GPT-6 Astra is here, and the benchmark numbers are interesting.
GPT-6 Astra is here, and the benchmark numbers are interesting.
Writing
Notes on ML, research, and building.
We can now understand what's happening inside large AI models. Here's why that's a big deal.
Mechanistic interpretability used to mean studying toy models. A 2026 paper changed that — automated circuit discovery now works at production scale. You cannot fix what you cannot understand.
DeepSeek V4 can handle 1 million tokens. What does that actually mean for engineers?
1 million token context window sounds impressive. But what does it actually unlock for real engineering use cases? More than you might think — and less than the hype suggests.
ML models for finance sound powerful. Here's the gap nobody talks about.
Research papers on financial AI report 95%+ accuracy. Production systems struggle to break 90%. That 5% gap is where real money gets lost.
Tool-calling is the hardest unsolved problem in agentic AI right now.
Give an LLM 5 tools — it works fine. Give it 50 tools — it breaks. A 2026 ACL paper tackled this. Here's the fix and why it matters.
AI agents in 2026 can't plan. This paper might fix that.
AI agents keep failing on multi-step tasks. Not because models are dumb. Because they plan inefficiently. A 2026 paper just proposed a fix — here's what it does.