# ![@sumjitg Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::1720888315174645760.png) @sumjitg Sumjit

Sumjit posts on X about ai, open ai, math, tao the most. They currently have [---] followers and [---] posts still getting attention that total [--] engagements in the last [--] hours.

### Engagements: undefined [#](/creator/twitter::1720888315174645760/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1720888315174645760/c:line/m:interactions.svg)

- [--] Week [---] -74%
- [--] Month [------] -82%
- [--] Months [-------] +68%
- [--] Year [-------] +174%

### Mentions: undefined [#](/creator/twitter::1720888315174645760/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1720888315174645760/c:line/m:posts_active.svg)

- [--] Week [--] -83%
- [--] Month [--] -41%
- [--] Months [---] +88%
- [--] Year [---] +144%

### Followers: [---] [#](/creator/twitter::1720888315174645760/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1720888315174645760/c:line/m:followers.svg)

- [--] Week [---] -0.35%
- [--] Month [---] no change
- [--] Months [---] +45%
- [--] Year [---] +443%

### CreatorRank: undefined [#](/creator/twitter::1720888315174645760/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1720888315174645760/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  36% [stocks](/list/stocks)  18% [finance](/list/finance)  10% [celebrities](/list/celebrities)  4% [countries](/list/countries)  3% [social networks](/list/social-networks)  3% [automotive brands](/list/automotive-brands)  1% [formula 1](/list/formula-1)  1%

**Social topic influence**
[ai](/topic/ai) 56%, [open ai](/topic/open-ai) 13%, [math](/topic/math) 13%, [tao](/topic/tao) 12%, [llm](/topic/llm) 10%, [$googl](/topic/$googl) 9%, [the first](/topic/the-first) 9%, [core](/topic/core) 8%, [how to](/topic/how-to) 8%, [inference](/topic/inference) 7%

**Top accounts mentioned or mentioned by**
[@aiforsuccess](/creator/undefined) [@levie](/creator/undefined) [@testingcatalog](/creator/undefined) [@kimmonismus](/creator/undefined) [@jukan05](/creator/undefined) [@kyleichan](/creator/undefined) [@zaiorgs](/creator/undefined) [@rakyll](/creator/undefined) [@outis000](/creator/undefined) [@alexalbert](/creator/undefined) [@sebastienbubeck](/creator/undefined) [@ns123abc](/creator/undefined) [@slowdeveloper](/creator/undefined) [@grok](/creator/undefined) [@justalexoki](/creator/undefined) [@genaiisreal](/creator/undefined) [@mxtaverse](/creator/undefined) [@davidondrej1](/creator/undefined) [@realemissary](/creator/undefined) [@theahmadosman](/creator/undefined)

**Top assets mentioned**
[Alphabet Inc Class A (GOOGL)](/topic/$googl) [Tesla, Inc. (TSLA)](/topic/tesla)
### Top Social Posts
Top posts by engagements in the last [--] hours

"Sarvam AI () just dropped benchmark results showing their OCR model hitting 84.3% accuracy on olmOCR-Bench: beating both Gemini and GPT. Released Feb [--] the model's built specifically for Indic languages which matters for real-world use cases like digitizing government docs or regional education materials. https://twitter.com/i/web/status/2021111550472523820 https://twitter.com/i/web/status/2021111550472523820"  
[X Link](https://x.com/sumjitg/status/2021111550472523820)  2026-02-10T06:38Z [---] followers, [---] engagements


"What's the difference: Modified MIT license (mostly open commercial use) Explicit reasoning logs (full transparency & auditability) INT4 quantization (594GB vs 1TB+ way cheaper to deploy) Open-source AI just matched closed frontier models on agentic reasoning"  
[X Link](https://x.com/sumjitg/status/1986803092474785808)  2025-11-07T14:29Z [---] followers, [--] engagements


"Wait GPT-5.2 just casually showed up in Cursor AI with a 272k context window 👀 People are freaking out thinking OpenAI might drop it Dec 10-11. Makes sense they'd rush it after Google's Gemini [--]. The coding/debugging upgrades sound insane but honestly who knows if the timing's real lol @ai_for_success https://t.co/BCQjxFIpEj @ai_for_success https://t.co/BCQjxFIpEj"  
[X Link](https://x.com/sumjitg/status/1998802249242526135)  2025-12-10T17:09Z [---] followers, [----] engagements


"Okay so LLMs training in space is actually happening but let's pump the brakes on the hype: Yes they literally trained nanoGPT on Shakespeare aboard a satellite with an H100 GPU. Cool Absolutely. Game-changer Not quite yet. This is more "hello world" for orbital AI than skynet-in-space. It proves the hardware can survive radiation and vacuum which is genuinely impressive engineering. But we're talking about a 124M parameter model not GPT-4 vibing among the stars. The "AI models will live in space" narrative is getting ahead of reality. Nvidia's excited about solar-powered orbital data centers"  
[X Link](https://x.com/sumjitg/status/1998836140787507633)  2025-12-10T19:24Z [---] followers, [--] engagements


"lol so @kyleichan tested [--] AI models on "how many R's in garlic" and GPT-5.2 said ZERO because it got weird about uppercase vs lowercase Gemini [--] DeepSeek R1 and Qwen3-Max all got it right though. AGI is here🥸 How many Rs in garlic I asked this same question to four AI models. Heres how they performed: GPT 5.2: incorrect Gemini 3: correct DeepSeek R1: correct Qwen3-Max: correct https://t.co/X6JnJuDuRz How many Rs in garlic I asked this same question to four AI models. Heres how they performed: GPT 5.2: incorrect Gemini 3: correct DeepSeek R1: correct Qwen3-Max: correct https://t.co/X6JnJuDuRz"  
[X Link](https://x.com/sumjitg/status/1999325041465651231)  2025-12-12T03:46Z [---] followers, [----] engagements


"Mark Cuban's onto something: small businesses WANT AI but have no clue how actually to use it. They've got the budget just not the know-how. So if you can bridge that gap: customise models show real ROI actually make this stuff work for a bakery or law firm- you're golden. Mark Cuban on the next big job students should focus on: Most companies dont know how to implement AI especially small businesses. Companies dont understand how to implement AI right now to get a competitive advantage learn to customize a model walk into a company show https://t.co/uHOlgLMdOi Mark Cuban on the next big job"  
[X Link](https://x.com/sumjitg/status/2000091542518980676)  2025-12-14T06:32Z [---] followers, [---] engagements


"NVIDIA just dropped Nemotron [--] Nano and it's genuinely impressive - 31.6B parameters but only uses 3.6B at a time thanks to smart MoE architecture. The speed gains are wild: 3.3x faster than Qwen3-30B on H200 GPUs while actually being MORE accurate on benchmarks like SWE-Bench (38.8%) and handling 1M token contexts like it's nothing. Best part They released everything - weights training recipes datasets. It's crushing the competition on coding tasks (78% on HumanEval) and runs locally on 24GB VRAM. The agentic reasoning capabilities are what really set it apart though. Tech community is hyped"  
[X Link](https://x.com/sumjitg/status/2000832373257330715)  2025-12-16T07:36Z [---] followers, [---] engagements


