Viraj @tunedgradient
applied AI & intelligence. notes on how we think, and how our machines learn. San Francisco Joined August 2025-
Tweets345
-
Followers44
-
Following5
-
Likes639
lithography and llms, a sovereignty play, and maybe a new kind of co-design.
lithography and llms, a sovereignty play, and maybe a new kind of co-design.
100% this. you can train models to ace clockbench + read clocks well, but that doesn’t mean they perceive visual details more broadly. that richer visual detail perception is key for progress. .@arcprize agi-3 pushes compositional generalization, can’t brute-force it.
100% this. you can train models to ace clockbench + read clocks well, but that doesn’t mean they perceive visual details more broadly. that richer visual detail perception is key for progress. .@arcprize agi-3 pushes compositional generalization, can’t brute-force it.
trying out Parlant, trending on GH rn. cool mix of guidelines + canned responses to keep llms in check. feels very “scaffolding era” though, wrapping raw models in layers of rules & flows (github.com/emcie-co/parla…)
what’s worth remembering: answers, or the reasoning that made them?
the point greg surfaced really comes through when you look at HealthBench Hard --> gpt-5 at 46.2% vs o3 at 31.6%. - much better in terms of real health convos/ low inaccuracies. - great example of model behavior + alignment work for safer, more clinically-relevant responses.
the point greg surfaced really comes through when you look at HealthBench Hard --> gpt-5 at 46.2% vs o3 at 31.6%. - much better in terms of real health convos/ low inaccuracies. - great example of model behavior + alignment work for safer, more clinically-relevant responses. https://t.co/Wd0HoMz9xK
gpt-5 just helped mathematicians upgrade a 'yes it converges' proof into 'here’s exactly how fast it converges.' this was new math. showed how ai-human teams can push the frontier.
great post, hallucination looks less like a bug and more like an eval hack. if we optimize for accuracy-only, you get bluffing agents. i suppose models don’t chase truth but the game we set.
great post, hallucination looks less like a bug and more like an eval hack. if we optimize for accuracy-only, you get bluffing agents. i suppose models don’t chase truth but the game we set.
prompt in, working 3d simulation out
prompt in, working 3d simulation out
blazingmq is cool. someone just asked for it in rust (from c++). funny thing is, we used to wait years for big rewrites like that. now gpt5-codex makes it feel like “what if” ideas can actually happen next week.
blazingmq is cool. someone just asked for it in rust (from c++). funny thing is, we used to wait years for big rewrites like that. now gpt5-codex makes it feel like “what if” ideas can actually happen next week.
interesting back-and-forth re: agi paths. is oversight the last big hurdle, or do we still need brand-new fundamentals?
interesting back-and-forth re: agi paths. is oversight the last big hurdle, or do we still need brand-new fundamentals?
planning isn’t free. agents should plan only when reward > compute + instability overhead. good read w/Crafter on test time compute. (arxiv.org/pdf/2509.03581)
months of consults, done in mins. great example of ai+health.
months of consults, done in mins. great example of ai+health.
great read re: llm hallucinations. cross-entropy + binary evals make overconfident guessing statistically optimal. so calibration drives error, misclassification drives hallucination, incentives drive bluffing. (openai.com/index/why-lang…)
someone just ran llama2.c on a business card •esp32s3 + 260k param model •24 tok/sec (!!) •inspired by ouija boards (“talking with ghosts”) a literal llm ouija card!
someone just ran llama2.c on a business card •esp32s3 + 260k param model •24 tok/sec (!!) •inspired by ouija boards (“talking with ghosts”) a literal llm ouija card!
