The Omniscient Bulletin · 2026-05-02
The Omniscient Bulletin — May 2, 2026
The frontier reshuffles, a DeepMind legend bets against human data, distillation lands in a federal courtroom, and AI's generational toll becomes measurable.
A week that began with back-to-back frontier releases ended in a California courtroom — with Elon Musk conceding, under oath, that xAI built Grok partly by learning from OpenAI's own models. Elsewhere, the economics of AI's labor disruption are sharpening into hard numbers, and London is quietly becoming the address of choice for post-LLM bets.
Frontier
GPT-5.5 arrives as OpenAI eyes a 'super app' future
OpenAI released GPT-5.5 on April 23, rolling it out to Plus, Pro, Business, and Enterprise users across ChatGPT and Codex. President Greg Brockman called it "a big step towards more agentic and intuitive computing" and flagged the company's ambition to unify ChatGPT, Codex, and its AI browser into a single super app. Chief scientist Jakub Pachocki added a pointed note: "I would say the last few years have been surprisingly slow." The pace of model releases — November, December, March, now April — suggests OpenAI intends to keep proving him right.
Frontier
DeepSeek V4 Preview: 1.6 trillion parameters, frontier prices undercut again
One day after GPT-5.5 launched, DeepSeek released preview versions of V4 Flash and V4 Pro — mixture-of-experts models with 1M-token context windows. V4 Pro tops out at 1.6 trillion total parameters (49 billion active), making it the largest open-weight model available. Pricing undercuts every comparable closed-source model: V4 Pro costs $0.145 per million input tokens versus GPT-5.5's higher rate. The lab acknowledges a knowledge-benchmark lag of "approximately 3 to 6 months" behind frontier leaders — and released the weights regardless.
Industry
Musk admits in court that xAI used OpenAI models to train Grok
Testifying in a federal Oakland courtroom on April 30, Elon Musk confirmed that xAI used model distillation — training on outputs from OpenAI's models — to improve Grok. When asked directly if that was a "yes," he said, "Partly." The admission is significant given that OpenAI and Anthropic have spent months accusing China's DeepSeek of the same practice. Musk cast himself as the safety-minded party in the dispute, arguing OpenAI has become profit-driven. The testimony sharpens a brewing industry debate over whether distillation is legitimate knowledge transfer or proprietary theft.
Research
AlphaZero's creator raises $1.1B to build AI that needs no human data
Ineffable Intelligence, a British lab founded months ago by former DeepMind researcher David Silver, closed a $1.1B seed round at a $5.1B valuation. Sequoia and Lightspeed led; Index Ventures, Google, Nvidia, and the U.K.'s new Sovereign AI fund also participated. Silver — who built AlphaZero, which mastered chess and Go purely through self-play — wants to create a "superlearner" that builds understanding entirely from self-generated experience, bypassing human-labeled training data. The company's site likens the ambition in scope to Darwin's theory of natural selection.
Policy
EU AI Act deal collapses, leaving August deadline in doubt
After roughly 12 hours of trilogue talks on April 28, EU member states and Parliament lawmakers failed to agree on the AI Omnibus — amendments designed to ease compliance and push high-risk AI deadlines to 2027–2028. The sticking point: whether AI embedded in regulated products (medical devices, cars, machinery) should be exempt from the AI Act's additional conformity requirements. Talks resume in mid-May, but if no deal is reached before the original August 2, 2026 high-risk AI deadline, unprepared companies will have no runway.
Labor
Goldman Sachs puts a number on AI job loss: 16,000 net per month, Gen Z hardest hit
A Goldman Sachs note by economist Elsie Peng estimates AI substitution eliminated roughly 25,000 U.S. jobs per month over the past year while augmentation effects added back around 9,000 — a net loss of 16,000 monthly. The damage falls disproportionately on workers under 30, concentrated in routine white-collar roles — data entry, billing, customer service — that AI automates most readily. Goldman notes the figures likely undercount new infrastructure jobs, but the entry-level wage gap is already widening by an estimated 3.3 percentage points per standard deviation of AI exposure.