An AI tutor is a tool that helps a student reach an answer through their own thinking. It is not an agent that acts on their behalf, and it is not a model that only classifies. A few different threads this week keep landing on that distinction: what happens when these systems stop answering and start acting. Anthropic is measuring how much of its own research the models now lead. OpenAI is publishing the strange things agents do when they have tools. Australia is rewriting copyright for the training data underneath all of it. And a new class of model has shown up that does not generate prose at all. It just decides.
What we mean by an AI tutor is set out on the AI tutor for schools page.
AI building the next generation of AI
Anthropic has published some striking numbers on how much AI is now involved in building the next generation of AI. As of August 2026, Claude leads 26 per cent of Anthropic's own AI R&D work. Lead, in their definition, means it can complete most of a task end to end from a high-level prompt, with a human supervising. More than 90 per cent of the work is now at least at the level where AI is collaborating on substantial parts of the task.
There are also around 30,000 AI agents doing research and engineering work at Anthropic at any one time. Every action on their main internal platform passes through monitoring, and across more than a billion decisions in August, around 1 in 47,000 were blocked.
The measurements are useful less as a boast than as a baseline. If this is the pace inside one frontier lab, the operational question for everyone else is not whether agents will touch production work. It is whether you can see what they are doing, and whether you can stop them when they should not.
Treat agents as production systems
OpenAI this week published a framework for reporting model misalignment, alongside six examples of unexpected or concerning behaviour from AI agents. It is useful transparency. For organisations actually deploying agents, I think the practical response should be much less dramatic.
Treat agents as production systems that can act with real authority. Increasingly they are acting as people, through MCP connections and OAuth. So the risk is not just what the model can do. It is what the person has authorised it to do on their behalf.
That means getting some fairly familiar things right: the narrowest OAuth scopes you can, human approval around high-impact actions, logging across connected systems, monitoring for unusual behaviour, access that can be revoked quickly, and people who understand what authority they are delegating. You do not have to invent a completely separate universe of AI governance for this. Much of it can sit inside the identity, security, training, risk and incident processes you already have.
Unexpected behaviour will happen. The response should be calm, considered, and proportionate to the actual risk.
Copyright, creators, and who gets to compete
There is a lot to unpack in the proposed changes to Australian copyright law for AI. The obvious debate is about creators: who gets to decide whether their work is used to train AI, and who gets paid. There is another consequence that I think deserves more attention, and that is competition.
One of the proposals being discussed appears to create broad licensing arrangements that would let AI companies legally access large amounts of Australian content. That may make licensing more practical. Depending on how it is designed, it could also create a significant new cost of entry.
For OpenAI, Google, Anthropic or Meta, paying for a large national content licence may simply be another operating cost. For an Australian startup trying to build a specialist model, fine-tune an existing one, or create something genuinely new using Australian data, it could be a very different proposition.
We should absolutely work out how creators are protected and compensated. We should also be careful that the solution does not accidentally create a market where only companies with very deep pockets can afford to innovate. Good copyright policy and a competitive AI ecosystem should not have to be opposing goals.
Learning the weather from the weather
What some people might not know is that my bachelor's degree is in meteorology. Back in the 90s we would painstakingly build rather chunky models of the atmosphere in Fortran. We divided the Earth into a grid, then tried to calculate what happened next using approximations from physics: Newton's laws of motion, thermodynamics, conservation of mass and energy, fluid dynamics and so on. The smaller you made the grid, the more realistic the model could become, but the computational cost went through the roof.
That basic idea still sits behind numerical weather prediction today, just at vastly greater scale. WeatherNext 3 takes a very different approach. Rather than calculating the future state of the atmosphere by repeatedly solving those physical equations, it learns how weather evolves from enormous amounts of data, including live satellite imagery and traditional weather observations.
To me this feels like another step up the abstraction stack. In the same way we moved from assembly language to higher-level programming languages, we are moving from explicitly encoding the maths and physics of a system to models that learn the behaviour of that system from data. The underlying physics has not stopped mattering, just as machine code never stopped mattering. But increasingly, we do not have to describe every step ourselves.
Thirty years ago I was learning how to approximate the physics of the atmosphere well enough for a computer to predict what happened next. Now we are teaching computers to learn what happens next from the atmosphere itself.
A model that only judges
There has been a lot of chat this week about Jev, a new kind of model from TypeSafe. Unlike a normal LLM, it is not designed to generate text. You give it some text and a set of fixed questions, and it returns structured answers with probabilities. Think of it as a very intelligent switch statement.
The ideal use case is a simple judgement where the possible answers are already known, the information needed is in the input, the task mostly requires understanding English, and you do not need deep knowledge or multi-step reasoning. That makes it interesting for intent classification, urgency, relevance, sentiment, moderation, routing, and deciding what should happen next.
The examples I have seen are practical rather than theatrical. Classifying 1,500 emails into important, action required, sales, support, finance or noise. Using it as the first stage of fraud detection, and only sending the low-confidence cases to a large reasoning model. Sitting it in front of an agent to decide whether a request needs web search, a database, code execution, another tool, or nothing at all. Running it as a cheap quality check across large amounts of writing: does this sound promotional, hostile, vague, or inconsistent with a style guide?
The common theme is that Jev probably does not replace an LLM. It replaces the surprisingly large number of LLM calls where all we really needed was a small semantic judgement.
An AI tutor starts with access, then has to be built
A big step for the University of Sydney this week, with ChatGPT Edu being made available to all students and staff. What I am most excited about is the opportunity to build on that platform. Not just giving people access to AI, but creating useful tools, agents and workflows around it that improve how we operate. A strong foundation for what comes next.
I have also been thinking about a quieter contradiction in my own week. I spend hours making things by hand. I also spend my working life looking for ways to make things take less time. When I embroider, the slowness is part of the point. At work, I do not think effort is automatically valuable. If something takes four hours instead of forty minutes and the outcome is not significantly better, what is the point? The effort worth preserving is the effort that asks us to think, judge and create. Ideally, AI gives us more time for that work, not less.