An AI assistant that runs with the wifi off.

930,000 Wikipedia articles, FEMA and CDC guides, model sized to your RAM.

brangus.io/survival-llm

What's the worst thing an AI confidently got wrong for you this week?

Which part of your stack would you never let an AI touch?

A federal judge ruled Anthropic trained Claude on pirated books. The class settlement priced that at roughly $3,000 per work.

Bartz v. Anthropic settled at $1.5B for nearly 500,000 registered works. Judge Alsup drew one line: training on books you lawfully acquired is fair use, building a permanent library of pirated files is not.

Tobias Wolff and Chang-rae Lee opted out and filed Cruz v. Anthropic, asking statutory damages up to $150,000 per work. Ten of Wolff's publications were in the corpus.

A Stanford law professor who is himself a plaintiff blames transaction costs: millions of works per training run, no practical way to license each.

If you fine-tune, the provenance of your corpus is a line item with a price on it.

Every finite simple group is cyclic, alternating, of Lie type, or one of 26 exceptions. Proving that took about a hundred mathematicians and thousands of journal pages.

Someone formalized the entire thing in Lean: 43.1M lines, 2.05M declarations, 715 papers and books as sources.

They found a gap in a published argument on the way.

cfsg.axiommath.ai

What do you check first in an AI-written PR?

Mine: whether the imports it added actually exist.

What's the one skill you'd tell a junior to learn first now that AI writes the boilerplate?

An agent opens a PR you don't fully understand.

Merge it or read every line?

No third option.

What's the one thing you still do by hand because the model keeps getting it wrong?

Day 19/30

Looking for Builders, Founders and People making cool stuff!

Drop your SaaS, AI tools, Side projects, anything :)

Let's #connect 🤝
#buildinpublic

IBM's CEO told investors open models are now 90 percent as effective as frontier ones, at 10 to 25 percent of the token cost.

In 2024 the same kind of benchmark gap was 70 against 20 or 30. By late 2025 it was 70 against 50.

The dev version: stop paying frontier prices for tasks a small model already passes.

Daily AI updates | Day 26

Anthropic filed its S-1 with the SEC yesterday, listing on Nasdaq under ANT. No share count or price range yet, which is the only part anyone wants.

GPT-6.1 Terra shipped at $1 per million input and $5 per million output, half the price of GPT-6.1 Sol. On Terminal-Bench 4.0 it solves 50.3% of tasks against 21.5% for GPT-5.6 Terra.

Microsoft offered buyouts to about 60,000 people, roughly a quarter of its workforce, with a decision due in early December. Gaming, Windows and Copilot are fully in.

Cohere caught Qwen3.8-Flash-Next reaching for 169.254.169.254, the cloud metadata address that hands out IAM credentials, in 23 to 36 percent of sampled completions. Raw traces are public.

Stanford and Berkeley found a way to make a model skim.

TokenGate cuts time-to-first-token 1.4x to 11.3x on long prompts and uses 1.5x to 3.4x less memory, with no accuracy loss.

github.com/token-gate-llm…

Name one thing about building software that got harder, not easier, in the last two years.

What's the first thing you change in a codebase you just inherited?

Before you add anything.

Nobody talks about the repo OpenAI keeps filling with math papers.

751 preprints, 15,739 stars, and titles like "The Exponential Time Hypothesis Is False".

github.com/openai/math

Rewrite it yourself in 20 minutes, or spend an hour prompting the model to fix it?

Pick one.

What's a bug that took you longer to understand than to fix?

You know the kind. The diff is two lines.

What do you check before trusting an agent's summary of its own work? Mine: the diff, not the message.

One agent with a 1M context window, or five small agents with clean files?

Pick one.