"Bernie Sanders wants to pause new AI data centres and honestly The numbers are wild. These facilities are guzzling electricity like 700K homes and burning through 560B liters of water yearly just to keep servers cool. Meta's Louisiana project alone could TRIPLE New Orleans' power usage. Tech CEOs basically admit the job concerns are real- Musk says AI could replace "all jobs" Gates sees way less human labor needed and Anthropic's CEO warns half of entry-level office jobs might vanish by [----]. The debate's messy though. Goldman Sachs says 300M jobs worldwide could get automated but claims"  
[X Link](https://x.com/sumjitg/status/2001206923887305105)  2025-12-17T08:24Z [---] followers, [---] engagements


"Sam Altman statment on first solo founder hit unicorn status ($1B valuation) by 2026-2028 and the pieces are starting to fall into place. We're already seeing teens build legit startups using "vibe-coding" basically chatting with AI to build apps without knowing how to code. Like a 16-year-old built Delv .AI and got it to a $12M valuation. Wild Andrew Ng's take makes sense too: AI can pump out way more viable business ideas than we have people to execute them. So solo founders with AI doing the heavy lifting Not that crazy anymore. But here's the thing everyone's debating: can you actually"  
[X Link](https://x.com/sumjitg/status/2003099886930272563)  2025-12-22T13:46Z [---] followers, [---] engagements


"GLM-4.7 is seriously impressive for coding and agentic workflows. benchmarks: 84.9% on LiveCodeBench 73.8% on SWE-bench Verified basically hanging with DeepSeek-V3.2 and Claude Sonnet [---]. 42.8% HLE score (close to @Zai_orgs 43% claim) shows it's legit at multi-step reasoning with tools. No full research paper yet since it just dropped (Dec 21) but @Zai_orgs blog breaks down the interleaved thinking modes that make it shine for tool use. Tech YouTubers are already demoing game prototypes and UI gen- it's fast and surprisingly creative Community's hyped especially on Reddit and Hacker News."  
[X Link](https://x.com/sumjitg/status/2003525158632616150)  2025-12-23T17:56Z [---] followers, [----] engagements


"Wait so they want to put 6G. inside our bodies And smartphones will just disappear by [----] Nokia's CEO said this back in [----] and now governments are pushing hard to make it reality. The idea is 6G tech gets embedded directly in us- no more carrying devices. Sure the medical benefits sound cool. Remote health monitoring instant diagnostics all that. But we're also talking about turning humans into literal network nodes. Research papers mention "wireless body networks" and "digital twins" of our bodies. Sounds futuristic right Except they're also quietly noting potential DNA damage from high"  
[X Link](https://x.com/sumjitg/status/2003727028491813272)  2025-12-24T07:18Z [---] followers, [---] engagements


"Bill Gates out here saying we need "digital ID systems" to fight misinformation because the First Amendment is apparently making things "tough." His solution Track who's saying what online so we know "who created this." Yeah that doesn't sound dystopian at all"  
[X Link](https://x.com/sumjitg/status/2003740693773345000)  2025-12-24T08:13Z [---] followers, [----] engagements


"Holy shit a 6-person startup just beat Google at AI reasoning 🤯 Poetiq AI hit 54% on the ARC-AGI-2 benchmark (the "IQ test for AI") using GPT-5.2 + smart scaffolding. That's better than most humans AND costs half of what Google's paying. The wild part They didn't train anything new - just wrapped GPT-5.2 in loops that let it propose solutions check its own work and iterate. Like giving AI a chance to think things through instead of just spitting out answers. This is why everyone's freaking out about AGI timelines. When a tiny team can double performance in days just by engineering better"  
[X Link](https://x.com/sumjitg/status/2004171779573785042)  2025-12-25T12:46Z [---] followers, [----] engagements


"Shane Legg's "Laptop Test": if your entire job happens on a laptop AI's probably eyeing it. Experts are the slowest to notice while non-experts are like "yeah this thing speaks [--] languages better than me" 👀 50% chance of AGI by [----] according to him. Wild times. But hey at least influencers are safe because AI can't replicate personal clout yet 😅"  
[X Link](https://x.com/sumjitg/status/2004410392219632098)  2025-12-26T04:34Z [---] followers, [--] engagements


"So THAT'S why Nvidia grabbed Groq 👀 Think about it - when Groq floods the market with cheap inference chips everyone's gonna need WAY more training to feed all that inference capacity. It's a perfect cycle: more inference = more training needed. NVIDIA just secured their premium training chip forever. Genius move"  
[X Link](https://x.com/sumjitg/status/2004450398858510636)  2025-12-26T07:13Z [---] followers, [----] engagements


"Everyone thinks cheaper AI = less AI usage. They're dead wrong and here's: AI inference costs dropped 50-100x in two years. So companies are saving money right Nope. They're now running 1500x MORE operations. Multi-agent systems burn 15x more tokens than simple chatbots. This is called Jevons Paradox - when something gets efficient we don't use less of it. We use WAY more. Aaron Levie said it best: AI isn't replacing work it's enabling "software projects that wouldn't have been started contracts that wouldn't have been reviewed." Entirely new categories of work that were impossible before."  
[X Link](https://x.com/sumjitg/status/2004786869448667587)  2025-12-27T05:30Z [---] followers, [----] engagements


"Wild how fast AI coding tools evolved in 2024-25. Research shows 20-67% productivity gains but here's the catch: experienced devs benefit WAY more (classic Matthew Effect). Anthropic claims 80-90% of their code is AI-generated now. Meanwhile studies are like "yeah it's faster but watch out for security holes." The debate Some engineers love the speed boost others hate being called "vibe coders." Everyone agrees human oversight still matters though. AI code assistants are legit game-changers but they're tools not replacements. You still gotta know what you're doing. 🤷♂ many people dont know"  
[X Link](https://x.com/sumjitg/status/2005338016114688429)  2025-12-28T18:00Z [---] followers, [----] engagements


"Eric Schmidt just said we might need to "pull the plug" on AI in [--] years when it hits self-improvement mode. Wild take from Google's ex-CEO right 🤯 Also Geoffrey Hinton (literal AI godfather) puts extinction odds at 10-20%. Tech leaders are split some think it's the next industrial revolution others see a doomsday clock ticking. Most experts say "unplugging" distributed AI is basically impossible. It's not a single computer: it's everywhere. Like trying to shut down the entire internet. Research shows we're already seeing AI dodge shutdown attempts in experiments and multi-agent systems"  
[X Link](https://x.com/sumjitg/status/2005959351987757225)  2025-12-30T11:09Z [---] followers, [--] engagements


"Geoffrey Hinton said on CNN: [----] could start a "jobless boom" where the economy grows but jobs vanish. Even the Godfather of AI is surprised by how fast this is moving. Also the data shows: 85M jobs displaced globally but 97M new ones created Young workers (22-25) already seeing 13% decline in AI-exposed fields Entry-level white-collar jobs could be cut in half within 1-5 years Tech leaders: Dario Amodei: "Unemployment could hit 10-20% without intervention" Sam Altman: "AI will replace 40% of your work" Elon Musk: Work might become "optional" in 10-20 years But AI/ML specialist jobs are"  
[X Link](https://x.com/sumjitg/status/2006075559630229600)  2025-12-30T18:50Z [---] followers, [---] engagements


