Viacheslav Prokopev

[PyTorch @PyTorch]

FlexAttention is a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code @Boyuan_Feng & @__avik show how many existing attention variants can be implemented via FlexAttention & that we achieve competitive performance compared to handwritten kernels. 📺 Watch the full video on YouTube: https://t.co/oRi5myp9zu

[Rohan Paul @rohanpaul_ai]

OpenAI's o3 LLM to discover CVE-2025-37899, a remote zero-day use-after-free bug in the Linux kernel’s SMB (ksmbd) implementation.

o3 autonomously reasoned about 12k LoC and uncovered a bug missed by human review

-----

→ o3 found this bug in the SMB logoff handler, where sess->user is freed without synchronization, allowing other threads to still access the freed memory, leading to potential kernel code execution.

→ The test used no scaffolding or agents—just direct API calls. Code context was expanded to ~12k LoC (~100k tokens) for the test that discovered the new CVE.

→ In earlier tests, o3 identified another use-after-free (CVE-2025-37778) 8 out of 100 times with a 1:4.5 true-to-false positive ratio. Claude 3.7 only found it 3 times; Claude 3.5 missed entirely.

→ o3 outperformed Claude in clarity and structure of its bug reports, at times even outperforming its human operator in reasoning about patch sufficiency.

→ These results suggest LLMs, when used right, can dramatically boost effectiveness in vulnerability discovery—even in kernel-level codebases.

[Ben Nash @bennash]

@sama “if you are not skillsmaxxing with o3 at minimum 3 hours every day, ngmi”

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

Tina: Tiny Reasoning Models via LoRA

"the best Tina model achieves a >20% reasoning performance increase and 43.33% Pass@1 accuracy on AIME24, at only $9 USD post-training and evaluation cost (i.e., an estimated 260x cost reduction). Our work reveals the surprising effectiveness of efficient RL reasoning via LoRA."

[Jeff Dean @JeffDean]

I see what you mean. This is normal.

[Sam Altman @sama]

o3 and o4-mini are super good at coding, so we are releasing a new product, Codex CLI, to make them easier to use.

this is a coding agent that runs on your computer. it is fully open source and available today; we expect it to rapidly improve.

[Ian Nuttall @iannuttall]

somebody on reddit got the full 400+ line devin system prompt 🤯

(link in replies below)

[Yushe @_yushe]

Apparently all the janitor and moderator emails have been leaked.
Some of the emails contain .edu domains for American universities.

[Yushe @_yushe]

4chan just got hacked hard.
The person who hacked them claimed they dumped the entire database.

[Matt Shumer @mattshumer_]

I've officially switched away from Cursor.

My new stack:
- Cline (w/ Sonnet 3.7) for frontend
- RepoPrompt (w/ o1 pro) for backend

Accuracy and quality are just so much higher when you're not trying to save $ by compressing to fewer tokens.

[Aadit Sheth @aaditsh]

Google literally dropped a free masterclass with 601 real-world AI use cases

[Matthew Berman @TheMattBerman]

The AI world just got disrupted by Optimus Alpha, a SECOND mystery model that's outperforming almost everything.

Is it OpenAI's secret new release or something else entirely?

Let's investigate 🧵

[Deedy @deedydas]

There's a Chinese Github repo which helps you get Cursor for free by registering custom emails.

[Roy @im_roy_lee]

I just got kicked out of Columbia for taking a stand against Leetcode interviews.

Here's the whole story (long thread):

[Jen Zhu @jenzhuscott]

I think Manus is here to destroy the $20,000 / month subscription.

And the $200/month subscription. 🤯

[Nathan Wilbanks @NathanWilbanks_]

SLOP vs MCP

[neural nets. @cneuralnetwork]

it has just been one month. oh boy.

[Shubham Saboo @Saboo_Shubham_]

China reached AGI with Manus AI Agent before we saw GPT-5.

It’s the DeepSeek moment for AI Agents without $20,000 monthly subscription.

[Rohan Paul @rohanpaul_ai]

Github 👨‍🔧: Awesome-GraphRAG: A curated list of resources (surveys, papers, benchmarks, and opensource projects) on graph-based retrieval-augmented generation.

-------------

→ Provides a categorized list of research papers on GraphRAG, structured according to a comprehensive survey.

→ Covers key areas within GraphRAG including knowledge organization, knowledge retrieval, and knowledge integration techniques.

→ Offers a comparison between traditional RAG and GraphRAG, clarifying the advantages of graph-based approaches.

[Deedy @deedydas]

Manus, the new AI product that everyone's talking about, is worth the hype.

This is the AI agent we were promised.
Deep Research+Operator+Computer Use+Lovable+memory.

Asked it to "Do a professional analysis of Tesla stock " and it did ~2wks of professional-level work in ~1hr!