Saturday, June 20, 2026

Good Saturday, NOLA. June 20th is a quieter news day, but we've got a legit research finding worth thinking about, a major talent move at Anthropic, and a benchmark comparison that challenges what we think we know about model sizes. The week's story: bigger doesn't always mean better, and John Jumper just joined Anthropic — the Nobel Prize-winning chemist who showed how AI could solve structural biology.

Big Moves & People

John Jumper Joins Anthropic

Jumper announced he's joining Anthropic, bringing his Nobel Prize-winning work on protein folding and AI-assisted chemistry to the team. This is a major signal: Anthropic is building for scientific applications, not just chat. If you've been following the talent reshuffling all week — Noam Shazeer to OpenAI, Barret Zoph's departure — this shows the labs are all fighting for deep technical expertise.
Twitter / Community Signal

Research & Benchmarks

GPT-5.5 Hallucinates 3x More Than Smaller Open Model GLM-5.2

A developer benchmarked GPT-5.5 against GLM-5.2 on the same task and found the frontier model made 3x more factual errors. This isn't just "open models are catching up" — it's a real reminder that model size, training cost, and accuracy aren't linearly correlated. If you're building something where hallucinations matter, this is worth a read. The technical detail: both models were tested on the same evaluation set with identical prompting.
Hacker News

Using AI to Improve a Challenging Reaction in Medicinal Chemistry

OpenAI published a case study showing how AI helped optimize a tricky synthesis step in pharmaceutical research, improving yield and reducing time. This is the kind of applied science win that matters: not just "AI can think about chemistry" but "AI solved a real problem faster than humans alone." Given Jumper's move to Anthropic, expect more of these stories.
OpenAI Blog

Business & Operations

Companies Rein in AI Usage as Costs Strain Budgets

Enterprise spending on frontier models is being scaled back as API costs and token prices start to bite. This is the reality check after the hype: companies are learning that running GPT-4 or Claude on everything costs real money. Local models and smaller, cheaper alternatives are getting renewed interest. If you're building an AI product, this is a good reminder: your users care about cost-per-query as much as quality.
Financial Times

Interesting Reads

Is AI Ruining Our Skills? Early Results Are In

Nature published a roundup of research on how AI tools affect human skill development, and the early signals are mixed. Some studies show knowledge retention issues when people offload too much to AI; others show AI as a scaffold actually improves learning. Not doom, not hype — just nuance. Worth a read if you're thinking about how to integrate AI into workflows without breaking the humans.
Nature

Today’s Sources