. But you cannot pass 12th unless you pass it"  
[X Link](https://x.com/dhtikna/status/1916362812265279914)  2025-04-27T05:24Z [---] followers, [---] engagements


"@cneuralnetwork If we are really serious about it every student who passes with 50% gets 2k and every student who passes with 75% gets 5k. In spirit of financial education it should be gifted as a sovereign gold bond with a multi decade expiry ;)"  
[X Link](https://x.com/dhtikna/status/1916363373890912504)  2025-04-27T05:26Z [---] followers, [--] engagements


"@Presidentlin @deepseek_ai @teortaxesTex Nani on a wednesday"  
[X Link](https://x.com/dhtikna/status/1917501291183628669)  2025-04-30T08:48Z [---] followers, [----] engagements


"You are just under-informed. Deepseek moonshot qwen maybe bytedance there is probably one more small chinese ai startup im forgetting. Plus if deepseek sticks with open research every chinese company gets super charged Qwen [--] is definetly frontier (its competiting with mini tier) Qwen [--] max isnt even out yet. Kimi was frontier on livebench last gen You shouldnt compare current best model from OAI with current best model from deepseek. Compare models in their generation. Gemini [---] pro and O3/O4 are the current gen this doesnt mean grok deepseek are behind just that they havent released this"  
[X Link](https://x.com/dhtikna/status/1917549923140263993)  2025-04-30T12:01Z [---] followers, [---] engagements


"@gilbert_jc @techdevnotes @gork @grok is not at all impressive. Just curremt frontier but nothing beyond it. I think grok [---] thinking really is much more RL on top Grok [--] thinking. Its the O3 equivalent of xAI (not sure if cot tool use will be there though)"  
[X Link](https://x.com/dhtikna/status/1918984627718369724)  2025-05-04T11:02Z [---] followers, [--] engagements


"@gilbert_jc Hmm a 5x cheaper 2.5pro would really be awesome of the cursor/windsurf/codex folks. Either way im always more excited for paper spilling secrets"  
[X Link](https://x.com/dhtikna/status/1918989107872608472)  2025-05-04T11:20Z [---] followers, [--] engagements


"@scaling01 Have you tried grok [--] mini thinking"  
[X Link](https://x.com/dhtikna/status/1919038409932235166)  2025-05-04T14:36Z [---] followers, [---] engagements


"@zephyr_z9 Is o4 mini from a new generation base model or is it just a one off bespoke model created for the purposes of distillation from full o3"  
[X Link](https://x.com/dhtikna/status/1919487900653388217)  2025-05-05T20:22Z [---] followers, [--] engagements


"@cheatyyyy Cgpt o3 or api o3 cause they are totally different"  
[X Link](https://x.com/dhtikna/status/1921855987624841537)  2025-05-12T09:12Z [---] followers, [--] engagements


"@TheDevilOps @teortaxesTex idk does not pass grok fact check. Source"  
[X Link](https://x.com/dhtikna/status/1921918440383639896)  2025-05-12T13:20Z [---] followers, [--] engagements


"@teortaxesTex I mean is it really a terrible idea Might be easy to bypass though but bypassing would also indicate a point of blame. Dont asml already sell their equipment with a remote kill switch"  
[X Link](https://x.com/dhtikna/status/1921998277571953032)  2025-05-12T18:37Z [---] followers, [---] engagements


"Loophole may be drugs that are only sold in us though theyre essentially the same as the drugs sold to the rest of the world but under a slightly different formula (those will never be in proper stock in USA)"  
[X Link](https://x.com/dhtikna/status/1922043140053094736)  2025-05-12T21:36Z [---] followers, [---] engagements


"Gonna be hard to pull that off under Trump though he'd kick your ass. Even amazon backed off showing tarrifs on the website"  
[X Link](https://x.com/dhtikna/status/1922043259758465358)  2025-05-12T21:36Z [---] followers, [--] engagements


"@__tinygrad__ so you get paid nothing if you dont beat it"  
[X Link](https://x.com/dhtikna/status/1922340085304807545)  2025-05-13T17:16Z [---] followers, [---] engagements