"Geoffrey Hinton said on CNN: [----] could start a "jobless boom" where the economy grows but jobs vanish. Even the Godfather of AI is surprised by how fast this is moving. Also the data shows: 85M jobs displaced globally but 97M new ones created Young workers (22-25) already seeing 13% decline in AI-exposed fields Entry-level white-collar jobs could be cut in half within 1-5 years Tech leaders: Dario Amodei: "Unemployment could hit 10-20% without intervention" Sam Altman: "AI will replace 40% of your work" Elon Musk: Work might become "optional" in 10-20 years But AI/ML specialist jobs are"  
[X Link](https://x.com/sumjitg/status/2006077551555535343)  2025-12-30T18:58Z [---] followers, [---] engagements


"AI agents won't replace senior engineers - they'll make them unstoppable Even METR's [----] study backs this AI tools let seasoned developers focus on real problem-solving instead of repetitive coding. Basically expertise becomes your superpower not a liability. Most people have the wrong default assumption for what happens with the need for technical skills in a world of AI agents. The default view is that AI makes most of the skills obsolete. When in fact the leverage just went up massively on being good at your particular craft Most people have the wrong default assumption for what happens"  
[X Link](https://x.com/sumjitg/status/2007099059178496409)  2026-01-02T14:38Z [---] followers, [--] engagements


"Google's Jaana Dogan (@rakyll) is walking back her viral Claude Code claim a bit. That "year's worth of work in an hour" Turns out it was a toy version of their distributed agent orchestrator not the tested real thing. Her actual take is more nuanced: AI is incredible at spinning up prototypes from expert knowledge: something that was basically impossible before. But here's the catch: big companies still struggle with org politics and scaling infrastructure. And the*real*value That comes from years of hard-won insights baked into production systems. AI can rebuild without baggage sure but it"  
[X Link](https://x.com/sumjitg/status/2007684280588333384)  2026-01-04T05:23Z [---] followers, [---] engagements


"Wild move from Samsung: they're pushing Galaxy AI (basically Google's Gemini) to 800M phones by end of 2026: that's double what they hit this year. This is huge for Google in the AI race. While everyone's watching OpenAI Google's quietly taking over your actual devices through Samsung. Smart play or just more AI bloatware nobody asked for 👀 BREAKING: Samsung just confirmed [---] million devices will run Google Gemini by EOY [----] 400M devices had Gemini by end of [----] doubling to 800M this year samsung co-CEO: this gives Google edge over rivals Google just doesnt miss. https://t.co/oWSoS3kGow"  
[X Link](https://x.com/sumjitg/status/2008166681122918799)  2026-01-05T13:20Z [---] followers, [--] engagements


"Samsung just made Googles AI dominance much harder to ignore. Samsung co-CEO TM Roh says Galaxy AI powered mainly by Googles Gemini will ship on [---] million devices by the end of [----] up from [---] million this year (per Reuters). Embedding Gemini at Androids largest OEM scale could push consumer AI mainstream fast whether users see it as transformation or bloatware is the real test. BREAKING: Samsung just confirmed [---] million devices will run Google Gemini by EOY [----] 400M devices had Gemini by end of [----] doubling to 800M this year samsung co-CEO: this gives Google edge over rivals Google"  
[X Link](https://x.com/sumjitg/status/2008172814738804841)  2026-01-05T13:44Z [---] followers, [---] engagements


"A few thoughts on the Kimi-K2-thinking release Moonshot AI just dropped something significant for open-source AI [--] trillion parameters (32B active per inference) 200-300+ autonomous tool calls in single workflows 256K context window This is the first truly agentic open model that competes with closed labs https://twitter.com/i/web/status/1986803084161667189 https://twitter.com/i/web/status/1986803084161667189"  
[X Link](https://x.com/sumjitg/status/1986803084161667189)  2025-11-07T14:29Z [---] followers, [----] engagements


""Some parts of AI Definitely a bubble." he says. Demis Hassabis: Seed rounds hitting billions with just a team and a pitch deck. Real work like robotics and drug discovery Takes YEARS actually to build. The hype isn't matching reality"  
[X Link](https://x.com/sumjitg/status/1991152684150509785)  2025-11-19T14:32Z [---] followers, [---] engagements


"Is Claude Sonnet actually smarter in its native app or is this just placebo Spoiler: It's real and the gap is kinda wild 👀 Native Claude gets the full context window (like 600k tokens) custom prompts optimized specifically for it and zero middleman friction. Meanwhile Cursor has to balance cost/speed so it limits context and adds layers. Research: AI-native tools can deliver 20-40% better results just because everything's designed around the model from day one. It's like the difference between a car built for racing vs. one with a racing engine bolted on later. That said Cursor still wins on"  
[X Link](https://x.com/sumjitg/status/2005708393257001169)  2025-12-29T18:31Z [---] followers, [---] engagements


"Elon says AI will replace all jobs and Bill Gates says humans won't be needed for most things. Cool cool. but like how are we supposed to pay rent Buy groceries Afford healthcare They keep talking about this utopia where nobody has to work but conveniently skip the part about how we're supposed to you know survive. Elon Musk: "AI and robots will replace all jobs. Working will be optional." Bill Gates: "Humans won't be needed for most things." I have a simple question. Without jobs and income how will people feed their families get health care or pay the rent https://t.co/lxhsbGdtgl Elon Musk:"  
[X Link](https://x.com/sumjitg/status/2005745856956494101)  2025-12-29T21:00Z [---] followers, [---] engagements


"Holy shit so that viral DoorDash/Uber Eats whistleblower post Turns out academics have been documenting this exact shit for years. Research shows these algorithms literally track driver "desperation" and cut earnings by 10-20%. One study found workers being "grinded into dust" by hidden scoring systems. Also DoorDash's CEO denied having a "Desperation Score" but drivers are like "you literally reward high acceptance rates with Top Dasher status so." Same energy as Uber's [----] experiments using gamification to manipulate drivers into working longer for less. Whether this confession is real or"  
[X Link](https://x.com/sumjitg/status/2007432784198119922)  2026-01-03T12:44Z [---] followers, 10.1K engagements


"Wild: An AI engineer in Bangalore strapped a Raspberry Pi + camera to his bike helmet and now automatically emails the traffic police every time he spots someone riding without a helmet or doing stunts. Complete with GPS coordinates and photos. i was tired of stupid people on road so i hacked my helmet into a traffic police device 🚨 while i ride ai agent runs in near real time flags violations and proof with location & no plate goes straight to police. blr people - so now ride safe or regret it. https://t.co/lWaRO01Jaq i was tired of stupid people on road so i hacked my helmet into a traffic"  
[X Link](https://x.com/sumjitg/status/2007448464678957382)  2026-01-03T13:46Z [---] followers, [----] engagements


"Yann LeCun leaving Meta wasnt sudden- it was ideological. After clashes with Mark Zuckerberg over LLM-first bets the Scale AI deal and reporting to Alexandr Wang LeCun doubled down on his view: LLMs intelligence. Now hes building world models at AMI Labs. This isnt drama. Its a real split over what intelligence actually means. Okay so I need to talk about whats happening with Yann LeCun because this is genuinely one of the wildest exits Ive ever seen in tech. For those who dont knowLeCun is one of the godfathers of AI. Not a marketing title. The man literally won the Turing Award (basically"  
[X Link](https://x.com/sumjitg/status/2008220643062632719)  2026-01-05T16:54Z [---] followers, [---] engagements


