The Sequence Radar - Issue 936: Last Week in AI: Gemini Talks, Astra Practices Law, Figure Folds Laundry, and Crusoe Powers It All
TheSequence Jesus Rodriguez ● Covered by 10 sources
Google, OpenAI, and Figure all pushed AI closer to doing real work. The twist: the bottleneck is looking less like model IQ and more like context, timing, and power.
Based on reporting by TheSequence, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This week’s AI news read less like a model leaderboard and more like a systems diagram. Google worked on conversation itself, OpenAI narrowed a model to legal work, Figure pushed robots into unfamiliar homes, and Crusoe raised a fresh mountain of cash to keep the whole machine fed. The pattern is hard to miss: capability is only half the story. The rest is everything around it.
Google’s Gemini 3.8 Live and Live Extended Thinking are aimed at the awkward bits of voice agents, where a user keeps talking while the system is trying to do something useful in the background. The new models mix visual context with tool calls, and the Extended Thinking version can reason and speak at the same time. That matters because delays are not just annoying in a conversation; they break the illusion that the assistant is present at all. Google is treating latency as a design problem, not an afterthought.
OpenAI’s Astra for Law takes a different route. It wraps GPT-6 Astra in legal instructions and a legal search setup built for case law, statutes, regulations, court rules, and administrative decisions. On 200 questions from a private validation set, OpenAI says it hit 54% correctness, up from 38.7% when the system used ordinary web search. That is a real improvement, but not a miracle. It shows how much a strong model can gain when the information environment is tightened up, and how far it still has to go before anyone should hand it the keys.
Figure’s Helix 2.5 is the most physical version of the same idea. The company tested laundry, towels, and bed making in 30 unseen homes, and pretraining on its Index human-behavior dataset lifted complete-task success from 9% to 56% while the architecture and task-specific training stayed fixed. The homes were new, the objects were new, and the point was transfer: broad human data made the same robot brain a lot more useful. Even so, 44% failure is a stubborn number when the task is folding towels in somebody else’s house.
And then there’s Crusoe, which closed the initial part of a planned $3.9 billion Series F at a $30.9 billion post-money valuation. Its pitch is blunt: if AI keeps getting more capable, someone has to build the power, data centers, and cloud infrastructure underneath it. That is not the glamorous part of the story, but it is the part that decides how much of the glamorous part can actually run.
My take — AI-written commentary, not fact-checked reporting
The industry keeps pretending the model is the product, but this week made the opposite case. The real winners are the teams that can supply context, compute, and a place for the thing to exist without falling over. Open models may win the long game on adaptability, but closed systems are still the ones grabbing the revenue while everyone else argues about benchmarks over coffee.
Read more about this at: TheSequence