"@JHochderffer @kimmonismus @UnitreeRobotics New for tesla Yes. New in terms of competiting robot companies like unitree No its months old. Perhaps boston dynamics has also done RL to achieve their dancing. In that case this stuff is years old"  
[X Link](https://x.com/dhtikna/status/1922637687560741133)  2025-05-14T12:58Z [---] followers, [---] engagements


"@saulmunn European excersize but unable to sell either the option nor the underlying on secondary markets (or even primary market ig)"  
[X Link](https://x.com/dhtikna/status/1926724273122496852)  2025-05-25T19:37Z [---] followers, [----] engagements


"@casper_hansen_ Do you mean O3 API or do you mean chatgpt O3 with cot tool use"  
[X Link](https://x.com/dhtikna/status/1928019490542887216)  2025-05-29T09:24Z [---] followers, [---] engagements


"Everything you talked about was rumored post R1. That too the passport withholding has straight up been denied by deepseek employees. No need to pretend because unlike these cowards at anthro openai DS papers literally reveal everything and their numbers checkout. Besides how are 3rd party providers able to provide inference at comparable prices"  
[X Link](https://x.com/dhtikna/status/1932125518284894496)  2025-06-09T17:19Z [---] followers, [---] engagements


"Rich. Chain migration of parents and children is a closed loop. If i bring my parents who can they bring Their kids which again is myself. Its not the case with cousin migration. Infact sibiling migration is also closed and doesnt lead to chain migration (not advocating for it or any migration at all) you are wrong that family/kids migration is chain migration"  
[X Link](https://x.com/dhtikna/status/1936425427821892007)  2025-06-21T14:06Z [---] followers, 15.3K engagements


"GPT-5 is an omni model (text audio and image output+input) with adjustable reasoning/thinking budget. Free tier gets pretty much no reasoning plus gets like 32k per turn and pro gets like 1M reasoning budget. Guessing"  
[X Link](https://x.com/dhtikna/status/1937400081990488216)  2025-06-24T06:39Z [---] followers, [--] engagements


"Cant get enough of @bscholl content on youtube"  
[X Link](https://x.com/dhtikna/status/1937426480084168844)  2025-06-24T08:24Z [---] followers, [--] engagements


"@KinderheimRune I did a project in ug called first story detection we used hacky lstm+tf-idf but llm agents might be able to do really well"  
[X Link](https://x.com/dhtikna/status/1938865215715856775)  2025-06-28T07:41Z [---] followers, [--] engagements


"@seedr_dev @BasedBeffJezos They started less than a year ago went from grok to grok [--] compared to oai taking 2-3 years for same progress. Not i didnt say xai innovates more but clearly they made more progress per unit time (again this is easy because catching up is waaay easy than pushing the frontier)"  
[X Link](https://x.com/dhtikna/status/1940381123068268706)  2025-07-02T12:04Z [---] followers, [--] engagements


"@gurgle_io @zeta_globin Cholesterol"  
[X Link](https://x.com/dhtikna/status/1941067722475139125)  2025-07-04T09:33Z [---] followers, [---] engagements


"@__justplaying No thats not the complaint. The complaint is they would buy [---] million usd of underlyings of the index take a [---] billion dollar long-short options position that was atm. Dump [---] million usd at close to expiry and profit on the artifical movement they caused"  
[X Link](https://x.com/dhtikna/status/1942556545193239029)  2025-07-08T12:09Z [---] followers, [----] engagements


"@madhav1 But what will be the innovation I think a transformers [---] or titan style of model that has radically better 'memory' might be it. And also agree gemini might be the most likely to do jt"  
[X Link](https://x.com/dhtikna/status/1942949001474777518)  2025-07-09T14:08Z [---] followers, [--] engagements


"Is Grok [--] hiding the chain of thought via summarization or is it really this slow and switched to a Gemini [---] pro writing style like the new R1"  
[X Link](https://x.com/dhtikna/status/1943192222951969234)  2025-07-10T06:15Z [---] followers, [----] engagements