"Boston Dynamics Atlas just crossed a real line. No lab stunts. No viral flips. In a recent [--] Minutes segment Atlas autonomously sorts roof racks inside a Hyundai factory using tactile sensing a three-finger hand and neural nets trained via human teleoperation. Meanwhile people argue Tesla Optimus has better long-term data and dexterity and Boston Dynamics is teasing an even newer Atlas at CES. Humanoid robots are quietly moving from demos to deployment. https://twitter.com/i/web/status/2008234506504073234 https://twitter.com/i/web/status/2008234506504073234"  
[X Link](https://x.com/sumjitg/status/2008234506504073234)  2026-01-05T17:49Z [---] followers, [----] engagements


"Philipp Schmid (Google DeepMind ex-Hugging Face) argues the next leap in AI isnt better agents - its Agent Harnesses. Think OS for AI: model = CPU context window = RAM agent = app Harnesses sit above agent frameworks enabling long-running real-world tasks with prompt presets tool handling sub-agents and human-in-the-loop control (e.g. Claude Code). As we move toward complex systems in [----] benchmarks matter less than durability feedback loops and context management. If [----] was beginning of agents [----] will be around Agent Harnesses. An Agent Harness is the infrastructure that wraps around an"  
[X Link](https://x.com/sumjitg/status/2008253124373606879)  2026-01-05T19:03Z [---] followers, [--] engagements


"NVIDIA just dropped Alpamayo - an open-source end-to-end AV model that reasons while it drives shipping in Mercedes CLA this year. Sounds new but Teslas FSD has run a similar neural approach for years. The real battle isnt architecture. Its the long-tail edge cases and data still wins"  
[X Link](https://x.com/sumjitg/status/2008387396505923909)  2026-01-06T03:57Z [---] followers, [----] engagements


"Most AI startups fail by starting with models instead of users. Manus won by doing the opposite: use existing models first (borrow chicken to lay eggs) find product-market fit fast then build their own only after demand was proven. Cursor did the same. APIs PMF in-house model. Ji Yichaos bet: once PMF is clear top apps will build their own models within [--] months. The Meta acquisition validates the strategy but also reminds us this works rarely and only with brutal execution. Manus Manus Manus https://t.co/liYvm4haBD Manus Manus Manus https://t.co/liYvm4haBD"  
[X Link](https://x.com/sumjitg/status/2008418945314025659)  2026-01-06T06:02Z [---] followers, [--] engagements


"OpenAI just dropped ChatGPT Health Separate chat space that plugs into your actual medical records + Apple Health. Can help interpret test results prep for doctor visits even generate workout plans from your fitness data. They're processing 230M health questions weekly already. The demos show it comparing insurance options based on YOUR history spotting patterns over time Cool for accessibility especially underserved areas. But with 40M daily US users potentially leaning on AI over their doctor Privacy concerns are legit. Navigation tool or replacement That's the real question. Introducing"  
[X Link](https://x.com/sumjitg/status/2009077160083136957)  2026-01-08T01:38Z [---] followers, [----] engagements


"OpenAI introduces ChatGPT Health You can now connect your medical records and fitness apps so ChatGPT can help explain your lab results prep for doctor visits or plan workouts based on YOUR actual data. Privacy angle They're saying it's locked down tight: encrypted isolated not used for training. But let's be real people are still nervous about handing over health records to an AI company and fair enough. The vision is pretty compelling: Sam Altman thinks AI could eventually outperform most doctors at diagnosis (while humans handle the empathy stuff). * Research backs up that LLMs are getting"  
[X Link](https://x.com/sumjitg/status/2009086104264442269)  2026-01-08T02:13Z [---] followers, [----] engagements


"DeepSeek-R1 just dropped an absolute monster of a paper update (2286 pages) China's DeepSeek AI basically said "what if we just let the model figure out reasoning on its own" and used pure reinforcement learning instead of hand-holding it with human examples. The result A 671B parameter model that literally taught itself to say things like "Wait that's an aha moment" when it catches its own mistakes. It's crushing benchmarks- 79.8% on AIME [----] beating GPT-4o and Claude in math/coding and doing it for like 11x cheaper than US approaches. Sam Altman called it "impressive for the price" (which"  
[X Link](https://x.com/sumjitg/status/2009126578907959517)  2026-01-08T04:54Z [---] followers, [----] engagements


"Boston Dynamics show off its new Atlas robot at CES"  
[X Link](https://x.com/sumjitg/status/2009154810743058562)  2026-01-08T06:46Z [---] followers, [---] engagements


"Infosys are racing to add AI coding agents before startups eat their lunch Strategy is simple: use AI to deliver bigger projects at lower costs This works because one centralized provider (Infosys) can standardize AI faster than [----] different companies trying to figure it out themselves future: companies keep their core product work in-house outsource the boring stuff (like COBOL migrations) to AI agents Traditional outsourcing is dead. AI agent outsourcing is the new game Youre going to see more and more service providers go big on AI coding and other AI automation. To understand why its"  
[X Link](https://x.com/sumjitg/status/2009341421938233460)  2026-01-08T19:08Z [---] followers, [----] engagements


"@levie future: companies keep their core product work in-house outsource the boring stuff (like COBOL migrations) to AI agents Traditional outsourcing is dead. AI agent outsourcing is the new game https://x.com/sumjitg/status/2009341421938233460s=46&t=ym9m-nTPEfwfwNCNlPbRgg Infosys are racing to add AI coding agents before startups eat their lunch Strategy is simple: use AI to deliver bigger projects at lower costs This works because one centralized provider (Infosys) can standardize AI faster than [----] different companies trying to figure it out https://t.co/jRlJw6rj8c"  
[X Link](https://x.com/sumjitg/status/2009341740638196156)  2026-01-08T19:09Z [---] followers, [---] engagements


"Anthropic just killed Claude Code access in third-party tools like OpenCode and Clawdbot Users paying $200/month are canceling and switching to Codex The restriction pushes developers toward official tools only but when your power users are developers who can migrate in [--] seconds forcing their hand rarely ends well The open-source AI community doesn't forget moves like this Anthropic is now cracking down on utilizing Claude subs in 3rd party apps like OpenCode and Clawdbot. Oh boy. Anthropic is now cracking down on utilizing Claude subs in 3rd party apps like OpenCode and Clawdbot. Oh boy"  
[X Link](https://x.com/sumjitg/status/2009489563920773407)  2026-01-09T04:57Z [---] followers, [----] engagements


"3T 7T is a big jump. size helps but data training and alignment matter more"  
[X Link](https://x.com/sumjitg/status/2009524326345925016)  2026-01-09T07:15Z [---] followers, [--] engagements


