A mix this week of headline capability claims that deserve a closer read, some genuinely novel approaches to how computation might work, and a couple of reminders that the questions raised by all this reach well beyond the technology itself. I want to slow down on a few of these rather than skim them, because the interesting detail is usually a layer below the summary.
Reading the GPT-5.6 system card carefully
The GPT-5.6 system card rates every version of the model High capability in biology, and left there it sounds alarming. But the evaluations underneath tell a more measured story. The rating rests on three of four biology evals clearing threshold, and the card itself notes that two of those may be saturated, with differences from the previous model possibly down to noise.
The one that stayed with me is the open-ended test, where the model has to troubleshoot a flawed lab protocol in its own words rather than pick from multiple choice. There it scored 43.5% against a 54% expert threshold. Human experts still win that one.
That gap matters. Multiple-choice benchmarks saturate and then stop telling us anything useful about frontier progress. If we want an honest picture of capability, we need more open-ended evaluations like this one.
The GPT-5.6 preview
OpenAI has previewed the GPT-5.6 series: three models named Sol, the flagship; Terra, a balanced option pitched as competitive with GPT-5.5 at half the cost; and Luna, the fastest and most affordable.
The performance claims are notable. Sol Ultra leads Terminal-Bench 2.1 at 91.9%, and Sol is said to shift the performance-efficiency frontier on cybersecurity tasks, reaching parity with Mythos Preview on ExploitBench while using roughly a third of the output tokens. OpenAI states Sol does not cross the Cyber Critical threshold under its Preparedness Framework.
What I found more telling was the safety framing. The models launch with what OpenAI calls its most robust safety stack yet: model-level refusal training, real-time cyber and biology misuse classifiers, and more than 700,000 A100-equivalent GPU hours of automated red-teaming. OpenAI also says the government access process used here should not become the long-term default, which is a candid thing to put in writing.
An image generator that runs on oscillators
A startup called Unconventional AI has released Un-0, an image generator that does not run on a neural network in the usual sense. Instead it runs on a simulated system of coupled oscillators, and the patterns they settle into become the image.
The mental picture that helped me was thousands of metronomes on a shared table, each nudging the others into sync or opposition until the whole system finds a stable arrangement. The work is simulated and early, and I would not read too much into a single release. But if the approach scales, it hints at something interesting: a computer that runs on oscillators rather than GPUs. You can read their introduction for the detail.
Reading a scroll Vesuvius turned to carbon
Two thousand years ago the eruption of Vesuvius buried Herculaneum and burnt a scroll to carbon. A team has now used an AI trained to spot ink on hidden layers, working from high-resolution X-ray scans, to virtually unwrap it, recovering 20 columns of text across more than a metre of charred papyrus.
The recovered text is a stoic treatise on impulse and practical wisdom, possibly by Chrysippus. I find this one quietly moving. So much of the AI conversation is about capability and risk that it is easy to forget the same tools can hand us back something thought lost for two millennia. The full account is in the Guardian.
A rare Five Eyes warning
A rare joint statement from the Five Eyes intelligence agencies warned that frontier AI models capable of disrupting governments and businesses are only months away. Their argument is that AI will accelerate the speed, scale and sophistication of cyber threats and lower the barriers for bad actors.
The line I keep returning to is that cyber resilience can no longer be treated as a purely technical issue. It is a core business risk and a leadership responsibility. That reframing sits a little outside my usual focus on education, but it applies just as squarely to schools and universities as to any other organisation holding sensitive data. The statement is reported here.
The case against banning AI in primary schools
Finally, a question closer to home. There is a strong instinct to simply ban AI for younger students, and I understand it. But a ban achieves very little for a Year 5 or 6 student. They will use AI at home, alone, on tools built for adults, with no teacher visibility. The ban removes the school's ability to shape how the technology is used, not the risk itself.
I would argue Years 5 and 6 are precisely the right window for scaffolded exposure. This is when children form their first mental model of what AI is and how far to trust it, and that model is better built with a teacher in the room.
It is also worth being precise about what we are comparing. A bounded, teacher-designed, logged and curriculum-aligned agent is a genuinely different category of risk from a child alone with a general-purpose chatbot at nine at night. Pretending those two are the same does not keep anyone safer.