Morning Digest, September 8, 2026
3 newsletters, 3 overlapping stories
Top Stories
An AI-designed drug shows early signs of slowing aging
(2 newsletters)
Insilico Medicine published trial data on rentosertib, a drug its AI designed to treat idiopathic pulmonary fibrosis, showing that treated patients read as biologically younger on all six of the “aging clocks” used to score them. The trial was small, 42 patients, with blood samples compared against the UK Biobank, and one group’s age estimates dropped by roughly 2.7 to 3.5 years. Notably, the dose with the clearest aging signal was not the dose with the biggest lung-function gain, which hints at an effect beyond treating the disease itself. Regulatory approval is still years away even for sick patients, so this is a signal, not a product.
ByteDance joins the race to build world models
(2 newsletters)
ByteDance is preparing a real-time spatial video model that generates interactive virtual worlds, built on top of its Seedance video AI, with founder Zhang Yiming personally running the project. The worlds are meant to respond to Pico VR headset users’ voices and movements, putting the company directly against the other frontier labs chasing world models. A launch could come as soon as next month.
Building got cheap, thinking did not
(2 newsletters)
Two separate pieces landed on the same argument from different angles: when agents remove the friction of implementation, weak ideas stop dying on contact. The product side calls it “slop-creep,” the accumulation of features, abstractions, and infrastructure that exist only because they were cheap to produce, a problem that compounds inside large companies where software never gets removed. The engineering side reaches the same place from data: a study of how agents use test and verification techniques found that telling an agent to use a specific technique mostly produces the tests it would have written anyway, wrapped in the new framework, with the actual value of the technique left on the table. The common thread is that the scarce resource has shifted from execution to judgment.
Also Worth Knowing
- OpenAI published an internal look at its own agent usage. Coding agents now log 3.1 workdays for every human workday, the typical researcher burns over $600 a day in agent tokens (90th percentile above $7,000), token output is up 124x since December, and about 80% of researchers run four or more agents at once.
- Voters’ fear of AI is now bipartisan. An NBC News poll of 7,105 adults found 70% more worried than excited, 69% opposed to a local data center, and 44% trusting neither party on AI policy, even as 52% report using the tech regularly.
- ChatGPT may soon learn your writing style from Slack or Gmail. Some users have had the unannounced feature for weeks, which pulls writing samples from connected apps so the model can mimic your voice.
- XPeng started production of its IRON humanoid robot. The line runs with over 80% of core processes automated, with mass production targeted by year end, while Tesla’s Optimus stalls.
- Finding and exploiting bugs is becoming cheap. As offensive capability gets abundant, the defender’s hard problem shifts to patching critical bugs fast, and AI-assisted formalization only partly closes the gap.
- UN rights chief warns on AI concentration. Volker Turk said he shares fears that advanced AI could pose an existential risk, and that a few men hold near unlimited power over it.
- The Seattle Times and Newsday sued OpenAI and Microsoft over training data, arguing AI is breaking journalism’s economics beyond repair.
- Four questions about AGI. The argument: AGI announcements amount to labs declaring they have nearly reached a destination they defined themselves, and the industry would do better making a long list of specific things reliable enough to trust.
- How the leading AI companies do product. Cheaper execution and faster experiments do not replace product expertise; the advantage moves toward judgment, customer understanding, taste, and verification.
Quick Hits
- Applied AI often fails because the process is broken first: speeding up a bad workflow creates little value, so redesign the workflow before deciding where software, agents, and humans each belong.
- Model and tool releases: OpenBMB open-sourced MiniCPM5-2B, now first among open models under 4B parameters on the AA Intelligence Index; Google shipped Gemini 3.8 Flash; Meta released Muse Voice Transcribe for real-time speech to text; xAI opened a Grok Bot marketplace.
- Foldables: Huawei launched the Mate XT 2 tri-fold, a 10.2-inch display, two days ahead of Apple’s foldable iPhone.
- Tesla Cybercab specs surfaced: Supermanifold V3, no rare earth magnets, plus custom thermal hardware.
- Prompt archaeology: a read of Claude’s Fable 5.1 system prompt tracking how model quirks and product design shift between versions.
- Craft reading: how to name things argues precise naming exposes design flaws, and a definition of done for AI features makes the point that evals tell you whether AI works, not whether it creates value.