This week I have been thinking about the gap between how we actually communicate and how we are told to talk to machines. That runs from a new take on pointing and shared context, through the stubborn advantage of bigger models, to a thoughtful essay on the purpose of education. It ends with two questions worth sitting with: what writing trains in us that a tool can shortcut, and whether AI toys belong in children's hands at all.
The power of "this" and "that"
In everyday life we rarely speak in long, careful paragraphs. We say "fix this", "move that here", "what does this mean", and we lean on gesture and shared context to carry the rest. Most of our communication is shorthand, and it works because the other person can see what we see.
Almost every AI interface still asks us to do the opposite: to spell out in words what a glance and a pointed finger would settle in a second. DeepMind's recent work on combining context, pointing and speech is a glimpse of something better, an interface that lets people make genuinely complex requests in natural shorthand, with none of the fiddly prompting we have come to accept as normal. It matters because the barrier to using these tools well is rarely the underlying model. It is the effort of translating a simple intention into the kind of explicit instruction a machine can act on. Close that gap and the technology finally starts to meet people where they are. You can read the DeepMind post here.
Why bigger models keep winning everywhere
One property of large language models we quietly take for granted is that newer, bigger models tend to be better at almost everything. The labs pour their effort into economically valuable areas like coding, and yet the same larger models also turn out to be better at negotiation, at strategy, even at poetry. Capability seems to arrive as a broad tide rather than a set of narrow wins.
As Ethan Mollick has noted, this complicates the understandable push towards smaller, local models. Smaller models are cheaper, more private and easier to run, and for well-defined, repetitive tasks they are often more than enough. But if you want a good answer to an unexpected problem, frontier models still generally outperform their smaller counterparts across a wide range of tasks. That tension is worth naming honestly in education, where we are rightly drawn to the control and cost of running things ourselves. The pragmatic answer is probably not one or the other, but knowing which questions are routine enough for a small model and which deserve the larger one.
Renegotiating the education social contract
Kim Smith's essay, "Renegotiating the Education Social Contract for the Age of AI", is a must-read, and two passages have stayed with me. The first is her argument that the choice between the individual learner's private good and the collective civic good is a false one. We have long treated it as an unavoidable trade-off, when in truth it is a design failure we now have the capacity to solve.
The second is her framing of agency. Learners and families, she argues, need the power to construct learning pathways that genuinely meet their needs, cultivating not only self-directed agency but morally directed agency: the capacity and the desire to contribute to something beyond oneself. That second half is the part we tend to forget. It is not enough to help a student steer their own learning; the point is to help them want to use that capability in service of others. Parts of the piece are uniquely American, but I suspect you will recognise Australia in it too. It is worth reading in full, here.
What writing trains that AI can shortcut
A fiction-writing lecturer recently read two student stories and told the class he knew AI had written them, without any detection software. The tells were simple: the prose was too polished, the arcs too tidy, every metaphor a well-turned pastiche with no context behind it. One student then admitted what had happened, and it had not started as cheating. It started with fear of looking stupid: a grammar check, then line edits, then structural edits, and finally a full rewrite.
What strikes me about that slide is how reasonable each step felt. But it points to something we underrate. Writing is not just the production of sentences. It is the training of endurance by way of sustained attention, the slow work of holding an idea in mind long enough to shape it. That is precisely the capacity a tool which produces the sentences for you quietly removes. The finished text can look better while the thinking that writing was meant to develop never happens. In education that is the trade we most need to watch, because it is invisible in the output.
The case for regulating AI toys
The AI toy market would benefit from regulation and frequent independent testing, and parents should expect that safer products will cost considerably more. Early testing from groups like PIRG has already surfaced the core problem: guardrails that hold up in short, scripted interactions tend to break down in the longer, open-ended play sessions children actually have.
If it were up to me, I would want three things:
- mandatory safety evaluation against a published standard;
- continuous re-testing, because the toy tested in March is not the toy on the shelf in September;
- clear age gating with transparent disclosure of what the toy is and how it behaves.
There is a deeper uncertainty underneath the safety question. We do not yet know whether young children benefit from chatbot companions at all, and there is good reason to think they attach to responsive agents in ways we should understand properly before scaling them into millions of homes. So on this one I think it is worth asking the harder question: whether unsupervised AI companions for young children should exist in the first place.