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
Claude took a laser from 58% to 99.3%. What it shipped was a decision tree.
That framing undersells it, and the number I keep coming back to is not the integration time. At QuEra, four engineers had spent months on a bespoke script that recovered a laser's lock…
By Pranay Mahendrakar Read postThomson Reuters spent $40M on a model. The training run cost $450K. That gap is the lesson.
Thomson Reuters launched Thomson on 24 August: its own frontier model, built for legal and tax work. It debuted inside exactly one feature, the Tabular Analysis tool in CoCounsel Legal…
By Pranay Mahendrakar Read postAlibaba open-sourced one video model this month and closed the other. That's the strategy.
On August 24, the same company launched Wan 3.0. Thirty-second clips generated in a single pass, up to 1080p at 30 frames per second, audio produced in the same run, and inputs that now…
By Pranay Mahendrakar Read postOpenAI retires o3 tomorrow. In production, a model is a dependency with an expiry date.
Neither of those is a disaster on its own. Put them on a calendar with the rest of the year and the shape changes. On 23 October, sixteen more models go dark in a single wave, including…
By Pranay Mahendrakar Read postEurope's AI transparency rules went live three weeks ago. Only half of them are enforceable.
The first is disclosure. If a person is interacting with an AI system, the system has to say so. Chatbots, voice agents, interactive assistants — they identify themselves at the first…
By Pranay Mahendrakar Read postWe spent a decade pushing AI to the edge. The agent stack is quietly pulling it back.
It's called Kitesurf. It runs entirely in V8 isolates on Cloudflare Workers, and it exists for one reason: agents were driving Chromium, and Chromium is a terrible thing to hand an…
By Pranay Mahendrakar Read postCoSnitch didn't hack Copilot's model. It hacked everything Copilot was allowed to touch.
Varonis Threat Labs reported it in December 2025. Microsoft patched it on August 18, 2026. The bug has a name now: CoSnitch, CVE-2026-24301, CVSS 8.8. Here's the mechanic, stripped down…
By Pranay Mahendrakar Read postHot take: the scariest thing in AI this week isn't a model release. It's a User-Agent header.
Three days. For a library that's basically invisible to anyone outside ML engineering. If you haven't run into Ray, it's the distributed computing framework that Amazon, Apple, and…
By Pranay Mahendrakar Read postCerebras' CS-4 is 30x faster than GPUs. It still won't help most production AI I've shipped.
That's not a typo. That's roughly 2,000x the bandwidth of a next-gen GPU on that one metric, according to Cerebras' own numbers and the early third-party coverage I've read…
By Pranay Mahendrakar Read postOpenAI's Astra solved 10 open math problems. The number that matters is the zero.
The proofs weren't just generated. They were formalized as Lean 4 certificates and published on GitHub, and the repository's "sorry" count sits at zero. In Lean, "sorry" is the…
By Pranay Mahendrakar Read postWashington wants to ban Kimi K3. That's solving the wrong security problem.
I build deployment infrastructure for a living, including systems that run in defence and government settings where "what does this model do with our data" is not a hypothetical…
By Pranay Mahendrakar Read postHot take: DARPA's AI flew an F-16. The switch that took it back is the real story.
Here's what happened, stripped of the headline. DARPA and the US Air Force flew a modified F-16 at Eglin Air Force Base under the VENOM program — Viper Experimentation and Next-gen…
By Pranay Mahendrakar Read post