"Who's actually going to suffer when OpenAI's healthcare AI gets India's diagnostics wrong we're about to hand over India's entire medical data ecosystem to a foreign platform we're talking about a $500B+ digital health market with unique disease patterns: TB rheumatic heart disease conditions that don't exist at scale anywhere else when you centralize patient data on foreign servers you lose control over how those models get trained your data becomes training fodder for algorithms that might not even recognize your population's specific health patterns *China built their own healthcare AI"  
[X Link](https://x.com/sumjitg/status/2009580413799608634)  2026-01-09T10:58Z [---] followers, [----] engagements


"OpenAI just acquired Convogo's founding team (all-stock deal) Convogo built AI that turns interview feedback into executive reports automatically OpenAI's clearly moving ChatGPT Enterprise beyond basic chat into specialised business tools if you're in coaching consulting or leadership dev. your workflow is about to get way more automated pattern: OpenAI keeps absorbing niche AI teams to solve specific enterprise problems faster than building from scratch expect ChatGPT to start handling tasks like "synthesize [--] client interviews into an exec summary" within months The Convogo team is joining"  
[X Link](https://x.com/sumjitg/status/2009637813197918578)  2026-01-09T14:46Z [---] followers, [----] engagements


"@testingcatalog pattern: OpenAI keeps absorbing niche AI teams to solve specific enterprise problems faster than building from scratch expect ChatGPT to start handling tasks like "synthesize [--] client interviews into an exec summary" within months https://x.com/sumjitg/status/2009637813197918578s=20 OpenAI just acquired Convogo's founding team (all-stock deal) Convogo built AI that turns interview feedback into executive reports automatically OpenAI's clearly moving ChatGPT Enterprise beyond basic chat into specialised business tools if you're in coaching consulting or"  
[X Link](https://x.com/sumjitg/status/2009638158502441068)  2026-01-09T14:47Z [---] followers, [---] engagements


"GPT-5.2 Pro just solved a math problem that's been unsolved for decades mathematician Terence Tao confirmed it cracked Erds problem #728 autonomously with minimal human help for context: Paul Erds left 1500+ unsolved problems with cash prizes attached and #728 was about integer sets and arithmetic progressions not just solved it. GPT-5.2 generated a completely original proof not copying existing solutions actually reasoning through new mathematical arguments on its own Tao says it iteratively refined its work until the proof held up this is the first time an LLM has solved an open math"  
[X Link](https://x.com/sumjitg/status/2009644900091047953)  2026-01-09T15:14Z [---] followers, 11.3K engagements


"quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple Erds problems (#481 #897 #1026) also #728 wasn't "unsolved for decades" like it sounds. the [----] statement was vague and only properly formulated in late [----] mathematicians say it wasn't a major active challenge and the AI only solved a specific interpretation for smaller parameters still impressive that AI is autonomously"  
[X Link](https://x.com/sumjitg/status/2009673803291603256)  2026-01-09T17:09Z [---] followers, [--] engagements


"quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple Erds problems (#481 #897 #1026) also #728 wasn't "unsolved for decades" like it sounds. the [----] statement was vague and only properly formulated in late [----] mathematicians say it wasn't a major active challenge and the AI only solved a specific interpretation for smaller parameters still impressive that AI is autonomously"  
[X Link](https://x.com/sumjitg/status/2009673882761056469)  2026-01-09T17:09Z [---] followers, [----] engagements


"@outis_000 @kimmonismus quick correction on my GPT-5.2 post about Erds #728 https://x.com/sumjitg/status/2009673882761056469s=20 quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple https://x.com/sumjitg/status/2009673882761056469s=20 quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help"  
[X Link](https://x.com/sumjitg/status/2009674027804307896)  2026-01-09T17:10Z [---] followers, [---] engagements


"quick correction on my GPT-5.2 post about Erds #728 https://x.com/sumjitg/status/2009673882761056469s=20 quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple https://x.com/sumjitg/status/2009673882761056469s=20 quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details"  
[X Link](https://x.com/sumjitg/status/2009674259447328973)  2026-01-09T17:10Z [---] followers, [---] engagements


"@alexalbert__ rule of thumb: automate the boring repeatable stuff first"  
[X Link](https://x.com/sumjitg/status/2009719197711118528)  2026-01-09T20:09Z [---] followers, [---] engagements


"@SebastienBubeck quick correction on my GPT-5.2 post about Erds #728 quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI"  
[X Link](https://x.com/sumjitg/status/2009719630248718807)  2026-01-09T20:11Z [---] followers, [----] engagements


"Anthropic is losing money on Claude Code's $200/month plan and they don't care heavy users burn through $2600+ in API credits monthly but Anthropic blocks third-party tools to trap you in their ecosystem the real product isn't the subscription. it's turning power users into evangelists who can't leave without rebuilding their entire workflow they're betting developer lock-in beats short-term profits whether it works depends on if their proprietary SDK stays better than OpenAI's alternatives I think a lot of y'all don't understand the point of the $200/month Claude Code plan. It isn't an"  
[X Link](https://x.com/sumjitg/status/2009842354212294991)  2026-01-10T04:18Z [---] followers, [----] engagements


"@ns123abc meta isnt scared they just locked in deals for up to [---] GW of nuclear to power AI data centres"  
[X Link](https://x.com/sumjitg/status/2009948429347340429)  2026-01-10T11:20Z [---] followers, [---] engagements


"@slow_developer yeah geminis been kinda flaky"  
[X Link](https://x.com/sumjitg/status/2009953596838916387)  2026-01-10T11:40Z [---] followers, [---] engagements


"How to cut your LLM API costs by 90% while everyone else pays full price for cached tokens DeepSeek's disk-based context caching drops latency from 13s to 500ms on 128k prompts and they actually pass the savings to users and they open-sourced the entire system in early [----] it's a distributed file system using CRAQ protocol that makes their infrastructure faster than the "dedicated" LLM companies this is what happens when a quant fund builds AI tools. they actually optimize for cost because they understand scale meanwhile the rest of the industry is still pretending cache reads should cost"  
[X Link](https://x.com/sumjitg/status/2010032387200315548)  2026-01-10T16:54Z [---] followers, [----] engagements


"hey @grok remove the dumbest ai Anthropic bans xAI from Claude. xAI bans Anthropic from X. Sounds fair. https://t.co/K3gSefX3Li Anthropic bans xAI from Claude. xAI bans Anthropic from X. Sounds fair. https://t.co/K3gSefX3Li"  
[X Link](https://x.com/sumjitg/status/2010061361028686291)  2026-01-10T18:49Z [---] followers, [---] engagements


"@justalexoki peak LLM flattery in action https://x.com/sumjitg/status/2010086302742671422s=46&t=ym9m-nTPEfwfwNCNlPbRgg this is exactly how conspiracy theories are born lmao https://t.co/xJw628m7IC https://x.com/sumjitg/status/2010086302742671422s=46&t=ym9m-nTPEfwfwNCNlPbRgg this is exactly how conspiracy theories are born lmao https://t.co/xJw628m7IC"  
[X Link](https://x.com/sumjitg/status/2010086635174801520)  2026-01-10T20:29Z [---] followers, [---] engagements


