ma2tic

The brief · 28 July 2026

The AI brief, 2026-07-28

3 min read J / K to navigate

Anthropic says no to a ban on Chinese open models

Dario Amodei just settled a debate that had been simmering in Washington for weeks. His stance shifts the calculus for any team already running DeepSeek or Qwen in production: no legal crackdown appears on the horizon in the US, and the protectionist camp just lost the heavyweight backer it was counting on.

practitioners › Position published July 27, 2026; no announcement of technical restrictions on Anthropic's own closed models.

www.anthropic.com →

OpenAI Study Shows ChatGPT Is Expanding What Jobs Actually Involve

Employees using ChatGPT aren't sticking to their job descriptions anymore. They're picking up adjacent skills, which is blurring the lines between roles on the same team.

practitioners › Internal OpenAI study based on ChatGPT usage data; scope and methodology not detailed in the announcement.

OpenAI →

A cortical implant learns just by listening to the brain at rest

Neuroprosthetics are starved of labeled data because every perception trial burns clinical time. This work shows a model can instead be trained on hours of spontaneous brain activity, with the patient performing no task at all, and still be used to decode what someone perceives.

practitioners › A masked autoencoder trained on 14.6 hours of spontaneous activity in the V1 visual cortex of a blind patient reached 84.1% decoding accuracy on a general psychometric task.

arxiv.org →

An algorithm that reads your motor thoughts works well on average, poorly for you specifically

Motor imagery brain-computer interfaces promise to let users control devices by thought alone, but each brain responds differently to the same decoding methods. A research team tested tens of thousands of combinations across three datasets and confirmed that the best-performing algorithm changes from person to person almost every time.

practitioners › 216,714 evaluations tested, 42 different winning pipelines across 52 subjects in Cho2017, arXiv study from July 24, 2026.

arXiv →

A Simple Boolean Network Beats a Neural Net at Predicting Gene Behavior Over Time

The model that's most accurate at a single step isn't necessarily the one that stays faithful over an entire trajectory. For anyone building predictive models in biology or elsewhere, this is a reminder that a good point-in-time score guarantees nothing about long-term behavior.

practitioners › Threshold Boolean network: perfect accuracy on the binarized trajectory, versus 0.708 for Random Forest regression.

arXiv →