i spent Monday and Tuesday, and a bit of Wednesday, at a spa outside Hamburg. i brought an old Nokia and left the rest of the tech alone. it felt good to be away from it for a few days.
meanwhile, GPT-6 Sol came out. this week i moved the agents i run on Hermes over from Astra, which had been eating through my usage allowance in a day. Sol feels fast and strong so far.
OpenAI says Sol and Luna's API prices are 50% below their GPT-5.6 promotional pricing. Anthropic released Opus 5.5 on the same day, with input and output token prices 20% below Opus 5. i haven't tried Opus yet. coming back from a few days offline to a model i can already run across my agents, at a lower price tier, made the pace of all this feel a little absurd.
enjoy this edition!
The Build
tools, models, and dev releases
a product motion-graphics workflow for Opus 5.5, from a friend
a friend tried the "one prompt" approach to making a product motion-graphics video with Opus and got a mediocre result. they went back through videos they liked and put together a better process. i haven't tried it myself, but the steps are worth stealing:
1. pick one or two launch-film references from whatships.com. show Opus the pacing, type and transitions instead of describing a style from scratch.
2. use HyperFrames or Remotion to build the scenes in code and render an MP4. the video stays editable rather than becoming a screen recording of an HTML preview.
3. bring screenshots of your real product; use 21st.dev for UI components where they help. don't let the model invent an interface your product doesn't have.
4. give Opus the logo, colors, fonts, screenshots, references and a short description of the video. ask for three storyboard directions before choosing one.
5. review one still frame per scene before animating. fix the composition while a change is cheap.
6. render, then give precise notes: slow a zoom, cut hard between scenes, or push in on a button. "make it better" gives the model too much room to guess.
also worth opening:
- OpenCodeReview checks pull requests with rules and an agent - run Alibaba's open review tool against code and inspect line-level findings for issues such as injection, null handling and concurrency. (GitHub)
- WeKnora turns documents into a searchable knowledge base - self-host Tencent's open platform to test retrieval, question answering and a maintained wiki against your own files. (GitHub)
- Hindsight gives agents a memory service - inspect the server, integrations and cookbook to test whether an agent recalls past decisions across runs. (GitHub)
Founder Fuel
funding, acquisitions, and market moves
- Nscale secures $3.36B before a planned US IPO - the British AI cloud company gets $2.36B immediately and expects another $1B from Nvidia in November, with notes set to convert after listing. (More)
- Ando raises $20M for a team chat with agent identities - its app gives agents inboxes, channels and conversations alongside human workers, rather than placing them behind a separate bot interface. (More)
- Lightspeed targets $250M for an India AI fund - the proposed early-stage vehicle is half the size of its previous India fund, according to a letter seen by TechCrunch. (More)
Longevity & Health
aging, biotech, and health research
- a blood test detects some cancers but misses others - the PATHFINDER 2 study followed adults aged 50 or over; its multi-cancer test found 39.3% of cancers diagnosed in the study window, and longer randomized trials are still needed to establish clinical benefit. (More)
- NMN raises a blood marker, not measured function - in a 24-week randomized trial in older adults, the supplement raised NAD-related metabolites but did not significantly improve the primary hearing outcome or physical performance. (More)
- a brain atlas spans infancy to old age - Nature reports that researchers mapped gene activity in prefrontal cortex samples from nearly 1,500 donors, creating a reference for studying aging and Alzheimer's disease. (More)
The World
regulation, policy, and geopolitics
- the EU seeks input on data-centre performance standards - the European Commission opened a consultation through December 14 and plans a legislative proposal in 2027; a separate EU rating and label package was published this week. (More)
- Brussels' industrial AI adviser pushes deployment over model rivalry - POLITICO reports that Jim Hagemann Snabe urged commissioners to find where AI can lift productivity in health, manufacturing and public services. this is advice, not an adopted EU policy. (More)
Go Deeper
long reads, podcasts, and documentaries
- improved token efficiency for longer agent runs - Cursor | Jediah Katz, Connor O'Keefe and Calvin Yee. an engineering breakdown of tool loading, cache boundaries and why the company reports a 7% reduction in user token costs. (Read)
- priorities and principles for third-party assessments - OpenAI | OpenAI. a detailed proposal for outsiders to test safety claims and safeguards, with the unresolved question of how independent those tests will be. (Read)
- heaven's melodic strains - n+1 | Colin Vanderburg. a tour of the people buying six-figure sound systems in pursuit of a cleaner signal. (Read)
after those days with the old Nokia, Vlady and i have been talking about focus. we're putting more of our attention into one project instead of trying to move several forward at once. if you struggle with focus too, hit reply and tell us what actually helps. we can compare notes in a future issue.
have a good week.
stay curious,
AI BEAVERS