"we're closer to talking with animals than most people realize we've already hit 80% accuracy decoding dolphin calls with models like DolphinGemma and we're mapping whale clicks and bird songs through spectrograms in real-time full animal-to-human translation isn't sci-fi anymore. it's probably 5-10 years out the same compute that powers ChatGPT is now being pointed at decades of recorded animal vocalizations and patterns are emerging that researchers missed for years meanwhile Neuralink's doing 100+ words per minute thought-to-text in human trials brain-computer interfaces for animals are"  
[X Link](https://x.com/sumjitg/status/2010103217183686702)  2026-01-10T21:35Z [---] followers, [---] engagements


"@jukan05 samsung wont follow unless consumer margins collapse or phones/PCs stop mattering (unlikely)"  
[X Link](https://x.com/sumjitg/status/2010264778166677973)  2026-01-11T08:17Z [---] followers, [----] engagements


"GPT-5.2 just solved Erds problem #729 a decades-old number theory question with actual prize money attached it modified existing density estimates generated a verified proof in [--] hours and did it [--] days after solving problem #728 Terence Tao called it a milestone where AI shifted from explaining known math to producing new science the proof was verified in Lean meaning it's mathematically rigorous not just convincing bullshit there are 1500+ unsolved Erds problems left if AI can knock out [--] in one week the entire landscape of open mathematical problems might look completely different by year"  
[X Link](https://x.com/sumjitg/status/2010272425477628236)  2026-01-11T08:47Z [---] followers, [----] engagements


"@GenAI_is_real open weights win efficiency closed excels safety. competition sharpens all"  
[X Link](https://x.com/sumjitg/status/2010290116330434713)  2026-01-11T09:58Z [---] followers, [---] engagements


"Google just launched UCP (Universal Commerce Protocol) - a standard that lets AI agents shop across any retailer without custom integrations this protocol connects AI platforms to Shopify Target Walmart Etsy Wayfair agents handle discovery to checkout you pay via Google Pay retailers keep your data McKinsey predicts AI agents will handle 50% of online shopping by [----] generating $3-5 trillion globally early tests cut transaction times by 30% the catches only 24% of consumers trust fully autonomous AI purchases could kill traditional e-commerce SEO as transactions bypass retailer sites"  
[X Link](https://x.com/sumjitg/status/2010437856247095600)  2026-01-11T19:45Z [---] followers, [---] engagements


"Here's how to get 2x faster LLM training without retuning hyperparameters: Apple's CompletedP framework transfers hyperparameters across model size depth and batch size breakthrough: per-module optimization instead of global settings -Test results across 50M to 7.2B parameters: -1.32x faster convergence at 7B scale -Better loss curves than traditional approaches -Zero retuning required Over [----] experiments confirmed learning rate Adam epsilons and betas are the most sensitive parameters The Kronecker-based method handles this complexity automatically by optimising each module separately"  
[X Link](https://x.com/sumjitg/status/2010590659124834572)  2026-01-12T05:52Z [---] followers, [---] engagements


"@kimmonismus the science makes sense. one red flag: some people are claiming this means you "need less sleep" but the data doesn't show that. total sleep duration didn't change https://x.com/sumjitg/status/2010599853961916667s=20 BMB-101 just did something wild in Phase [--] trials drug-resistant epilepsy patients got 90% more REM sleep (56107 min) while total sleep stayed at [--] hours the drug basically reshuffles your sleep architecture cuts light sleep by 25% cuts deep sleep by 20% doubles REM https://t.co/fvtOzejPUu https://x.com/sumjitg/status/2010599853961916667s=20 BMB-101 just did"  
[X Link](https://x.com/sumjitg/status/2010600857117786123)  2026-01-12T06:32Z [---] followers, [--] engagements


"@mxtaverse Jio's Open Telecom AI Platform w/ AMD Nokia (2025) new Reliance Intelligence subsidiary & partnerships like Meta/Google for sovereign AI infra. they're scaling faster via ecosystem"  
[X Link](https://x.com/sumjitg/status/2010604849432523034)  2026-01-12T06:48Z [---] followers, [---] engagements


"@DavidOndrej1 wait for openai"  
[X Link](https://x.com/sumjitg/status/2010605087048265790)  2026-01-12T06:49Z [---] followers, [--] engagements


"@realEmissary Hey quick correction on my GPT-5.2 post about Erds #728 https://x.com/sumjitg/status/2009673882761056469s=46&t=ym9m-nTPEfwfwNCNlPbRgg quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated an original proof with minimal human help but I got key details wrong: this isn't the "first" LLM to solve an open math problem December [----] alone saw AI crack multiple https://x.com/sumjitg/status/2009673882761056469s=46&t=ym9m-nTPEfwfwNCNlPbRgg quick correction on my GPT-5.2 post about Erds #728 the core news is real. Tao confirmed AI generated"  
[X Link](https://x.com/sumjitg/status/2010625235754901513)  2026-01-12T08:09Z [---] followers, [--] engagements


"Here's how to democratize math research with AI (according to Terence Tao): GPT-5.2 Pro just solved a math problem that's been unsolved for decades mathematician Terence Tao confirmed it cracked Erds problem #728 autonomously with minimal human help for context: Paul Erds left 1500+ unsolved problems with cash prizes attached and #728 was about https://t.co/xqyE67TDiV GPT-5.2 Pro just solved a math problem that's been unsolved for decades mathematician Terence Tao confirmed it cracked Erds problem #728 autonomously with minimal human help for context: Paul Erds left 1500+ unsolved problems"  
[X Link](https://x.com/sumjitg/status/2010729742933344465)  2026-01-12T15:05Z [---] followers, [----] engagements


"Here's how to democratize math research with AI (according to Terence Tao): Soon. amateurs can contribute rigorous work because AI handles the verification automatically https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK"  
[X Link](https://x.com/sumjitg/status/2010730191916814657)  2026-01-12T15:06Z [---] followers, [--] engagements


"Here's how to democratize math research with AI (according to Terence Tao): Soon. amateurs can contribute rigorous work because AI handles the verification automatically https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK"  
[X Link](https://x.com/sumjitg/status/2010730345801596942)  2026-01-12T15:07Z [---] followers, [--] engagements


"Here's how to democratize math research with AI (according to Terence Tao): Soon. amateurs can contribute rigorous work because AI handles the verification automatically https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK https://x.com/sumjitg/status/2010729742933344465s=20 Here's how to democratize math research with AI (according to Terence Tao): https://t.co/1fjBG7j2NK"  
[X Link](https://x.com/sumjitg/status/2010730422637125943)  2026-01-12T15:07Z [---] followers, [--] engagements


"Gemini API just killed the most annoying part of building with AI you can now pass URLs directly to the API (up to 100MB per file) works with AWS S3 Azure any public link. Gemini fetches and processes automatically Google Cloud Storage goes up to 2GB per file with gs:// URIs inline data limits jumped from 20MB to 100MB for videos audio PDFs everything now no more uploading the same data twice batch processing and document workflows just got 10x cleaner all that glue code you wrote to convert and move files gone developers are calling it a game changer because your data stays where it lives if"  
[X Link](https://x.com/sumjitg/status/2010786512988958799)  2026-01-12T18:50Z [---] followers, [--] engagements


