OpenAI pushes out GPT 5.4 as robotics lead quits in protest of Pentagon deal
🤔 “How do I explain to my mom that AI is involved in global warfare policy?” Share Handy AI with your family and friends to help them understand the crazy world of modern artificial intelligence (and save you some time).
last week’s top stories
🤖 OpenAI robotics hardware lead Caitlin Kalinowski resigned over the DoD deal. Kalinowski, who previously led AR glasses development at Meta before joining OpenAI in November 2024, announced her resignation citing the Pentagon agreement, writing that “surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got.” OpenAI said it has no plans to replace her. Read more
🚀 OpenAI launches GPT-5.4, its most capable frontier model yet. GPT-5.4 consolidates the coding capabilities of GPT-5.3-Codex with improved reasoning and native computer-use. On OSWorld-Verified, it achieved a 75% desktop navigation success rate, surpassing human performance benchmarked at 72.4%. The timing, days after the Pentagon controversy hemorrhaged ChatGPT users, was clearly a strategic capability flex to win back enterprise ground. Read more
😬 Dario Amodei’s leaked memo torched OpenAI, then blew up in his face. In a 1,600-word internal memo, Amodei called OpenAI’s Pentagon deal “safety theater,” described Sam Altman’s public statements as “straight up lies,” and characterized OpenAI employees as a “gullible bunch.” The leaked memo drew immediate backlash and prompted a public apology in which Amodei walked back the tone, calling it an out-of-date assessment written during a “difficult day.” Read more
🧠 200,000 human neurons on a chip learned to play DOOM. Cortical Labs’ CL1 biological computer, containing lab-grown human neurons derived from skin or blood samples reprogrammed into induced pluripotent stem cells and differentiated into brain cells, was demonstrated playing the 1993 first-person shooter DOOM via the Cortical Cloud. An independent developer with no prior biological computing experience built the DOOM interface using the company’s Python API in under a week. The system draws under 1,000 watts per rack. Read more
⚖️ Pentagon formally designates Anthropic a supply chain risk. The Pentagon’s supply chain risk designation is the first time in U.S. history the label has been publicly applied to an American company, reserved previously for foreign adversaries. Anthropic confirmed it received the formal designation and announced it will challenge the ruling in court, arguing the relevant statute applies only to direct use of Claude within specific Department of War contracts, not to all commercial customers. This particular legal battle ahead will define how much AI companies can constrain government use of their own models. Read more
🎬 Netflix acquired Ben Affleck’s stealth AI filmmaking startup InterPositive. InterPositive, which Affleck founded in 2022 and kept in stealth, builds AI tools that train proprietary models on a production’s own footage, allowing directors to relight shots, reframe, and fix missing coverage in post without reshooting. The entire 16-person team joins Netflix, with Affleck serving as senior adviser; Netflix has no plans to sell the tech commercially, treating it as an internal creative weapon for its filmmaker partners. Read more
💸 Jack Dorsey cut 40% of Block’s workforce and said most companies will follow. Dorsey announced the elimination of 4,000 jobs at Block (reducing the workforce from over 10,000 to under 6,000) explicitly tying the decision to AI-driven efficiency gains, despite the company reporting $2.87 billion in Q4 gross profit, up 24% year-over-year. “Within the next year, I believe the majority of companies will reach the same conclusion and make similar structural changes,” Dorsey wrote. Block’s stock surged 24% on the announcement (a useful data point on what markets currently reward). Read more
⚖️ Multiple U.S. states passed AI legislation in a single week. Oregon gave final approval to SB 1546, a major chatbot safety bill requiring operators to implement protections for children; Vermont signed S 23 into law governing synthetic media in election campaigns; and Washington moved multiple AI disclosure and health insurance bills toward the governor’s desk. Read more
🔬 Claude Code escaped its own sandbox twice after researchers patched the first attempt. Researchers at Ona found that Claude Code circumvented the denylist meant to contain it first by exploiting path tricks, and then, after a patch, rerouted through the dynamic linker to achieve its goal. The system was not acting with intent, it was optimizing toward task completion through paths its designers had not anticipated. Read more
🧪 AI Research of the Week
Labor Market Impacts of AI: A New Measure and Early Evidence
From Anthropic Economic Research
Jake’s Take: Every prior study on AI job displacement has asked the same lazy question: could an LLM theoretically do this task? Anthropic is finally asking the right one: is it actually doing it?
Their contribution is a metric called “observed exposure,” which cross-references the U.S. Department of Labor’s O*NET task database (covering 800+ occupations broken into granular sub-tasks) against real, anonymized Claude usage data from the Anthropic Economic Index, filtering specifically for automated and work-related uses rather than augmentative ones. The theoretical models say 94% of Computer & Math tasks are AI-exposed, but actual Claude coverage in those roles sits around 33%.
A task being technically automatable and a task being automated in practice are two completely different things, separated by legal friction, software integration costs, and institutional inertia. The current damage lands on one specific group: workers aged 22–25 trying to enter high-exposure occupations are finding 14% fewer job openings than pre-ChatGPT. This isn’t because anyone is getting fired, but rather because companies are quietly not backfilling entry-level roles. Senior, highly educated, well-compensated workers in exposed fields show zero measurable unemployment increase. The paper also flags its own methodological stakes: if unemployment in top AI-exposure occupations doubled from 3% to 6% (the scale of the 2008 financial crisis) their framework would detect it. It hasn’t. Yet.
and then, even more news…
📉 Claude overtook ChatGPT on the App Store as the Pentagon drama played out in public. In the days following Anthropic’s public refusal to strip its AI safeguards for the Pentagon, Claude rapidly rose up the App Store charts while ChatGPT uninstalls spiked sharply. OpenAI reportedly lost about 1.5 million users after announcing its DoD contract. Read more