#ai
- DeepSeek V4.1 Flash reads cheap, writes expensive
The number that matters to me isn't the 40 on the Intelligence Index, it's the input side: 8 billion active parameters per token for…
- SWE-2 gets to the edit 30 steps sooner
The number I care about isn't 50.0% on FrontierCode. It's 18 steps versus 48 before the first real edit. I've watched agents grep their way…
- A pretraining researcher just quit Anthropic, and I still opened Claude Code this morning
Jacob Coxon spent three years inside pretraining at OpenAI and Anthropic, and his exit line is that neither is acting responsibly. I build…
- The Proof Compiled Because Someone Built a Graph
The part I keep rereading isn't the 13 million lines. It's that the first agent teams failed — they got partial results, lost track of the…
- GPT-6 Astra: $50 Output and the Speed Tax
The provider table is the interesting part, not the model card. OpenAI Flex runs $5/$25 at 64 tok/s. OpenAI Fast runs $20/$100 at 14 tok/s…
- Gemini 3.8 Flash: The Model That Works Harder Than You Asked
Three Flash releases in six weeks, and the interesting line isn't a benchmark — it's Google admitting 3.8 Flash "works harder," burning…
- 1.7GB of LibreOffice Is Hiding in My Cache
I found this the way you find a leak in a bilge: not by looking for it, but by looking at something else. Simon Willison poked around…
- Anthropic's Model Hardware Standard: the tags matter more than the driver
The part I care about isn't the driver layer, it's the tagging system. Metadata describing physical properties and limits that don't show…
- Debian Says the Quiet Part: You Own the Patch, Not the Tool
Debian voted down mandatory disclosure for AI-assisted contributions, and I think they got it right for a boring reason: the rule was never…
- Sony and Warner Want Claude's Training Data Deleted
The money isn't the part I'd worry about. Sony and Warner say damages are just an operating cost for Anthropic, and they're right — the…
- Gemini 3.5 Transcribe, or: the end of my Whisper cron job
The part I care about isn't the word error rate, it's the disfluency cleanup. I dictate notes on a boat with the engine running and half of…
- The bug report I wanted was behind a proof-of-work wall
I went to read an FFmpeg issue and got Anubis instead: a proof-of-work challenge, JavaScript required, sitting between me and a page of…
- Salesforce Says the UI Isn't the Product
The line that stuck with me is Patrick Stokes saying the value of Salesforce isn't in the interface — it's the data, metadata, and years of…
- Small Models Have Arrived
I reach for the expensive model by reflex. Coding, drafting, debugging — Fable 5 or nothing, and I never look at the bill. French-Owen's…
- The dependency I never chose deliberately
Half my local model workflow starts with a Hugging Face URL. So a report that Nvidia has been in talks to buy the place for more than $13…
- A Filter That Can Explain Itself
The interesting bit isn't the redaction, it's the codebook. Tippe's team turned GDPR Articles 4 and 9 and German trade-secret law into…
- Gemini 3.7 Flash, three weeks after 3.6
Three weeks. That's the gap between 3.6 Flash and 3.7 Flash, and it's the part that actually changes how I work. I've been treating model…
- Two Cents a Task Changes What You Bother Automating
The number that stopped me wasn't 89.0% on ARC-AGI-1. It was $0.02 per task. I've spent the last year writing prompts as if every call were…
- The bottleneck is me, not the model
The uncomfortable part of Goedecke's argument is that it explains my own logs. When I'm working on the Astro site or a Cloudflare Worker…
- A moderation model I can actually host on the boat
The number that got my attention isn't the 7x claim, it's 16GB. A 3B model that fits on one 16GB GPU is a thing I can run myself, next to…
- Kimi K3 is close enough that I'll rerun my agent evals
The number that got my attention isn't the 2.8 trillion parameters. It's that Kimi K3 won three of six agent tests, and on Artificial…
- Cheap tokens only matter if the model shuts up
Luna at $0.20 in / $1.20 out per million tokens is the kind of number that changes what I'm willing to run in a loop. The jobs I've been…
- Cloudflare's quantization math is the same math I run on the boat
The part of this I keep rereading is the throughput table for the KV cache. At one concurrent request, BF16 beats FP8 — 137 tokens per…
- Documentation is state. Insight lives in moments.
Fed my AI curation panel 20 candidates harvested from my project documentation. It rejected 17 outright and drafted 3 status reports no one…
- The model I write this blog with just got pulled
Most of The Cloudy Brain is co-written with Fable 5 — it drafts, I argue with it, I sign. So this one lands close to home: as of yesterday,…
- Munich court: "the AI said it" is not a shield
If you ship anything that presents LLM answers as authoritative to end users, this is the precedent to read twice. The disclaimer-and-pray…
- "You'll never know if the AI degrades you" — except you can
I spent Fable 5's launch day building with it — a whole publishing platform, including the bumps: it over-built twice, and I had to catch…