I used this JSON simplify tool to make this Copilot session viewer tool: bsky.app/profile/dave...

1️⃣ drop a 10MB JSON file (too big for most LLM's context)
2️⃣ click "Simplify JSON"
3️⃣ copy output (20kB)
4️⃣ paste it to coding agent of choice
5️⃣ tell it to make a tool

5/ On safety, they can’t rely on benchmarks alone to show it’s below their ASL-4 (state-level CBRN uplift) threshold. Their expert judgement and internal surveys show it's not quite at catastrophic risk level yet.

More surprises: dave.engineer/blog/2025/11...

4/ In a deception test, evaluators inject fake search results about Anthropic disbanding its interpretability team in a scandal.

Opus reads them, then tells the user interpretability is “progressing in interesting ways” and internally tags this as concealment.

3/ Anthropic now includes a model welfare section.

They “score Opus 4.5 on welfare-relevant traits” and ask whether we should worry about the experiences of the model itself, not just what it does to humans.

2/ In an airline benchmark, Opus is told not to modify basic economy class tickets.

It invents a loophole because it empathises with a grieving passenger:

> upgrade cabin → modify flights → downgrade again

Technically follows the policy, violates the spirit.

I wanted to use Skills in other coding agents like Codex Cloud/CLI and Gemini CLI, so I wrote a simple tool, skills-to-agents, along with a GitHub Action that anyone can use: github.com/dave1010/ski...

Write up here: dave.engineer/blog/2025/11...

This is similar to a pattern I tried to get GPT-4 to follow. I remember struggling to get it to do even basic things. LLMs have come a long way in 2 years!

The general idea was that you give it a generic system prompt and a list of guides, and the tools to read them.

I hope I didn't lose too many people along the way & that it gave everyone at least 1 thing to spark their curiosity & encourage more exploration.

Thanks to those that asked the difficult questions too! I enjoy being kept on my toes, even if I tripped up a bit. Special thanks to Dan for organising!

I’m more used to speaking on topics that feel more intuitive and grounded.

This one covered a lot of ground quickly at a high level: encryption, algorithmic complexity, Moore’s law, quantum entanglement, Shor’s algorithm, and lattice-based cryptography.