Peer review just sided with AI. A Harvard Medical School study published in Science on April 30 found OpenAI’s o1 hit 67% triage diagnostic accuracy on 76 real ER patients at Beth Israel Deaconess Medical Center, against 55% and 50% for two human attending physicians. On a separate set of 143 complex clinical vignettes, o1-preview included the correct diagnosis 78.3% of the time and a helpful diagnosis in 97.9% of cases, with every output assessed double-blind by two attending physicians who didn’t know which came from the model. Lead authors Arjun Manrai, Adam Rodman and Thomas Buckley still call for prospective trials before clinical deployment—and ER physicians have publicly noted the human comparison group was internal-medicine attendings, not ER specialists, which complicates the headline. AAAS journals and double-blind grading shift the burden of proof: hospital systems can no longer treat clinician-AI benchmarks as marketing claims when the data is sitting in Science.
The same week the labs stopped selling raw model access and started shipping vertical products. Mistral launched Vibe remote agents alongside Medium 3.5, a 128B-dense, 256K-context flagship that hit 77.6% on SWE-Bench Verified—ahead of Devstral 2 and Qwen3.5 397B—at $1.50 per million input and $7.50 per million output tokens, with open weights, a modified MIT license, and a 4-GPU self-hosting footprint. Vibe runs async in cloud sandboxes, opens GitHub PRs on its own, and ties into Linear, Jira, Sentry, Slack and Teams under a new Le Chat “Work Mode” layer. Two days later, Anthropic moved Claude Security into public beta, an Opus 4.7-powered scanner that traces data flows across files, returns patches with confidence, severity, impact and reproduction steps, and exports to Slack, Jira and CSV without a line of API code. Mistral takes the open-weights coding-agent flank; Anthropic takes the vertical-security flank. The era of selling tokens is over—the labs are now selling workflows.
Underneath those products, the institutional layer is rebuilding from both ends. IBM’s 2026 CEO study—2,000 chief executives across 33 geographies and 21 industries—found 76% of organizations now have a Chief AI Officer, up from 26% in 2025, with 64% of CEOs comfortable making major strategic decisions on AI-generated input and an estimate that 48% of operational decisions where guardrails can be codified will be AI-autonomous by 2030. The same study projects 29% of employees will need reskilling for a different role by 2028 and 53% will need upskilling, even as 83% of CEOs insist AI success is people-first. Pushing the other direction, the Academy ruled that AI-generated acting performances and AI-written screenplays will be ineligible at the 99th Oscars on March 14, 2027, requiring acting roles to be “demonstrably performed by humans with their consent” and screenplays to be “human-authored,” with producers certifying both under penalty of disqualification, while AI tools remain eligible in VFX, sound design and editing. The institutional layer is arriving from both directions at once: the C-suite codifies AI authority while the creative industries codify human authorship.
The Pulse
xAI launches Grok 4.3 at $1.25/$2.50 per million tokens with a new voice cloning suite
Always-on reasoning and a 40% price cut — xAI is trying to win on the bill, not the bench.
Anthropic in early talks to buy DRAM-less inference chips from UK startup Fractile
Anthropic is shopping for an exit from Nvidia’s memory tax — the silicon shortage just became a strategy.
Inference is giving AI chip startups a second chance to make their mark
The training era rewarded scale; the inference era rewards diversity — and Nvidia’s moat finally has a side door.
OPAQUE acquires Abu Dhabi-developed cryptographic AI tech from TII
Confidential AI just went cross-border — the UAE is exporting crypto IP, not just buying it.
That’s the signal.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.