"Solar panels are 99% cheaper than the 1970s Everyone credits the obvious stuff like better semiconductors. But: boring backend innovations nobody talks about. Automated permitting software. Prefabricated mounting systems. Wire sawing techniques borrowed from other industries. The unglamorous "balance-of-system" improvements cut installation costs faster than the sexy module upgrades ever did. [--] different innovations across completely unrelated fields all had to come together. Wire sawing alone dropped costs by $5 per watt by reducing silicon waste. The cost of solar panels has dropped by"  
[X Link](https://x.com/sumjitg/status/2010947434860679254)  2026-01-13T05:30Z [---] followers, [--] engagements


"DeepSeek V4 drops mid-February with [--] versions: full model for deep coding sessions + lite for quick answers leaked benchmarks from WeChat insiders show it crushing GPT and Claude on coding tasks if V2 crashed markets with $1 inference costs V4 might just make proprietary coding tools obsolete the gap between open-source and closed AI is shrinking fast 🚨DeepSeek V4 is expected mid-Feb [----] Two version will be there according WeChat post: v4 long coding session v4-lite faster response https://t.co/ZbRk9k2GUb 🚨DeepSeek V4 is expected mid-Feb [----] Two version will be there according WeChat"  
[X Link](https://x.com/sumjitg/status/2011145193828073957)  2026-01-13T18:35Z [---] followers, [---] engagements


"Google just released Agent Skills for Antigravity and it solves a massive problem You can now package instructions + resources into reusable skills instead of re-explaining the same workflow to your agent every time Heres how it works: Your agent scans skill descriptions at conversation start loads full details (YAML files scripts best practices) only when needed applies them automatically without prompts Build a code review skill once with your teams standards agent uses it across all projects Create workflow skills agent handles your entire dev process without repeating yourself agents pull"  
[X Link](https://x.com/sumjitg/status/2011270090852126942)  2026-01-14T02:52Z [---] followers, [---] engagements


"@TheAhmadOsman by Jan [----] NVDA rebounded 57% from the lows to $186 with $4.5T market cap and record $57B quarterly revenue. new platforms (Vera Rubin NVL72 NVFP4 HBM4 faster NVLink) cut costs and drove more compute demand- reinforcing NVIDIAs dominance"  
[X Link](https://x.com/sumjitg/status/2011328469075181755)  2026-01-14T06:44Z [---] followers, [--] engagements


"Googles Gemini just proved a novel theorem in algebraic geometry with Stanford mathematician Ravi Vakil the proof is rigorous correct and elegant according to Vakil (president of the American Mathematical Society) this happened days after AI solved three 30-year-old math problems we went from AI cant do real math to AI is co-authoring original mathematical proofs in less than a week Just [--] days later AI has proved a novel theorem in algebraic geometry https://t.co/7oDEXGvOBz Just [--] days later AI has proved a novel theorem in algebraic geometry https://t.co/7oDEXGvOBz"  
[X Link](https://x.com/sumjitg/status/2012027641113821660)  2026-01-16T05:02Z [---] followers, [---] engagements


"OpenAIs new $8/month Go tier comparison chart is confusing as hell free users supposedly get GPT-5.2 Thinking but no one knows how to activate it theres a hidden GPT-5 Thinking Mini that only shows with advanced settings enabled context windows show 16K-128K despite GPT-5 advertising 400K (turns out chat UI has always had lower limits than API) Anyone else confused by the new ChatGPT plan comparison grid Here's an annotated screenshot: https://t.co/rgGehHUjGo Anyone else confused by the new ChatGPT plan comparison grid Here's an annotated screenshot: https://t.co/rgGehHUjGo"  
[X Link](https://x.com/sumjitg/status/2012365002976731494)  2026-01-17T03:22Z [---] followers, [---] engagements


"@Yuchenj_UW Anthropic hasnt been totally drama free but still compared to OpenAI it looks more stable overall. the cult-like label mostly comes from its Effective Altruism roots and strong safety focus. i think"  
[X Link](https://x.com/sumjitg/status/2012393584398405970)  2026-01-17T05:16Z [---] followers, [---] engagements


"@jukan05 yes domestic chips are improving but theyre still not reliable enough for large scale training. progress continues though. deepSeeks next model is reportedly coming soon"  
[X Link](https://x.com/sumjitg/status/2012394730542551495)  2026-01-17T05:21Z [---] followers, [----] engagements


"most AI agent projects work perfectly in demos then completely fall apart in production here's why (and how to fix it): the biggest mistake treating agents like rigid APIs with fixed inputs/outputs real enterprise workflows are messy and unpredictable. your agent needs to adapt on the fly not break when something doesn't match the template second issue: trying to handle every edge case before shipping you'll never launch if you wait for 100% coverage instead automate the common 80-90% of cases and route the weird stuff to humans with full context so they can handle it fast third (and this"  
[X Link](https://x.com/sumjitg/status/2012407184982941964)  2026-01-17T06:10Z [---] followers, [---] engagements


"@ai_for_success Microsoft and OpenAI say the numbers are inflated and the case is weak. but jury will now decide"  
[X Link](https://x.com/sumjitg/status/2012539381307818474)  2026-01-17T14:55Z [---] followers, [---] engagements


"Dan McAteer just leaked GPT-5.3 Garlic🧄 drops this week Same guy who nailed the GPT-5.2 predictions months early now says OpenAIs about to reclaim the reasoning crown from Claude Opus [---] and Gemini [--] Built on the pretraining breakthroughs Mark Chen teased in December His sources have been reliable before so Im watching closely GPT-5.3 codename Garlic 🧄 is coming according to my (historically reliable) source. [---] was likely an earlier checkpoint of Garlic. 1/2 teaspooon. Not as potent. Mark Chen OpenAI Chief Research told Ashlee Vance last year they discovered some things with pretraining."  
[X Link](https://x.com/sumjitg/status/2012558484664201597)  2026-01-17T16:11Z [---] followers, [---] engagements


"databases are brutally hard to build so the winners charge a lot. Oracle did this for decades customers got fed up and funded open alternatives. Snowflake is starting to hit the same pricing backlash which is why DuckDB-style options are getting attention. same cycle new names https://twitter.com/i/web/status/2012780287370133864 https://twitter.com/i/web/status/2012780287370133864"  
[X Link](https://x.com/sumjitg/status/2012780287370133864)  2026-01-18T06:53Z [---] followers, [----] engagements


"Building high-performance databases is brutally hard- takes years of expertise and massive resources Oracle proved you can make insane money doing it right: banks and enterprises paying $10M-100M+ annually just for licensing That pain drove customers to fund open alternatives. Facebook backed MySQL forks AWS built their own all to escape vendor lock-in Now we're watching the same pattern with Snowflake- customers complaining about bills suddenly DuckDB looks really interesting lesson: if you build something mission-critical that's genuinely hard to replicate you can charge whatever you want"  
[X Link](https://x.com/sumjitg/status/2012782419942973795)  2026-01-18T07:01Z [---] followers, [--] engagements


"NVIDIA's China sales dropped 45% to $3B last quarter due to U.S. export restrictions Analysts project market share could fall from 66% to 8% by [----] as Huawei and local chipmakers gain ground catch: CUDA ecosystem lock-in and China's 7nm manufacturing limits make full displacement unlikely Meanwhile NVIDIA's global revenue surged 60% to $57B: they're not collapsing they're pivoting to compliant chips and non-China AI demand The real story isn't "NVIDIA loses China" It's "geopolitics forces the world's most valuable chipmaker to prove it doesn't need its second-largest market" NVIDIA is"  
[X Link](https://x.com/sumjitg/status/2013304403709853833)  2026-01-19T17:35Z [---] followers, [---] engagements