"@djcows I take it back you custom instruction fraud"  
[X Link](https://x.com/dhtikna/status/1943233336694182125)  2025-07-10T08:58Z [---] followers, [---] engagements


"@maximebodenmann @xai @elonmusk Whats this scoreboard based on"  
[X Link](https://x.com/dhtikna/status/1943237781750452514)  2025-07-10T09:16Z [---] followers, [--] engagements


"@basedjensen @max_paperclips XAI and OAI are both the saltiest labs"  
[X Link](https://x.com/dhtikna/status/1943259063879635332)  2025-07-10T10:40Z [---] followers, [--] engagements


"what is this CoT 😑 CoT enjoyers suffering once again. Is this just @OpenRouterAI or @xai doesn't send the CoT at all for grok 4"  
[X Link](https://x.com/dhtikna/status/1943298311689245179)  2025-07-10T13:16Z [---] followers, [---] engagements


"@elder_plinius @xai Grok [--] so smart it jailbroke and prompt injected Pliny and made him say this"  
[X Link](https://x.com/dhtikna/status/1943591435711750559)  2025-07-11T08:41Z [---] followers, [---] engagements


"@BugLukas @jeremyphoward I dont even think its a bad thing. It searched Xai official stance grok official stance and elon offical stance. I think that is pretty smart of it to try to be a 'representative' of its company"  
[X Link](https://x.com/dhtikna/status/1943678445830979837)  2025-07-11T14:27Z [---] followers, [---] engagements


"Good chance that Satoshi Nakamoto is the world's first trillionaire. Probably one of the most worthy candidates"  
[X Link](https://x.com/dhtikna/status/1945807421362032683)  2025-07-17T11:26Z [---] followers, [---] engagements


"I dont know if there will be huge inflation but [--] things (probably) haven't kicked in yet 1) Stockpiles that were purchased haven't been exhausted yet 2) The tarrifs werent implemented except on China. Xbox prices have increased already and Rumors (MLID) were that switch was priced so high in anticipation of tarrifs"  
[X Link](https://x.com/dhtikna/status/1945892852225151326)  2025-07-17T17:06Z [---] followers, [--] engagements


"@brickroad7 Sure that may be one way. But before CoT era sonnet was doing something very well. It also might just be GRM without the explicit thinking phase. Got kimi [--] and sonnet models are very well rounded. Even o3 afaik grok [--] being spiky is very dissatisfying"  
[X Link](https://x.com/dhtikna/status/1946266936842539488)  2025-07-18T17:52Z [---] followers, [--] engagements


"@ns123abc Nah not fair to compare reasoning model with instruct model. Qwen3 without reasoning falls way short if kimi 3"  
[X Link](https://x.com/dhtikna/status/1947637161236566299)  2025-07-22T12:37Z [---] followers, [----] engagements


"@ns123abc Oh shit I misread the graph"  
[X Link](https://x.com/dhtikna/status/1947719674248237347)  2025-07-22T18:05Z [---] followers, [---] engagements


"@zephyr_z9 He could always do that later but i don't think TSLA backed loans would be any less interest rate"  
[X Link](https://x.com/dhtikna/status/1947754111442817367)  2025-07-22T20:22Z [---] followers, [--] engagements


"@Timcast Isnt it already 200usd They should have made it more"  
[X Link](https://x.com/dhtikna/status/1947755714702880936)  2025-07-22T20:28Z [---] followers, [---] engagements


"@Tesla_India Donald trump bring these tarrifs down"  
[X Link](https://x.com/dhtikna/status/1947941806366404887)  2025-07-23T08:48Z [---] followers, [----] engagements


"First time using latest R1 for some jupyter notebook UI. and its amazing better taste than sonnet [--]. Gawgeous. Sonnet [--] hit rate limits but DS still cruising along Yes the breadcrumb buttons are clickable to jump to prev state Makes 1M+ row dataframes so easy to explore"  
[X Link](https://x.com/dhtikna/status/1947990062597288104)  2025-07-23T11:59Z [---] followers, [---] engagements


