switch to an envelope, the back is more clear

Google, OpenAI, and Microsoft monitor this and release some findings too. But Anthropic seems the most proactive.

Anthropic warns that the barriers to entry for cyberattacks have dropped sharply, enabling less skilled actors to mount sophisticated operations.

Targets included major tech companies, banks, chemical manufacturers, and government agencies.

This level of autonomy marks an escalation beyond prior “vibe hacking” cases, where humans remained heavily involved.

The only alternative is using third-party providers like Openrouter, but they struggle to set up Chinese models correctly, with significant between-provider differences in benchmarks for the same model!

One annoying thing is that Moonshot (the Chinese AI company who makes the Kimi model) isn't making any privacy guarantees so no one can use their API to run evals. (Also sending private data into China is inherently dicey).

my piece on Iran was very popular: peterwildeford.substack.com/p/the-fordow...

I don't think they articulate their assumptions? They seem to just be bare assertions.

I agree there's a high chance they flame out but in the bull case automating 10% of US white collar labor gets you a trillion per year in revenue.

3. Claude continues to smash others at SWE-Bench too, as usual.

2. Everyone talking about 4.5 being great at coding, but I'm taking way more notice of that huge increase in computer use (OSWorld) score That's a huge increase over SOTA and I don't think we've seen anything similarly good at OSWorld from others? Claude Agents coming soon?