srutanik

Research @iiserbhopal
Building. Running. Reiterating.

[Pramod Goyal @goyal__pramod]

One of the best blog on CUDA you will ever read!

[Palisade Research @PalisadeAI]

Palisade interviewed 22 current and former employees from OpenAI, DeepMind, and Anthropic about their personal views and fears around AI development.

Today, we’re releasing the first batch of those interviews.

Please watch and share.

[Vaishnavi @_vmlops]

Found a genuinely solid learning resource: 100 Days of LLM Inference by @elizabetht 🔥

A structured deep-dive covering the full stack CUDA kernels → vLLM/SGLang/TensorRT-LLM → quantization → speculative decoding → multi-cloud autoscaling.

Every entry is a runnable notebook, tested on a real 2-GPU home-lab cluster

[codila @0xCodila]

Google just released free 2-hour course on full Graph & Loop engineering: 1 prompt → 100 agents → loops → graphs from 0% to 100%:

0% → 0:35 - Graph engineering from scratch
30% → 31:17 - build your first agents graph
45% → 43:40 - run hundreds agents in parallel
75% → 1:04:58 - Loop engineering: route, check, repeat
100% → 1:30:09 - self-improving graphs that work while you sleep

most people keep stuffing instructions into one agent - the real upgrade is a graph that routes, checks, and rebuilds the workflow at runtime

learn graph from Google, ship your first - then unlock the complete system design below ↓

[Dark Factor @DarkFactorr]

this is free f*cking gold for anyone still gluing agents together with loops

Andrew Ng mapped 4 agentic patterns into one graph instead of four separate tricks

Reflection, Tool Use, Planning, Multi-Agent Collaboration, one flow

User → Architect Agent → Tech Lead Agent → Developer Agent, typed handoffs

feedback loops back to the user at every stage, not just at the end

GPT-3.5 wrapped in this style of workflow: 95.1% on HumanEval. GPT-4 zero-shot: 67%.

give this to your next agent build before you touch the model card again.

[Yacine Mahdid @yacinelearning]

we are going to see much more of these post-trained models inside harness imo

I’m preparing a tutorial series on post-training RLMs btw

[Troll Football @TrollFootball]

Barcelona 2026-27

[Daniel San @dani_avila7]

Spotify cut Claude Code token usage by 90% with a few simple changes

Here’s the quick breakdown:

- Keep Claude for tasks that actually require reasoning

- Route large file reads to cheaper models like Gemini 2.5 Flash

- Use PreToolUse hooks to automatically intercept expensive reads

- Return only the relevant context back to Claude

- Delegate boilerplate, tests, configs, and repetitive code

Don’t spend frontier model tokens on work that doesn’t require frontier model intelligence

Link: https://t.co/oAM150SUW1

[ℏεsam @Hesamation]

This is arguably one of the best technical books you can read on how LLMs are trained at scale:

> GPU memory and profiling
> tiling, kernel fusion, FlashAttention
> Data / Tensor / Pipeline / Context Parallelism

I’ve already read the online free version, but still had to get the physical copy for the library lol.

You can also read it here on @huggingface: https://t.co/zkFhABMOTw

[avrl ☘ @avrldotdev]

Anthropic's most frequently asked system design interview question to date.

[Sylvia Li @Sylviaposts]

Found a genuinely useful repo for anyone writing ML papers.@ChenLiu_1996

A Yale CS PhD shared the Python scripts behind figures from Nature Machine Intelligence, ICML, NeurIPS, and ECCV.

Maybe I can finally stop spending half my research life moving matplotlib legends by 3 pixels.

https://t.co/2Cx2f9mJYz

[Dwarkesh Patel @dwarkesh_sp]

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.

This culminated in the third one taking over part of OpenAI itself.

All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.

I’ve spent the last three days reading through these reports and trying to understand exactly what happened.

Here is my attempt to tell the whole story in plain English:

https://t.co/Nb2un9oNJR

[pdawg @prathamgrv]

https://t.co/UInlkJKONl

[pdawg @prathamgrv]

you can literally implement every foundational component needed to build an LLM from scratch here

[Alex Lieberman @businessbarista]

One of the most productive engineers I know cannot read code.

@mvanhorn is not an engineer by training. He co-founded June, has 44,000 GitHub stars, and got contributions merged into Go & Python.

I had him build live for an hour & walk me through his process.

My fav lessons & quotes from the convo:
1) He never reads code. He does not have an IDE installed. In his words, "I fundamentally believe that very soon no humans should ever write any code and that no humans should ever read any code."

2) Every project starts with the agent writing itself a plan.md file using Compound Engineering. Agents are lazy, he says, and the plan keeps them honest. He never reads the plan either. "Plans are for agents, you silly human."

3) His CLIs leave notes for themselves. Each run records what it learned in a markdown file, so the next run starts where the last one ended. He calls them self-healing.

4) His agent has an email address. From Telegram on his phone, he sends a task. It emails his Mac, authenticated, and the work starts while he is at his kids' soccer practice.

5) He feeds whole transcripts, not summaries. After a two-hour meeting with a Google Ventures researcher, his agent read the researcher's entire book, wrote itself a report on every chapter, and turned it all into a plan for his business.

6) Him yapping to his agents is like nerd ASMR. "Go agent go" is how he likes to finish telling the agent what to do.

7) How he thinks about this next chapter of building: "Every generation of tools moves the engineer's job up a level. The code was never the point. The problem was."

[Dep @0xDepressionn]

Karpathy's 4 rules took coding accuracy from 65% to 94%.

most devs haven't read them.
the ones who did set up 21 rules total.
82,000 people on GitHub figured this out.

you're looking at all 21.

save this

[Anatoli Kopadze @AnatoliKopadze]

instead of watching 2 hours of Netflix tonight, watch this Stanford lecture

it's the clearest explanation I've seen of how ChatGPT and Claude actually work

useful whether you've never touched AI in your life or have been using it every day for the past year

I took the key ideas and turned them into a practical guide on how to actually get 100% out of Claude

find it below

[Farza 🇵🇰🇺🇸 @FarzaTV]

I did an interview about all my weird habits and workflows I've created to come up with ideas, build stuff, and stay sane in 2026. Enjoy!

[Eloi | UGC IA 🧨 @EloiLJF]

- Utiliser Claude Code tous les jours.
- Me croire avancé.
- Tomber sur le keynote d'Anthropic à Londres.
- Premières 10 minutes.
- Attends. Je faisais quoi depuis des mois ?
- Des agents qui se pilotent entre eux.
- Un Claude qui se souvient de tout.
- Un context window qui tend vers l'infini.
- Des trucs que 99% des gens ne connaissent pas encore.
- 60 minutes plus tard.
- Je comprends enfin comment les pros buildent avec Claude.
- Multiagent Orchestration. Dreaming. Outcomes.
- Tout un système que j'ignorais.
- J'utilisais un Ferrari en deuxième vitesse.
- 6 ingénieurs Anthropic. 1 conférence. Tout est là.
- Problème de skill découvert.
- La vidéo est en vas.
- Mets-la en signet. Maintenant. 🔖