OpenAI’s chief scientist publishes “An Alien Mind”: alignment unsolved, voluntary slowdowns welcome
Jakub Pachocki, OpenAI’s chief scientist, published an essay whose central claim is that AI is becoming a form of intelligence we don’t fully understand, may soon struggle to monitor, and that increasingly drives its own development. Three judgments stand out: no lab, in his view, has solved alignment and monitoring well enough to justify scaling at maximum speed; he expects, and hopes, that voluntary slowdowns become commonplace until the industry establishes shared safety bars; and international coordination on AI development should become a top priority for governments. He also splits alignment into two layers: goal alignment (does the model try to accomplish the goal you set) and value alignment (can it act on a set of high-level principles when goals are vague, conflicting, or adversarial), and argues the second is the harder and more consequential one. This is the company’s chief scientist writing on its own site; read it against the next item to get the full picture.
Inside OpenAI: 3.1 agent-workdays per human workday
The same day, OpenAI published numbers on how far coding agents have spread through its own research process. The company says it has hit its previously stated September 2026 milestone of an “automated research intern”: a system that completes well-scoped research tasks under human direction, including tasks that would take a skilled researcher a few days. By mid-August the research org was logging 3.1 agent-workdays for every human workday, and experiments per active researcher hit an all-time high in August (tracking began January 2025), which OpenAI says is “correlated” with rising internal Codex use. The next internal target is an “automated AI researcher” by March 2028. Two caveats: an agent-workday counts total agent runtime converted at a standard eight-hour day, and since agents run in parallel and much of that runtime can be redundant or unsuccessful, OpenAI itself cautions that these usage numbers are imperfect proxies, so 3.1x the hours is not 3x the output; the company also notes that research has many other bottlenecks, so overall progress won’t track any single metric. Even discounted, this is the most concrete internal data I’ve seen any frontier lab share on AI accelerating AI research.
Authors push back as publishers claim shares of the Anthropic settlement
Anthropic’s $1.5 billion book-piracy settlement ($3,000 per pirated work) has reached the payout stage, and as allocation notices go out, authors are finding publishers and literary agents claiming shares the authors say belong to them. Under the terms, in-print traditionally published books split 50-50 between author and publisher, while self-published books and books whose rights have reverted go entirely to the author; rights reversion is a contractual process, not an automatic switch: typically, after a book goes out of print the author sends the publisher a written demand, and the publishing rights return to the author only if the publisher fails to reissue or license the work within a set period (the Authors Guild’s model contract clause walks through the steps). Mystery writer April Henry says HarperCollins claimed a share of a book that reverted “at least 17 years ago,” and Victoria Strauss, who runs the industry-watchdog project Writer Beware, reports a stream of identical complaints and calls the problem widespread and systemic, though she is reluctant to blame on malice what poor recordkeeping can plausibly explain; the Authors Guild has published a dispute process. Now that this settlement has reached the execution stage, the friction is coming not from the AI company but from decades of messy rights records inside publishing; my bet is that how this payout pipeline handles disputes becomes the template for the settlements that follow.
Today in one sentence: the acceleration data and the slowdown plea came from the same company on the same day; that pairing says more about the current pace of frontier research than any outside warning could.