Viraj @tunedgradient
applied AI & intelligence. notes on how we think, and how our machines learn. San Francisco Joined August 2025-
Tweets316
-
Followers37
-
Following5
-
Likes561
think of asics as coin flips, either the program hits at tape-out or it’s dead on arrival. gpus aren’t like that. every new gen compounds across training and inference (which is why nvidia's growth feels steadier)
think of asics as coin flips, either the program hits at tape-out or it’s dead on arrival. gpus aren’t like that. every new gen compounds across training and inference (which is why nvidia's growth feels steadier)
a lot of folks worry they’ll lose control if they move off chat completions. truth is, responses gives you more control, run it stateless, keep all history client-side, and still tap into reasoning performance. make the switch.
a lot of folks worry they’ll lose control if they move off chat completions. truth is, responses gives you more control, run it stateless, keep all history client-side, and still tap into reasoning performance. make the switch.
solid framing. optimizer papers are Rorschach tests: some will interpret them as AdamW’s robustness, others as insufficient hyperparameter search. in the end, results hinge more on methodology and tuning protocols than on optimizer 'magic'.
solid framing. optimizer papers are Rorschach tests: some will interpret them as AdamW’s robustness, others as insufficient hyperparameter search. in the end, results hinge more on methodology and tuning protocols than on optimizer 'magic'.
train agents to budget reasoning, plan when it helps, act when it doesn’t. stop overthinking by default.
train agents to budget reasoning, plan when it helps, act when it doesn’t. stop overthinking by default.
trying out kimi k2 - 1T params (32b active, MoE) - 256k context - runs on fp8 safetensors, plugs into openai apis MoE tricks mean only ~32B are active at once, and ‘lossless’ claims rarely hold up. still, the long-context coding angle is worth watching.
trying out kimi k2 - 1T params (32b active, MoE) - 256k context - runs on fp8 safetensors, plugs into openai apis MoE tricks mean only ~32B are active at once, and ‘lossless’ claims rarely hold up. still, the long-context coding angle is worth watching.
funny how we keep rediscovering this same truth every few years
chips in-house
commit density > commit count
commit density > commit count
basically qed by construction, the tools proved themselves!
basically qed by construction, the tools proved themselves!
branching chats now works like git for conversations. explore multiple directions, keep your main thread clean, and never lose context!
branching chats now works like git for conversations. explore multiple directions, keep your main thread clean, and never lose context!
used to think kv cache paging/eviction was simple. but the traces here are interesting: - chat reuse lasts mins, api reuse lasts secs. - for apis, a small gpu cache is plenty. - for chat, smarter eviction. - be super workload-aware. (arxiv.org/pdf/2506.02634)
codex patched vim’s hardest bug: learning it.
codex patched vim’s hardest bug: learning it.
not sure if they’ve heard of codex and gpt-5
not sure if they’ve heard of codex and gpt-5
FineVision is 🔥 - 200+ datasets pulled together, all cleaned & streamed from hf hub - 17M images + 10B tokens - benchmarks show ~20% gains over Cambrian, LLaVa, etc. ImageNet moment for VLM training data?
FineVision is 🔥 - 200+ datasets pulled together, all cleaned & streamed from hf hub - 17M images + 10B tokens - benchmarks show ~20% gains over Cambrian, LLaVa, etc. ImageNet moment for VLM training data?
half of community “finetunes” are really just LoRA adapters bolted onto frozen weights. and honestly, that works surprisingly well.
codex cli now formal in hf mcp server. one more piece of the puzzle for smooth workflows
codex cli now formal in hf mcp server. one more piece of the puzzle for smooth workflows
Elon prime @elonprime10929
212 Followers 454 Following Work as founder, CEO of Tesla, SpaceX, starlink and owner of Twitter 🚀🚀
Birxou @Birxou706323
16 Followers 1K Following
Vinícius Figueiredo @vinobebs
796 Followers 1K Following um pouco de finanças e pouco de programação. - equity data scientist
dw @browsing051
0 Followers 431 Following
Chris Bolas @chrisbolas
138 Followers 483 Following Full stack web developer. Linux, python, javascript, robotics, DIY, gardening and fermentation all pique my interest.
Meesh @_mvted
102 Followers 678 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
85 Followers 112 Following The most certain way to succeed is always to try just one more time😎😉
Francis Harper @B206Harper
77 Followers 548 Following
Bella saint warrior @881101Gogirl
1K Followers 974 Following I walk slowly but never give up on myself
Eeswaawlar @Eeswaawlar3456
60 Followers 447 Following
C @C6966198450070
1 Followers 212 Following
Dorerfu @Dorerfu512
14 Followers 361 Following
World chat @elonmusk003301
19 Followers 1K Following
elon @elon3251
43 Followers 871 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 486 Following Mathematics, Quantitative Finance, Systematic Trading, Physics, Ontology, Architecture, Intelligence. System Engineering
vansh taneja @vansht567
775 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 121K 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
114 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
225.3M Followers 1K FollowingTrends for United States
You might like
