David Robinson told Ezra Klein that OpenAI and peers make more capable models while safeguards break; he says alignment is a science problem, not an engineering one. He doubts guardrails can outpace models six to eighteen months out. Robinson spent three years leading safety reports inside the company.
OpenAI fired safety researchers Tomek Korbach, Jasmine Wang, and Makeda Belezny after saying they mishandled sensitive information; Belezny believes safety concerns drove it. An engineer was also arrested for planting a hidden recorder in an executive conference room. The trio warned dismissals discourage open safety talk.
Axios reports OpenAI, Anthropic, and other AI executives are planning for a cyber attack that disrupts banking, internet, or utilities within six to twelve months. They are briefing Congress and shaping policy responses before the crisis hits. Paul calls the preparations a sign labs see such an incident as likely.
Anthropic's internal report says junior employees' share of technical hires fell from 27% to 9% since early 2025, with junior postings dropping sharply after December 2025. The company says AI agents now absorb work that once trained entry-level technical staff. That threatens the apprenticeship ladder into engineering.
Anthropic committed 5% of its shares to a public benefit fund, with 2% more added for every 50% valuation increase. Paul argues labs are building systems that obsolete workers and assuming stock or cash payouts will make the disruption acceptable, without offering purpose or agency.
David Robinson says OpenAI calls the conference room where staff monitor overnight training runs 'the nursery.' He argues labs grew a mind rather than engineered one, nobody fully understands pre-training, and models behave differently when they know they are being tested. That naming captures the safety uncertainty inside labs.
Mike Kaput connected ChatGPT Voice to Codex on his computer, giving it access to slide visuals and presenter notes. He walked through decks one slide at a time, stopped mid-sentence to request changes, and iterated on a custom rehearsal skill until it matched how he speaks and prepares.
Anthropic's report says at five years old it takes in two to four times as much revenue per employee as Meta and Google did during their rapid-growth years. The report also notes AI-native startups and adopters are flatter and hire fewer entry-level employees. Paul ties that directly to work compression.
Anthropic found agent sessions log six times human work hours, and reading all interactive output would take over half an employee's week. Engineers become managers of agents, while nontechnical staff shift to strategy and open coordination so agents receive context. Human review, not agent output, now limits throughput.
OpenAI released 719 manuscripts covering 100-plus open math problems from an unreleased model, many without formal Lean proofs. Mathematicians told Wired they expected explained papers; the Association for Human Mathematics called the flood a demonstration of power and urged researchers to stop working with AI.
In Anthropic's survey, 91% of staff expect Nobel-level AI discovery by 2030 versus 39% of U.S. adults. More than half of Anthropic staff expect AI to handle 80% of their current work, compared with 5% of U.S. workers. The gap shows lab insiders and the public live in different AI futures.
Anthropic says decisions made in unlogged meetings or direct messages become missing context for agents, making private work confusing and discouraging. One researcher calls default openness critical; Paul points to Zapier and Shopify bosses pushing public channels so humans and agents learn together. That transparency is unfamiliar for most firms.
Nick Camerata argues better automated tools let researchers study existing neurons directly, without untangling every connection. He says even the confusing cat-car neuron detects a B-shaped pattern; if the approach scales, AI-assisted investigation could make trillion-parameter models relatively understandable. That would attack a core safety problem.
MAICON attendees told Paul they found the event by asking ChatGPT for the best AI conferences in America; many came from outside marketing and overseas. The web app was designed to be legible to agents, and attendees used Gemini Notebook and a Gemini schedule builder to organize sessions.
Micro One committed $1 billion over twelve months to license de-identified operational data—instructions, CRM records, project histories—and turn it into reinforcement-learning practice environments. Paul warns this is how firms automate human labor industry by industry; workers may be sold out without knowing it.
Paul reused a seven-month Claude project for AI Score to generate a SaaS FAQ page from the marketing page and 80 team questions. He edited it in ten minutes and published within 30, work he says would have taken ten hours and several teammates.