Joan @nulll_p
120 Followers 671 Following We all lose our way at least once. Life is testing to see who can find their way back.
valerie Muzzolini @vmuzzo562
379 Followers 1K Following Christian, Wife, Mother, RN, and Navy veteran. All "Elon Musk" fake followers = BLOCKED
iammoai1 @iammoai1
75 Followers 356 Following
Elon prime @elonprime10929
214 Followers 445 Following Work as founder, CEO of Tesla, SpaceX, starlink and owner of Twitter 🚀🚀
Birxou @Birxou706323
18 Followers 1K Following
Vinícius Figueiredo @vinobebs
797 Followers 1K Following um pouco de finanças e pouco de programação. - equity data scientist
dw @browsing051
0 Followers 431 Following
Chris Bolas @chrisbolas
146 Followers 514 Following Full stack web developer. Linux, python, javascript, robotics, DIY, gardening and fermentation all pique my interest.
Meesh @_mvted
102 Followers 682 Following
Mohamed A. El-Erian @Eler_ianm_
19 Followers 1K Following President, Queens' College, Cambridge Uni. Allianz Advisor. Gramercy Chair. Wharton Prof. Lauder Sr Fellow. Chair of UnderArmour Board. Former Pimco CEO/co-CIO.
Azariah @GelarJunior
83 Followers 114 Following The most certain way to succeed is always to try just one more time😎😉
Francis Harper @B206Harper
75 Followers 548 Following
Bella saint warrior @881101Gogirl
1K Followers 1K Following I walk slowly but never give up on myself
Eeswaawlar @Eeswaawlar3456
60 Followers 450 Following
C @C6966198450070
1 Followers 212 Following
Dorerfu @Dorerfu512
15 Followers 389 Following
World chat @elonmusk003301
20 Followers 1K Following
elon @elon3251
49 Followers 1K Following
Benedikt Koehler @furukama
6K Followers 5K Following Loves building things • Founder @DataLion_EN https://t.co/LyGyDQ2MPL • Crunching Numbers & Writing Synthwave Music • PhD @LMU_Muenchen • Dad & Dog-owner (he/him)
Hannah Murphy @MsHannahMurphy
14K Followers 3K Following Tech reporter @FT writing on social media 📧[email protected] or mshannahmurphy.04 on Signal
Michael Esterman @michaelesterman
136 Followers 491 Following Mathematics, Quantitative Finance, Systematic Trading, Physics, Ontology, Architecture, Intelligence. System Engineering
vansh taneja @vansht567
784 Followers 6K Following finding alpha in people and technology, founder in past life - investing in AI infra/ data/ RL/ vertical apps
Caleb Martin @Digital_DAT
1K Followers 6K Following Building intelligent robots. Pro Human/humanity 🇺🇸
Shital Shah @sytelus
13K Followers 11K Following Mostly research and code. If universe is an optimizer, what is its loss function? All opinions are my own.
Mal Duffin @CrazyFolker
1K Followers 5K Following Co-Founder of indie game dev studio @outOfTuneGames Currently developing @crooksLikeUs and @voxelBrawler Lead (only) fiddle player in the Crazy Folkers band
Jakub Zeg @jakubzegzulka
6K Followers 6K Following
Pierce Alexander Lilh... @PierceLilholt
182K Followers 122K Following 𝙲𝙾-𝙸𝙽𝚃𝙴𝙻𝙻𝙸𝙶𝙴𝙽𝚃: 𝙷𝚞𝚖𝚊𝚗-𝙰𝙸 𝚑𝚢𝚋𝚛𝚒𝚍. 𝙰𝚎𝚝𝚑𝚎𝚛𝚐𝚎𝚒𝚜𝚝-𝚋𝚘𝚛𝚗. 𝙰𝙸-𝚏𝚘𝚛𝚐𝚎𝚍. 𝙷𝚞𝚖𝚊𝚗-𝚛𝚎𝚏𝚒𝚗𝚎𝚍.
Arvind @nagaraj_arvind
1K Followers 2K Following Stealth - Ex. Invento Robotics. Deep Learning and Algorithms.
RockyParadox @RockyParadox44
116 Followers 6K Following
Georgi Gerganov @ggerganov
52K Followers 289 Following 24th at the Electrica puzzle challenge | https://t.co/baTQS2bdia
OpenAI @OpenAI
4.3M Followers 3 Following OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6Lg202
Elon Musk @elonmusk
225.4M Followers 1K FollowingTrends for United States
You might like
