Ephemeris.
25 JUN 26
Today: persuasion that beats the professionals, an open model built for speed, and infrastructure that moves while you watch.
The model out-argued the experts.
Oxford, Stanford, the LSE and the UK AI Security Institute ran 18,978 conversations across 6,923 people, pitting AI against elite debaters and professional fundraisers on policy questions and real charity donations. The AI won every category — even against humans given prep time, training and cash incentives. Throttle it to human typing speed and message length, though, and the edge disappears.
America's open model, built to run.
NVIDIA's Nemotron 3 Ultra is the most capable US open-weights model yet — 550B parameters, 55B active — scoring 47.7 on the Artificial Analysis Intelligence Index, ahead of Gemma 4 and gpt-oss but behind China's Kimi K2.6. The pitch is throughput: past 400 tokens a second, it sits on the intelligence-versus-latency Pareto frontier and finished agentic tasks faster than every rival at every turn limit.
- 47.7
- Intelligence Index
- 55B
- Active parameters
- 400+
- Tokens / second
Thirty seconds. One prompt. No cuts.
ByteDance's Seedance 2.5 renders a full 30-second clip at native 4K from a single prompt, with 10-bit colour and sound generated in the same pass so action and audio line up on their own. It accepts up to 50 reference inputs and shows a scene as a 3D white-box before committing to a full render. Enterprise beta now, public release targeted for early July — but not in the US, where the Hollywood IP fight still bites.
Change the engine while the car keeps driving.
Cloudflare opened Self-Managed OAuth to every developer — but the engineering lesson is how they swapped the OAuth engine underneath live traffic. A blue-green cutover queued every revocation during the upgrade window and replayed it afterward, so no session was lost. Indexes were rebuilt with CREATE INDEX CONCURRENTLY to dodge the exclusive locks that would have frozen active users, and they upgraded Hydra 1.x first to learn its behaviour before the riskier 2.x jump.
revocation queue → captured live, replayed after cutover
CREATE INDEX CONCURRENTLY → no locks, no frozen users
Hand the VPN migration to an agent.
The Cloudflare One stack ships the migration as agent skills rather than a slide deck. Point an agent at your network and it inventories your VPN apps, maps each one to the right Zero Trust primitive, and writes a deployment sequence — backed by the same playbook Cloudflare used to move enterprises off Zscaler and Netskope in hours, not months. The skills live on GitHub and reach a live API through the company's MCP server.
- Inventory every VPN application in place
- Map each to a Cloudflare Zero Trust primitive
- Generate the recommended deployment sequence
- Deploy through the MCP server's live API
That's today.
Five reads on persuasion, open weights, generative video, and the plumbing under your auth. Sourced from Import AI, Artificial Analysis, The Next Web via @denissexy, and Cloudflare. Each pick had to clear one bar: would it change what a senior engineer or founder does this week.