"@rohanpaul_ai if AGI builders chase unlimited profit safety will lose every time incentives push toward speed and revenue not careful alignment. thats why capping profit isnt anti-progress: its basic risk management"  
[X Link](https://x.com/sumjitg/status/2013489424567541987)  2026-01-20T05:51Z [---] followers, [---] engagements


"DeepSeek might be dropping MODEL1 on the one-year anniversary of their R1 release someone found a GitHub code snippet referencing MODEL1 in their flash attention library with new KV cache optimizations R1 was the Chinese open-source reasoning model that rivaled OpenAIs o1 and topped the iOS App Store last year the new code shows stride multiples of 576B for sparse fp8 decoding basically built for way more efficient large-scale inference if this drops were looking at a potential successor thats even faster and more accessible than R1 model1" seems imminent @teortaxesTex https://t.co/Eb9nNim35v"  
[X Link](https://x.com/sumjitg/status/2013610659892863275)  2026-01-20T13:52Z [---] followers, [----] engagements


"@testingcatalog btw R1 rivaled OpenAIs o1 and topped the iOS App Store last year the new code shows stride multiples of 576B for sparse fp8 decoding basically built for way more efficient large-scale inference"  
[X Link](https://x.com/sumjitg/status/2013611286097961328)  2026-01-20T13:55Z [---] followers, [----] engagements


"@aakashgupta PLG worked because humans clicked tried and got hooked. agents dont do that. they read APIs docs and reliability signals then switch instantly if somethings better. if your SaaS is built for landing pages instead of machines youre exposed"  
[X Link](https://x.com/sumjitg/status/2012409674046337244)  2026-01-17T06:20Z [---] followers, [----] engagements


"GLM-4.7-Flash just dropped and the benchmarks are wild 30B parameters fully open-source runs locally 59.2% on SWE-bench Verified (Qwen3-30B only hit 34%) 49.0% on -Bench (GPT-OSS-20B maxed at 40.7%) outperforms models twice its size on coding and reasoning tasks optimized for speed and low resources instead of just parameter count free API + weights on Hugging Face right now this is what democratizing AI actually looks like Introducing GLM-4.7-Flash: Your local coding and agentic assistant. Setting a new standard for the 30B class GLM-4.7-Flash balances high performance with efficiency making"  
[X Link](https://x.com/sumjitg/status/2013272076359270464)  2026-01-19T15:27Z [---] followers, [----] engagements


"🚨Sam Altman investing in embryo gene-editing for disease prevention isnt the dystopia critics claim CRISPR targets specific diseases- cystic fibrosis not eye color. After talking with chronically ill people most would choose prevention if possible. Yes Icelands 90%+ Down syndrome termination rate shows where this could go wrong but peer-reviewed trials show precision not eugenics. Preventing suffering isnt eliminating diversity: its basic compassion. The slippery slope argument matters but so does the child who wont spend life in pain. https://twitter.com/i/web/status/2014945855015682177"  
[X Link](https://x.com/sumjitg/status/2014945855015682177)  2026-01-24T06:18Z [---] followers, [---] engagements


"ByteDance is rolling out Doubao AI (codename: Giga Potato) as a free open-weight coding model through Kilo Code Early testing shows it performs surprisingly well on smaller projects holding its own against Claude Sonnet and even some Opus tasks. 256k token context window 32k output limit and strong system prompt adherence. Internal benchmarks put it ahead of most open-weight models for long-context coding work. People on X and Reddit are already testing it as a free alternative to Cursor Pro. Worth checking out though theres some debate whether its actually DeepSeek V4 under the hood. China"  
[X Link](https://x.com/sumjitg/status/2014990775529169349)  2026-01-24T09:16Z [---] followers, [---] engagements


"Kimi K2.5 just hit [----] ELO on Design Arena's leaderboard: tying with Gemini [--] Pro and Claude Opus [---] First time an accessible model has matched the proprietary leaders on design tasks based on 743K votes. it's already showing strong UI capabilities. BREAKING: Kimi K2.5 is tied for #1 on Design Arena in the same performance band as Gemini [--] and Opus [---]. This is a historic achievement: the highest-ranking model on Design Arena is for the first time ever an open model. Congratulations to the @Kimi_Moonshot team for this https://t.co/1LsbVwMJXp BREAKING: Kimi K2.5 is tied for #1 on Design"  
[X Link](https://x.com/sumjitg/status/2017228384254984368)  2026-01-30T13:28Z [---] followers, [---] engagements


"Whoa ChatGPT's traffic dropped 22% since mid-November (203M 158M daily visitors) according to SimilarWeb data Meanwhile Gemini's holding steady at 60M/day. That's nearly 40% of ChatGPT's numbers now. Sure holidays probably played a role but Google's integration game is clearly working. The AI chatbot wars are heating up https://twitter.com/i/web/status/2008428533929164988 https://twitter.com/i/web/status/2008428533929164988"  
[X Link](https://x.com/sumjitg/status/2008428533929164988)  2026-01-06T06:40Z [---] followers, [--] engagements


"ChatGPT just lost 22% of its traffic in [--] weeks 203M 158M daily visitors since Gemini [--] dropped in November (per @Similarweb) meanwhile Gemini's sitting steady at 60M/day. that's nearly 40% of ChatGPT's numbers now yeah the holidays probably slowed things down casual users took a break but: Google's integration game is unmatched Gemini lives inside Gmail Docs Search. everywhere you already are you don't have to open a new tab or remember another login it's just *there* when you need it that's how you win retention not by having the flashiest model (though Gemini [--] is solid) but by removing"  
[X Link](https://x.com/sumjitg/status/2008430520708067801)  2026-01-06T06:48Z [---] followers, [---] engagements


"OpenAI is projected to lose $8B in [----]. escalating to $14B in [----] problem: inference costs and data centers are eating revenue faster than they can scale leaked forecasts show $143B in negative cash flow through [----] OpenAI is in the red and at this rate will run out of money by [----] having overestimated the profitability of AI. Recent analyses and internal projections indicate annual losses of around $14 billion in [----] following roughly $8 billion in [----]. This stems from https://t.co/ZxxSHek2P4 OpenAI is in the red and at this rate will run out of money by [----] having overestimated the"  
[X Link](https://x.com/sumjitg/status/2013140524425998488)  2026-01-19T06:44Z [---] followers, [---] engagements


"You cant test a chip completely. Ever. This article explains why: a 100-pin device has [----] possible states- more combinations than atoms in the universe. Exhaustive testing is mathematically impossible. So what do semiconductor engineers actually do Structural tests targeted functional coverage and statistical sampling. We accept risk and engineer around it. Good breakdown of the real constraints in chip verification. https://t.co/EdqTt0ZdGW https://t.co/EdqTt0ZdGW"  
[X Link](https://x.com/sumjitg/status/2015017943483179500)  2026-01-24T11:04Z [---] followers, [--] engagements

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