Yam Peleg
@Yampeleg
Followers
39K
Following
28K
Media
3K
Statuses
15K
The only AI researcher they sent a missile for ๐ฎ๐ฑ | Co-host @thursdai_pod โข AI news every Thursday
Joined July 2012
Lore TL;DR: > spent summer in SF > missile hit my place back home > officially refugee, returned, officially displaced > snuck into collapse hazard building, rescued GPUs > patched together a working setup > technical posts coming back to x .com > apologies for the inconvenience
16
9
344
We just shipped the first VS Code extension built specifically for @Solana developers. Static analysis detectors that catch vulnerabilities in the IDE + fuzzing coverage visualization. The missing layer between development and audits, integrated into devs' workflow โ
9
22
89
PewDiePie went from Minecraft let's plays to building 10xGPU rigs, training 120B LLMs, distributed vLLM inference, coding webapps, RAGs, speech recognition pipelines, linux sysadmin.. Millions of kids are watching him compile CUDA now. 110M+ subs. WHAT AN ABSOLUTE LEGEND!!
173
1K
21K
Just found ๐๐๐๐๐๐-๐๐๐ is goated for literature reviews Ask an open research question, give it a little push toward arXiv, and it goes off batch calling its ๐๐ผ๐ผ๐ด๐น๐ฒ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต tool, reading papers, multi-step reasoning with 1M context.. It finds crazy insights
2
5
86
plz stop pretending you run โ1000 agents in parallelโ so the rest of us can code thx
0
0
28
wow this whole thing is even interactive and also explain details about all the different LLM architectures, this is above and beyond!!
1
0
3
hf are doing godโs work fr
We've just published the Smol Training Playbook: a distillation of hard earned knowledge to share exactly what it takes to train SOTA LLMs โก๏ธ Featuring our protagonist SmolLM3, we cover: ๐งญ Strategy on whether to train your own LLM and burn all your VC money ๐ชจ Pretraining,
4
6
126
LIVE: ๐ ThursdAI - Halloween Special! | Weekly AI News: SCARY FAST Agents (MiniMax M2, Cursor 2.0, SWE 1.5), OpenAI's ASI Plan, 1X NEO Humanoid
1
1
8
Iโve been saying this for years: Instead of causal attention, do: N ร N ร N = O(Nยณ) โฒ โฒ โฒ โ โ โ fully connected attention โ for every token up to current token less efficient training but insanely more powerful with same inference cost as today O(Nยฒ)
just out of vain curiousity ; what happens if you increase the complexity of attention? like, has anyone tried cubic attention lol
26
19
437
every time i see an emoji in a github readme i just assume itโs a grift and close the tab
2
0
8
man i can actually count on codex itโs wild told it โmake my training code fasterโ left it in yolo mode controlling 8x A100s and went to lunch lmao
2
0
9
TL;DR: โข SFT: Mimics teacher but learns only from teachers states โข RL: Explores on its own but gets sparse reward Solution: 1. Student generates trajectory 2. Teacher scores each token (logprobs) 3. Reward = -KL(ฯ_ฮธ || ฯ_teacher) per token 4. Policy gradient update Wow nice
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other
20
32
514
LIVE: ๐ ThursdAI | Weekly AI News | ChatGPT Atlast Browser, DeepSeek OCR, Claude Code Web, Browserbase Director
0
0
8
OpenAI shipping social networks, browsers, app stores, video gen, code editors, agents, frontier LLMs, open weights models, CLI tools... ...is NOT because some AI is vibe-coding like crazy behind the scenes. It's because their people are good.
In less than a month @OpenAI launched a totally new social media app / format AND a new browser. Can anyone remember a time when companies this size were shipping so fast? Itโs a new benchmark for the velocity of world class software teams.
1
2
26