They deleted it from a yeast cell. The cell broke. Then the human version fixed it.🌿

For most of modern biology, finding a protein's relatives meant comparing its genetic spelling. Similar sequence, assumed relation. The problem is that evolution can preserve a shape while scrambling the letters, so anything that drifted far enough simply vanished from view.

Researchers at the Sylvester Comprehensive Cancer Center at the University of Miami searched by shape instead. Using AI-predicted structures, they compared more than 214 million proteins in three dimensions and pulled out hidden members of the GPCR family, the receptor class that roughly a third of all approved drugs already target. One of them, TM184C, turns out to sit inside cells on vesicles that travel along microtubules into the bridge-like projections where neighbouring cells trade metabolites, vesicles and even mitochondria.

They then went and proved it. The team deleted the equivalent protein in yeast, and the cells broke. They inserted the human version and the cells recovered, a function held intact across roughly a billion years of evolution. The AI changed what the scientists could see. The wet lab still did the proving, and that order is the whole story.

What we are still uncertain about: a newly named receptor is a starting line, not a drug. Whether TM184C or any of its hidden siblings proves targetable is years of work away, and "related to a druggable family" has disappointed before. Senior author Daniel Isom drew the line better than we could: "AI cannot be blindly trusted, but can lead to really big things in the hands of experts and prepared minds."

Source: AZoLifeSciences, University of Miami

Governor Maura Healey signed Executive Order 658 on 8 September. Any data centre above 25 MW must generate clean power on site, fund new generation nearby, or pay into a Ratepayer Protection Fund that flows back to ordinary customers. The quieter clause is the sharper one: state agencies may no longer sign NDAs with developers, and communities are told not to either.

Source: TechCrunch

"Right-Sizing AI at Work," from a joint Trinity College Dublin and TU Dublin centre, found 47.4% of respondents use AI daily, 64.7% feel confident with it, and 41.5% do not believe it could replace significant parts of their role. Professor Taha Yasseri's reading: task and workflow transformation, not immediate large-scale displacement. A national survey of workers, not a vendor statistic.

Source: Trinity College Dublin

Coefficient Giving opened Project Tailwind on 9 September: $200k to $2M at pre-seed, $2M to $20M and beyond at seed, for people who want to build AI safety organisations. It ships with a catalogue of specific organisations they wish existed, from rogue-incident detection to misalignment red teams, and pays bounties for successful founder referrals. Applications are open.

Source: Coefficient Giving

The tax nobody was charging on purpose

Hugging Face's tokenizers library shipped v0.23.2 this week with a new component called ParityBpeTrainer. That is an unglamorous sentence about an unglamorous piece of plumbing, and it quietly closes one of the most regressive pricing quirks in AI.

A tokenizer chops text into the units a model actually reads, and standard ones are trained on frequency. Dominant languages get efficient chunks. Lower-resource languages get tokenizations that are disproportionately longer, morphologically nonsensical, or padded with unknown-token placeholders. Because models are priced and context-limited per token, a Swahili or Telugu speaker literally pays more and fits less into the same window to say the same sentence. The parity-aware variant optimises each merge step for whichever language is currently worst compressed. The underlying research, from the Swiss AI Initiative and presented at ACL 2026, reports up to an 89% reduction in tokenization inequality, measured as the Gini coefficient of per-language token costs, with negligible loss of global compression and no systematic downstream degradation. The research is not new. What is new is that it stopped being a paper and became a default anyone can call.

Our PM + Risk Manager lens

This was a pricing and product-equity bug that lived for years in a layer with no owner. It was on nobody's roadmap because it was in nobody's product. Worth asking of your own stack: which shared component silently sets a cost or quality floor for a segment of your users, and whose name is against it? The second lesson is about how fairness fixes actually ship. This one did not arrive as a policy, a pledge or a paragraph in a model card. It arrived as a default in a library, which is the only form most teams will ever adopt without being asked twice.

The cost transfer here was real, quantifiable and entirely unmeasured. Nobody was doing anything wrong on purpose. There was simply no metric, so there was no finding, so there was no fix. That pattern generalises badly. A harm nobody has instrumented does not appear in a risk register, an audit or a regulatory filing, and its absence reads as an all-clear rather than as a blind spot. Note too that shipping a trainer is not the same as anyone using it. Existing models keep the tokenizers they were trained with, and retrofitting means retraining.

The next 12 to 24 months Watch two things. First, whether major open model families adopt parity-aware tokenizers at their next training run, since that is the only moment the fix can be applied. Second, whether per-language token cost becomes something buyers ask about in procurement. It is now cheap to measure and hard to argue with, which is usually what converts a fairness question into a contractual one. The wider version of the question is the uncomfortable one: what else is quietly priced this way, and who is holding the bill?

Source: Hugging Face, arXiv

OpenAI published a case study this week in which MIT researcher Beatriz Yankelevich used Codex with GPT-5.6 Sol to run superconducting-qubit experiments: choosing parameters on an uncalibrated six-qubit chip, finding transition frequencies, calibrating control and readout pulses, estimating coherence times, and iterating overnight. The interesting part is not quantum computing; it is the shape of the task. Calibration is fiddly, serially dependent work that sits between a researcher and the experiment they actually want to run, and it is exactly where an agent that will retry a thousand times without getting bored earns its place. Caveats worth keeping: a vendor case study, one researcher, one chip, and a human stepped back in whenever the signals got noisy.

→ Read more: OpenAI

There is a hidden theme in this issue: costs nobody was measuring. A protein family invisible because we were searching by the wrong feature. A token tax paid entirely by people who do not speak English. Data-centre terms negotiated under NDAs that a state has just banned. So: in your work, what is the cost that everyone quietly absorbs because nobody has ever put a number on it?

Hit reply. We read every response.