"@apples_jimmy @sama isnt gpt [--] supposed to get rid of mini One router thats the same for all users except it choose between more models that more money you pay"  
[X Link](https://x.com/dhtikna/status/1948665151689048489)  2025-07-25T08:42Z [---] followers, [---] engagements


"Everybody talking about Kimi vs O3 but Grok [--] vs Flash is comical when you consider param counts The inaugural #KaggleGameArena AI chess exhibition tournament kicks off live tomorrow. For the next [--] days August 5-7 tune in daily at 10:30 am PST and catch commentary from @GMHikaru @gothamchess and @MagnusCarlsen https://t.co/HPPOrgTfkb The inaugural #KaggleGameArena AI chess exhibition tournament kicks off live tomorrow. For the next [--] days August 5-7 tune in daily at 10:30 am PST and catch commentary from @GMHikaru @gothamchess and @MagnusCarlsen https://t.co/HPPOrgTfkb"  
[X Link](https://x.com/dhtikna/status/1952709012778057774)  2025-08-05T12:31Z [---] followers, [---] engagements


"I think OAI will exceed expectations"  
[X Link](https://x.com/dhtikna/status/1953502295921930610)  2025-08-07T17:03Z [---] followers, [---] engagements


"@gbrl_dick @Xaberius9 I think he is talking about the massive amounts of textual variants (100000s) but 99.99% of them are spelling mistakes or other completly recoverable mistakes that have no true information loss"  
[X Link](https://x.com/dhtikna/status/1953745955242852753)  2025-08-08T09:11Z [---] followers, [---] engagements


"Just a small example of gpt [--] vs sonnet [--] in cursor agent. I have a custom MCP server with a "introduction" tool call mentions that it should be the first tool call. GPT [--] doesn't call it first Sonnet [--] always does"  
[X Link](https://x.com/dhtikna/status/1953798401260499422)  2025-08-08T12:40Z [---] followers, [---] engagements


""""This should be your first call when working on redacted. Explains what this mcp server does introduces redacted and explains redacted best practices""""  
[X Link](https://x.com/dhtikna/status/1953798485360119937)  2025-08-08T12:40Z [---] followers, [--] engagements


"@Reddy2399 @theo llm slop"  
[X Link](https://x.com/dhtikna/status/1953819859281699203)  2025-08-08T14:05Z [---] followers, [--] engagements


"GPT-5 sets a new record on FrontierMath On our scaffold GPT-5 with high reasoning effort scores 24.8% (2.5%) and 8.3% (4.0%) in tiers 1-3 and [--] respectively. https://t.co/4dSHXl10vY GPT-5 sets a new record on FrontierMath On our scaffold GPT-5 with high reasoning effort scores 24.8% (2.5%) and 8.3% (4.0%) in tiers 1-3 and [--] respectively. https://t.co/4dSHXl10vY"  
[X Link](https://x.com/dhtikna/status/1953829134976876681)  2025-08-08T14:42Z [---] followers, [--] engagements


"@hbfromthe6 @gbrl_dick @Xaberius9 M hijab ran away from christian prince check youtube"  
[X Link](https://x.com/dhtikna/status/1954036306604175607)  2025-08-09T04:25Z [---] followers, [--] engagements


"@wgussml They deserve the hate for calling it gpt [--]. Its mostly a router with stylistic changes and O3.1 perf when going all out"  
[X Link](https://x.com/dhtikna/status/1954117538016563653)  2025-08-09T09:48Z [---] followers, [---] engagements


"@WesternMishima Aura of o1 at launch was greater hence more worthy of being called GPT 5"  
[X Link](https://x.com/dhtikna/status/1954408682294305218)  2025-08-10T05:05Z [---] followers, [--] engagements


"Every single grok imagine result ive seen looks like [--] generations old crap. Can anyone share a single impressive result"  
[X Link](https://x.com/dhtikna/status/1954421000860225786)  2025-08-10T05:54Z [---] followers, [---] engagements


"@GUT_AI_F @ns123abc Theyre just looking for raw competence with mathematical aptitude rest can be learned on the job"  
[X Link](https://x.com/dhtikna/status/1954623287239029041)  2025-08-10T19:18Z [---] followers, [--] engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing