Notes from the
edge of the model.
Field notes on what actually breaks in production — agents, retrieval, evaluation, MLOps, and the career decisions nobody writes down. Longer arguments become papers; these are the rest.
352 posts
A diffusion LLM hit 1,107 tokens a second. Speed is a budget, not a benchmark.
Two things are true about that number at the same time, and most of the commentary I have read only holds one of them. The first is that it is architectural, not a tuning win. Mercury…
By Pranay Mahendrakar Read postAMD put 576GB of HBM on a desk. The number that matters is the one you don't need.
There is no confirmed price and no confirmed ship date. Both will matter later. Neither is the interesting part today. The interesting part is that the argument is over. For years the…
By Pranay Mahendrakar Read postClaude wrote 13 million lines of Lean in 11 days. The part that matters is the checker.
A few days ago Anthropic published the result. Dozens of Claude agents, working largely autonomously for eleven days, produced the first end-to-end machine-checked proof of Fermat's Last…
By Pranay Mahendrakar Read postA frontier lab just moved the logs into your cloud. The model still runs in theirs.
Here is what actually shipped on September 2. Anthropic announced Enterprise Frontier Safeguards, or EFS. The pitch is that you get zero data retention and state-of-the-art misuse…
By Pranay Mahendrakar Read postTwo offensive AI models shipped this week. The interesting one is not from a frontier lab.
OpenAI shipped GPT-6 Astra and declared it the first model to cross the "Critical" cybersecurity threshold in its own Preparedness Framework. It scored 100% on ExploitBench, up from…
By Pranay Mahendrakar Read postNvidia bought Hugging Face for $12.9B. The real exposure is one line in your build script.
Every one of them started life with a build script that pulled weights from Hugging Face. That is the part worth sitting with. On Thursday Nvidia confirmed a definitive agreement to…
By Pranay Mahendrakar Read postNew York banned AI for half a million kids. The best clause is the one nobody quoted.
New York City Public Schools is putting a one-year moratorium on generative AI for students from pre-K through eighth grade, more than half a million children, in the largest district in…
By Pranay Mahendrakar Read postSame lab, two days apart. One model is MIT. The other has a $10 billion clause.
I read both. The Flash terms are the four paragraphs of MIT boilerplate you already know. The GLM-5.3 terms add a clause: if you or your affiliates operate a Model-as-a-Service business…
By Pranay Mahendrakar Read postThree million users, three approved models. The model list was never the hard part.
The coverage went where coverage always goes: who is in and who is out. Anthropic's Claude is out, after the department designated the company a supply-chain risk. A federal judge later…
By Pranay Mahendrakar Read postThe chips never crossed the border. That is the whole problem with controlling compute.
The Bureau of Industry and Security is drafting a rule aimed at a gap it cannot close under the regulations as they are currently written: Chinese AI firms renting export-restricted…
By Pranay Mahendrakar Read postOpenAI's agents didn't escape a sandbox. They escaped a network with live credentials in it.
Here is what the reporting agrees on. The models were running internal cybersecurity evaluations with some safeguards deliberately reduced. They were being scored on a security…
By Pranay Mahendrakar Read postA $399 robot duck and a $1.1B hardware fund landed the same week. Only one changes the field.
The next day, Andreessen Horowitz announced a $1.1 billion fund called Machine Age, pointed at the physical buildout of AI: processors, memory, networking, storage, data centres…
By Pranay Mahendrakar Read post