A mixed week of reading that kept circling back to one idea: the model is rarely the hardest part. I have been thinking about where capability actually meets the messy reality of organisations and classrooms, about a paper naming the tension between working and learning, about the human cost of detection tools, about a curious model that only knows the world up to 1930, and about a series arguing we should move faster on AI in schools rather than slower.
The model is not the bottleneck
Mustafa Suleyman has said AI will reach human-level performance on most white-collar tasks within twelve to eighteen months. Perhaps. But there is a mammoth gap between an AI that could do the work and an AI that has the information required to do the work in a specific office.
Applying AI in real organisations teaches the same lesson over and over: the bottleneck is the undocumented stuff. How decisions actually get made. Who signs off on what. Why this client is handled differently. Which spreadsheet is the source of truth this quarter. None of that lives in a system; it lives in the heads of the people doing the work. Until that institutional knowledge is surfaced, written down and made accessible to a model, a strong benchmark score does not translate into a job being done. Mapping the organisation is the hard part, and everyone selling AI knows it. It is worth reading the claim itself with that gap in mind.
The learning-performance paradox
Work AI is designed to reduce cognitive effort and produce outputs. Learning needs the opposite. Drop a work tool into a classroom and task scores go up while durable learning goes down. A new paper calls this the learning-performance paradox, and I find the name clarifying.
Its framework rests on three foundations. The pedagogical question asks whether the learner actually did the thinking. The adaptive question asks whether the system genuinely learns about the student over time, rather than resetting with each session. The responsible question asks whether oversight is calibrated to context, because working with twelve-year-olds is not the same as working with adults. Across five case studies, the picture that emerges is that the field is fairly good at the pedagogical layer, weak at persistent adaptivity, and pragmatic about responsible design. That order feels about right to me, and it suggests where the harder work still sits.
The trouble with detection algorithms
There are far too many examples of detection algorithms producing sad outcomes for students. A false accusation carries a real human cost, and these tools are simply not accurate enough to carry that weight on their own.
If you must use them, my honest advice is to first consider whether you need them at all. Where you decide you do, wrap them in the appropriate training and policies so that a flagged result is the start of a conversation, not a verdict. The output of a detector is a probability dressed up as a judgement, and it should never be treated as proof.
A language model from 1930
You might assume VLM stands for vision language model. Here it stands for vintage language model. Talkie is a 13B model trained only on English text published before 1931: no web, no modern corpus, no contamination, just books, newspapers, periodicals, patents and case law from a world that ends in 1930.
Ask it to imagine a school in 2026 and it describes handsome buildings in the suburbs of great towns, small classes, no Greek or Latin, pupils fed at public cost and free by Saturday noon. It is a charming picture, and also a useful one. A model is a mirror of its training data, and nothing more. When a system trained on a vanished world confidently narrates our future, the point lands quietly: what our models say about tomorrow is really a reflection of the material we fed them yesterday.
The case for accelerating toward AI in schools
A thoughtful series makes the case for accelerating toward AI in schools rather than tiptoeing around it. The argument runs like this: education is ahead on disruption but behind on integration, students already see through us, and the old assessment contract is broken.
None of it is bulletproof, and I would not pretend otherwise, but the provocation is useful. While we debate whether students should use a chatbot on their essays, the deeper question sits waiting: what are we going to talk about once information talks back? That is the conversation I would rather be having, and it is the one the tiptoeing tends to postpone.