Updates
Updatesand Insights.
Thoughts on AI infrastructure, agent memory, and building production ML systems.
The harness is the product. GPT-6 Astra just proved it.
ARC Prize verified GPT-6 Astra at 99.9% on ARC-AGI-3 and at 62.7% on the same benchmark. Same weights. The difference was the harness, and that should change...
What it takes to trust an operations agent
We spent five days measuring our operations agent against AIOpsLab, Microsoft's benchmark for autonomous incident response, and ended at or above the best pu...
Qwen's agentic coding is the real story in open models
Qwen's results on long-horizon agentic coding matter more than headline scores. Post-training like Ornith shows how much further strong open bases can go wit...
Intelligence is cheap now. Trust is the bottleneck.
An a16z crypto essay argues the agent economy is now limited by identity and verification, not by intelligence. I think that's right, and it applies just as ...
Software written by agents needs somewhere built to run it
A general assistant agent is lovely for booking a table and triaging an inbox. Production software has a different shape: it has to survive failures, keep ma...
AI made code cheap. Rework is where the money goes.
The inference bill is the smallest cost in an AI coding programme. The large ones are rework, review and the work of proving what happened afterwards. Why th...
When finding vulnerabilities gets cheap, where your software runs matters
AI agents have made probing software for weaknesses cheap enough to run continuously. That changes the maths on dependencies, patching and isolation. Here is...
Voxeltron: where AI-built software goes to run
AI made writing software cheap. Running it didn't get any cheaper. Voxeltron is where the software your agents build gets deployed, isolated, secured and kep...
Kizuna: where AI agents build software under your rules
Code forges were built for people who run a bit of automation. Agents now write a growing share of the code, and the forge still can't tell you which one did...
Know what your agents will cost before they run
Gartner expects a large share of agentic AI projects to be cancelled by 2027, with cost near the top of the reasons. Watching spend on a dashboard doesn't pr...
One record per change: what auditors actually want from AI-built software
When agents write the code, the audit question doesn't go away. It gets harder. What a single record per change needs to answer: who made it, what it touched...
Approvals that keep up with your agents
When agents work well, people become the queue. Human approval only works at agent speed if it's reserved for decisions that matter and arrives with the cont...
On your own ground: running AI-built software where you choose
Where AI-built software runs should be your decision, not a side effect of which tools you picked. A calm look at the choices, from your own infrastructure, ...
EU AI Act Article 12: what to keep when AI agents change your systems
Article 12 of the EU AI Act requires high-risk AI systems to record events automatically, by design. Here is what the article asks for in plain language, why...
AI writes the first 20%. Shipping is the other 80%.
AI made the first working version of a system cheap. What still decides whether it reaches production is everything around the code: who did the work, whethe...
Set the rules once, let agents build within them
A developer with a coding assistant carries the rules in their head. A team of agents working overnight doesn't. Why the interface for agent-built software i...
The price of writing software with AI just fell again
DeepSeek V4 brings a hybrid attention stack, constrained hyper connections, the Muon optimiser and FP4 quantisation-aware training to a pair of open-weight, ...
Autonomous agents need a record of what they tried
Andrej Karpathy's autoresearch lets an agent run a hundred training experiments overnight with no human in the loop. It's a glimpse of how autonomous agents ...
Running AI models on your own ground: a practical guide
How to run open models on hardware you control, which quantisation to pick, and a measured experiment fine-tuning Gemma 4 E4B on medical and finance data. On...
Residual connections have been broken this whole time
Kimi's Attention Residuals paper replaces the fixed sum in residual connections with learned attention over depth. What it fixes, what it costs, and what the...