The brief · 16 June 2026
Seven fronts advancing at once today: banks are multiplying their AI models rather than picking one, a biotech turns its failure into a blueprint, and a Chinese model holds a million tokens without flinching. Brussels stalls on the AI Act but won't budge on deepfakes.
A failed trial costs hundreds of millions and usually ends up shelved for good. This biotech instead reused the trial data to train a predictive model, which changes what a pharma failure can actually be worth. Whether other labs follow suit or this stays a one-off is still an open question.
practitioners › STAT News does not name the biotech or the model: worth watching once the full paper comes out.
STAT News →ABB Robotics is pairing PSYONIC's Ability hand with its GoFa cobots. The concept: train industrial robots on grasping and movement data from real prosthetic users, not simulations. A robot handling fragile or irregular parts on the factory floor becomes more reliable, without engineers hand-coding every motion.
practitioners › Partnership announced June 16, 2026: PSYONIC's Ability hand integrated with ABB's GoFa cobots, using data from real prosthetic users.
ABB Robotics →Wolfram has released version 15 of its language and Mathematica, with an AI assistant built into every notebook. The goal: stop copy-pasting between ChatGPT and your symbolic computations, and let AI suggest solving methods right inside the tool. For anyone doing scientific modeling or engineering work, this shifts where the work actually begins.
practitioners › A new ModelFit superfunction and AI-powered methods for DSolve ship natively, and Mathematica now also plugs into external AI environments.
Wolfram →An open model that keeps a million-token context without degrading means you can paste an entire codebase or full technical doc and get answers that stay coherent from start to finish. Z.ai also claims the top open-source score on coding and long-task benchmarks. Still unclear how it holds up against Claude or GPT on real-world debugging, not just leaderboard numbers.
practitioners › GLM-5.2, Z.ai, released June 16 2026, open-source, 1M-token context window with no loss, open-source SOTA on coding and long-horizon tasks.
Z.ai →The June 16 vote pushes some high-risk obligations back to late 2027, possibly August 2028. Companies get breathing room to comply, but the ban on non-consensual intimate deepfakes kicks in immediately. If you're tracking generative image AI, this is the first real legislative crackdown on nudifier apps since they took off.
practitioners › Vote: 423 in favor, 57 against, 174 abstentions. Watermarking rules also get pushed back, and machinery/AI obligations are merged to avoid duplicate compliance work.
www.europarl.europa.eu →Washington restricted access to certain Anthropic models, and now London is asking whether its national security is too dependent on tech that another country can switch off overnight. Baroness Kidron pressed the government to take a stance on sovereign AI and to give the AI Security Institute real legal teeth. Nothing changes for you tomorrow, but this signals where the UK's next AI legislation is headed.
practitioners › Debate on 16 June 2026, House of Lords, oral question from Baroness Kidron on US restrictions on Anthropic.
hansard.parliament.uk →