More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
What’s worth your attention in AI right now? As of October 31, 2026, three developments tell very different stories about where the technology is going: deeper into scientific research, further into unusual computing environments, and closer to the everyday problem of understanding what someone said. Let’s start with the one that carries the highest stakes.
On October 30, OpenAI published a system-card addendum for GPT-Rosalind-6, a model offered to approved organizations for biology, drug discovery and related research. It isn’t a new set of model weights: OpenAI says it uses the same weights as its GPT-6 Astra model, without additional training for this release. The difference is how it’s made available for advanced life-sciences work. Applicants must pass a review of their proposed use, governance and security controls, and customers must put qualified humans between model outputs and consequential research decisions.
Why pay attention? In OpenAI’s GeneBench evaluation, which tests planning and carrying out analyses of biological data, Rosalind-6 succeeded on 37.7 percent of single attempts at the highest reasoning budget, versus 16.7 percent for Rosalind-5.5. That is a substantial improvement on this particular test, not proof that it can reliably run a real laboratory project. The evaluation used a configuration different from production.
There’s a reason this isn’t simply being handed to everyone. OpenAI classifies the launch as having High capability in its Biological and Chemical risk category, though below its Critical threshold. It says the model is intended to help with legitimate research, including some dual-use topics, while continuing to refuse requests that would enable biological weaponization. The consequential development is not just a better benchmark score. It’s an attempt to make more capable scientific assistance available under tighter controls—and a real test of whether those controls hold up in use.
Now, a sharp change of scenery. Google reported on October 30 that its Project Suncatcher prototype satellite had returned an AI-generated response from a data-center-class TPU in orbit. Engineers sent a prompt to an onboard Gemma 4 model during a ground-station pass, asking for a haiku about space. The reply came back as the satellite passed near Svalbard, Norway. It began: ‘Blue marble below.’
A poem from orbit is delightful. Its technical significance is narrower, but still striking: Google says it has powered up the satellite’s computing hardware and run early machine-learning tests. This is a prototype milestone, not a working orbital AI data center. Google says it still needs to test how the chips perform under thermal stress and radiation. In its October 30 update, it described two interconnected satellites in 2027 as a *next milestone*, not an accomplished result. The interesting question now is whether a successful prompt and response can become reliable computing under space conditions.
Back on Earth, NVIDIA’s October 28 release of Nemotron 3.5 Transcribe tackles a less glamorous problem that millions of people recognize: a transcript can get the words roughly right and still be wrong for the job. Was that a person’s name? Should the transcript keep every ‘um,’ or clean them up? Who spoke during an interruption?
NVIDIA says its open-weights model can follow an instruction to produce different transcription styles, identify speakers, translate speech and provide word-level timestamps on request. It supports 40 languages and recordings up to an hour long. The company reports strong results on public speech-recognition leaderboards and says giving the model a list of expected names sharply reduced name errors in one benchmark. Those are NVIDIA’s reported test results, not a guarantee for every noisy meeting or accent. But the practical idea is compelling: instead of accepting one rigid transcript, you ask for the version you actually need.
Put these stories together and the pattern isn’t simply ‘AI got bigger.’ One development asks who should be allowed to use powerful scientific assistance. Another asks where AI computing can work at all. The third asks whether a model can adapt to a human’s task rather than making the human adapt to its output. The next results that matter will be measured in research practice, sustained performance in orbit, and transcripts people can trust—not just launch-day demonstrations.