The Omniscient Bulletin · 2026-05-26
The Omniscient Bulletin — May 26, 2026
OpenAI's CEO walks back his white-collar apocalypse warnings; a free tool strips Meta and Google safety controls in ten minutes; Huawei reframes the chip race around system-level scaling; the routing layer for the model market goes from $547M to $1.3B in a year.
The CEO walks back his warnings the week the labor data starts to confirm them. The frontier lab uses its first-ever profit quarter to insist the jobs apocalypse is not happening. The labor data is moving in one direction, and the people who built the technology have decided that direction is no longer convenient to talk about.
Policy
Free tool strips safety guardrails from Meta's Llama 3.3 and Google's Gemma models in under ten minutes
The FT and AI safety group Alice published an investigation Monday showing Heretic, a free open-source tool, strips guardrails from Meta's Llama 3.3 in under ten minutes on consumer hardware. FT testing of a Gemma 3 model produced ricin dosing, chlorine gas instructions, credit-card malware, and child sexual abuse material. Heretic's creator also stripped Google's Gemma 4 within ninety minutes of its release. Modified models created with the tool have been downloaded 13 million times. Safety controls at the model layer cannot restrict open-weight misuse once the weights are out.
Irish Times / FT investigation→
Labor
Sam Altman walks back the AI white-collar jobs apocalypse: 'delighted to be wrong'
At Commonwealth Bank of Australia's Accelerate AI conference in Sydney, OpenAI CEO Sam Altman told CBA chief Matt Comyn he had overestimated how fast AI would eliminate entry-level white-collar work: 'I'm delighted to be wrong about that.' Altman's own grade on his record: 'roughly right on technological predictions and pretty wrong on the social and economic implications.' The reversal arrives the same month BLS data showed the eighteen most AI-exposed occupations shrinking 0.2% against 0.8% overall U.S. employment growth. The frontier CEO and the labor data are now telling opposite stories.
Commonwealth Bank of Australia→
Compute
Huawei proposes the Tau Scaling Law and stacked-logic chips as its answer to the sanctions wall
At IEEE ISCAS in Shanghai on Monday, He Tingbo, president of Huawei's semiconductor business, introduced the Tau Scaling Law: abandon transistor miniaturization, instead compress signal-propagation delay across devices, circuits, chips, and systems. The companion LogicFolding architecture stacks chip layouts onto multiple layers and ships in this fall's Kirin processors. Huawei projects 1.4nm-equivalent transistor density by 2031 despite being locked out of EUV lithography. It is a concession the geometric race is over, and a bet that system-level scaling can substitute for litho leadership.
Industry
OpenRouter raises $113M from Alphabet's CapitalG at $1.3B as weekly tokens hit 25 trillion
AI inference router OpenRouter closed a $113M Series B led by Alphabet's CapitalG, with NVentures and co-investors including ServiceNow, Snowflake, MongoDB, and Databricks Ventures, plus existing backers a16z and Menlo. Weekly volume has run from 5 trillion tokens six months ago to 25 trillion today, with 8 million users routing across 400-plus models from Anthropic, Google, OpenAI, xAI, and DeepSeek. Alphabet's growth fund is now bankrolling the layer that disintermediates Alphabet's own model. Full analysis at Omniscient Media.
Agents
Catena Labs raises $30M and applies for a national trust bank charter purpose-built for AI agents
Catena Labs, founded by Circle co-founder Sean Neville, has raised a $30M Series A led by Acrew Capital and a16z crypto, and simultaneously filed for a national trust bank charter with the OCC. The pitch: financial rails purpose-built for AI agents, with programmable accounts, per-agent spending limits, payee whitelists, and rails that span traditional payments and onchain stablecoins. If the charter is approved, Catena would be a regulated fiduciary for autonomous agents, not a fintech wrapper around someone else's bank. The category is no longer hypothetical.