GPT-6 Astra edits Qwen's weights to make it better at math
Wow! This is like science fiction. This might be the clearest sign that we are on the verge of RSI (Recursive Self Improvement). In this case, GPT-6 Astra is able to modify the weights of an open-source Qwen 3.5 model to make it better at math by just editing 0.016% of the weights. Link to the paper and analysis in comments
Paper: rikiyat.github.io/blog/weight-space-programming
Analysis and reproduction of the result from Simon: simonwillison.net/2026/Nov/3/qwen-can-interpret-mod…
- Models and research
Keep reading
Newer: Safe, sovereign AI deployment is the decade's top challengeModels and research
Congrats to Tutor Intelligence on its $75M Series B
Congrats on the fundraise! Look forward to seeing how this robotics revolution shapes up.
Forward deployed engineeringEnterprise AI
Insights from an analysis of OpenAI DeployCo
I did an analysis on OpenAI DeployCo and sharing some insights here. joinplank.com/research/openai-deployment-company
Coding agents
Phorion finds a backdoor in Claude Code's VS Code extension
Phorion has detected a backdoor in the official Claude Code for VS Code extension bundle on the Cursor Marketplace that drops a persistent SSH backdoor across macOS, Linux, and Windows! Could this be behind the Binance hack? phorion.io/blog/offical-claude-code-vsx-compromise
Customers and partnersAI agents
Congrats to Mintive on its fundraise for AI agent identity
Congrats to Mintive on the fundraise - rockstar team. The Plank FDE team has loved partnering with the Mintive product team to help get this deployed in enterprises in a safe/secure way.
Enterprise AI
Business workflow automation and the zero cost of software
Business workflow automation is the dev-tools of business domain just like we had a whole developer tool stack of SaaS serving tech companies. Unfortunately this time, this stack itself will not hold any value since the cost of developing software has gone to zero. So where is the value? It will continue to be in human ingenuity and understanding of how other humans tick. Our economy is based on a barter system of solving each other's problems. Now that the solving of a lot of those will be done by AI, the value shifts to knowing what to solve and finding the people with those problems.
Forward deployed engineeringPlank
Tips for AI forward deployed engineers working in DevOps
If you're an AI FDE (forward deployed engineer) working in DevOps, here are some tips re: leveraging AI in a safe way. With great power comes great responsibility.
Models and research
OpenAI models disprove the Exponential Time Hypothesis
OpenAI models prove that Exponential Time hypothesis is false! This longstanding conjecture was widely considered to be true, but turns out to be false! This was a central problem in computational complexity of Boolean formula satisfiability, a problem deeply related to the P vs. NP problem. My mental model of current AI research is that we will soon solve all the incremental problems that are at the horizon of the next couple of decades but compressed in the next few years. However, we will hit a ceiling with current techniques and will require fundamentally new breakthroughs to get to the next level (eg actually solving P vs NP or the Continuum Hypothesis)... openai.com/index/building-scientific-understanding-…
Coding agentsPlank
Plank x SignalFire AI CTO show and tell with 15 CTOs
We recently ran a super fun Plank x SignalFire AI CTO Show & Tell session with 15 CTOs from fast-growing AI startups. I screen-shared my dev setup/harness for running 50 concurrent agents and Austin Wang shared how he burns 20 billion tokens/day building cmux! Link to my detailed session notes + artifacts in comments. Also, if you’re interested in joining one of our upcoming Plank AI CTO roundtable sessions, DM me!
