AI digest: chips, agents, and an awkward equity offer
Anthropic chases custom silicon, AI agents quietly eat freelance work, Microsoft bets big on enterprise deployment, and OpenAI tries to get into bed with Washington.
A lot moved this week. Hardware ambitions, agent capability benchmarks, and some eyebrow-raising politics.
Anthropic is talking to Samsung about custom chips
Anthropic is reportedly in early discussions with Samsung about manufacturing its own AI chip, and has already hired chip engineers to back that up. This comes a week after OpenAI announced its own custom silicon deal with Broadcom. The pattern is clear: every major AI lab wants to cut its dependence on Nvidia and control its own compute costs.
AI agents can now complete 16% of freelance jobs at pro quality
Eight months ago that number was 2.5%. The Remote Labor Index, which measures AI agents completing paid freelance work at professional quality, shows the top automation rate has more than quadrupled. This is the kind of benchmark that actually matters, because it tracks real output on real tasks, not synthetic evals.
Microsoft launches a $2.5 billion deployment unit
Microsoft is not just selling AI tools anymore. Its new “Frontier Company” unit puts 6,000 engineers directly inside enterprise clients to integrate AI into core processes. The pitch is measurable ROI, not more pilots that go nowhere. This is a significant shift from platform provider to hands-on delivery partner.
OpenAI offered the US government a 5% stake
Sam Altman reportedly proposed giving 5% of OpenAI’s equity to a US sovereign wealth fund, framed as sharing the gains from the AI boom with the public. What the government gives in return is unclear. It reads less like generosity and more like a very expensive insurance policy.
Anthropic cut Claude Code’s system prompt by 80%
According to Anthropic staffer Tariq Shihipar, the new Fable 5 models work better with far fewer instructions. The old guidelines were actually holding the models back. That tells you something useful about prompt engineering: more instructions is not always better, and newer models may need less scaffolding, not more.