Daily Signal · 2026-05-04
The Daily Signal — May 4, 2026
AI systems approaching the ability to build themselves, Google goes all-in on Pentagon AI, and Musk's last-ditch settlement text backfires.
Research
Jack Clark: 60% chance AI systems build their own successors without human input by 2028
In a sweeping essay published today, Anthropic co-founder Jack Clark argues that all the engineering components required to automate AI R&D end-to-end are already in place. His evidence: Claude Mythos Preview now scores 93.9% on SWE-Bench, AI systems can complete tasks requiring 12 human hours independently (up from 30 seconds in 2022), and Anthropic's agents have already beaten human baselines on alignment research tasks. Clark stops short of predicting a 2026 event, but gives 60%+ odds that a non-human AI system will plausibly train its own successor by end of 2028.
Policy
Google signs classified Pentagon AI deal as 600+ employees revolt — and management doesn't blink
Google amended an existing DoD contract on April 28 to allow Gemini models to operate on classified defense networks for "any lawful government purpose" — a phrase that gives the Pentagon, not Google, final say over use cases. More than 600 employees signed a letter urging CEO Sundar Pichai to pull back. He didn't. Google must adjust AI safety filters at Pentagon request and holds no veto over lawful uses. A Fortune analysis published today notes the backlash is notably weaker than the 2018 Project Maven revolt, which succeeded: employee leverage over military AI has materially eroded.
Industry
Musk texted Brockman a settlement offer, then threatened him when he pushed back
A court filing unsealed Sunday night reveals Elon Musk texted Greg Brockman two days before trial to probe whether OpenAI would settle. When Brockman suggested both sides drop their claims, Musk replied: "By the end of this week, you and Sam will be the most hated men in America. If you insist, so it will be." OpenAI's lawyers are seeking to introduce the message as evidence of competitive motive, not genuine grievance. Trial resumes this morning with Brockman expected on the stand; Altman and Satya Nadella are scheduled to testify later this month.
Industry
OpenAI's 'DeployCo' closes $4B from PE giants at a $10B valuation
TPG, Bain Capital, Advent International, Brookfield, and Goanna Capital are set to invest $4 billion into DeployCo, a Delaware LLC majority-controlled by OpenAI via super-voting shares. OpenAI commits up to $1.5B of equity and guarantees investors a 17.5% annual return. The venture, run by former COO Brad Lightcap, will embed forward-deployed engineers into PE portfolio companies. The move directly targets Anthropic's enterprise push; the FT reports Anthropic is separately in talks with Blackstone and Hellman & Friedman on a comparable vehicle.
Financial Times / PE Insights →
Agents
Anthropic's agents beat human researchers at their own alignment benchmark
Anthropic's Automated Weak-to-Strong Researcher showed autonomous AI agents proposing ideas, running experiments, and iterating on scalable oversight problems — then outperforming the human-designed Anthropic baseline. Jack Clark flags this in today's Import AI as a landmark proof-of-concept: frontier agents can now contribute novel results to cutting-edge AI safety research, not just automate rote tasks. The system doesn't yet generalize to production models, but every comparable benchmark started this way before improving sharply within months.
Frontier
METR's task-horizon curve: AI went from 30 seconds to 12 hours of autonomous work in four years
A chart embedded in Clark's Import AI 455 captures the steepest capability slope in recent AI history. The median task horizon — how long AI can work reliably without human check-ins — grew from 30 seconds (GPT-3.5, 2022) to 4 minutes (GPT-4, 2023), 40 minutes (o1, 2024), 6 hours (GPT-5.2 High, 2025), and 12 hours (Opus 4.6, early 2026). METR forecaster Ajeya Cotra projects 100-hour autonomy by year-end — the threshold at which AI can independently complete most multi-day research tasks.