More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
If an AI service says you’re using a particular model, how would you know? A polished answer won’t tell you whether a cheaper model was quietly substituted. That question is behind FlashProve, a research prototype Anthropic released on November second.
The idea is to attach a cryptographic proof to a model response. Someone checking it could confirm that the response matches a specified set of model weights without seeing those weights or rerunning the enormous model. That could matter to customers using third-party AI services and to independent auditors trying to establish which model they actually tested.
Anthropic demonstrated the system on a quantized version of Llama 4 Maverick. In its reported test, proof generation reached about 110 tokens per second, with proofs around 19 megabytes. But producing them took roughly 70 to 95 times the GPU time of ordinary inference in Anthropic’s comparisons. The prototype also checks a narrower kind of computation than many everyday chats, and its fast executable code does not carry the same formal guarantee as its mathematical specification. So this is promising work on model identity—not a feature that makes an AI answer true, safe, or automatically trustworthy.
The second development is less abstract and much closer to a school desk. In a November fourth update, Google said teachers can add Guided Learning activities to Google Classroom assignments. The rollout began November third and is expected to take more than 15 days; access depends on the school having Gemini enabled.
Teachers can select Classroom sources to give the activity curricular context. Rather than simply supplying an answer, Guided Learning is designed to offer step-by-step guidance and hints when students get stuck. Teachers can see AI-generated summaries of class and individual activity, as well as individual student chat transcripts.
That combination is what makes this worth watching. An AI tutor built into an assignment could give a student help during homework while showing a teacher where the class is struggling. But it also makes decisions about student conversations and teacher oversight part of the everyday classroom workflow. Google describes how the feature works; its availability is not, by itself, evidence that students learn more from it.
Taken together, these releases point to two different tests for useful AI. Can we establish which system actually gave the answer? And can we design the interaction so the person receiving it learns rather than merely copies? Neither question is settled. Both are becoming practical.