The Five Frontiers of Frontier AI
I’ve never liked the word "futurist" and I don't think of myself as one. So being named the UK’s number 1 futurist keynote speaker recently was both gratifying and slightly uncomfortable. Rather, what I seem to have (as a result of long years of experience in the field), is the ability to articulate possible trajectories and propose practical responses.
I live, as it happens, in the Islington house that was once home to John Nelson Darby, the nineteenth-century Bible teacher considered to be the father of modern dispensationalism and futurism. Unlike mine, Darby's futurism was theological - more prophecy than projection. I'm not religious, but I find it hard to sit in my home office thinking about the future without noticing that, under the same roof and the better part of two centuries earlier, John was doing the same thing.
The difference between Darby and me is important. His future was presented as a fixed truth - it needed no evidence and could not be wrong. Mine is a forecast, and a forecast is only worth reading if it tells you what would prove it wrong. This piece does two things. It sets out the Frontiers I think frontier AI will pass through over the coming years, and, for each, it names the signals that would tell us whether we are on the trajectory or have fallen off it.
Ask most people what the next frontier in AI looks like and they will describe a bigger model: more parameters, more data, more GPUs. Bigger is not always better, of course. So what’s the alternative to size? For me it’s the ability to learn.
My preferred definition of intelligence is Sternberg and Salter’s ‘goal-directed adaptive behaviour’ - with a strong focus on the word adaptive. A system that cannot change in response to its own experience is not intelligent in any interesting sense, however dazzling its output. On that definition, today's frontier models are extraordinary but not truly intelligent. They are trained, frozen, and shipped. Everything that looks like learning after that point is context management and adaptation that happens in a single conversation.
So the frontier of frontier AI is not scale, it is the machinery of adaptation. How cheaply a system can think, how well it can deliberate, whether it can learn in situ, whether it can remember, and eventually whether it can improve its own way of learning. I think those five capabilities will arrive roughly in that order, each making the next affordable.
The two trajectories
Two trajectories run through everything that follows.
The first is cognitive depth: how generally a machine can run the scientific method. I have described this before as a ladder from intoxicated graduate to professor.*
The second is breadth and embodiment: a machine cognitively and physically capable of what an average person can do.
These two trajectories pass through the development of the five capabilities - or the five Frontiers as I will call them.
Frontier 1: Cheap cognition
Everything else depends on this. Deliberation among many agents, learning during deployment, persistent memory - each multiplies the computation per unit of useful work. If thinking stays expensive, the later Frontiers won’t happen.
The best benchmark we have for efficient cognition is our own biology. A human brain runs on roughly twenty watts and acquired language from tens of millions of words, not trillions. Progress towards closing the gap between current LLMs and cheaper, more efficient architectures is already becoming visible in three places:
Neuromorphic hardware: chips that compute with spikes and events rather than dense matrix multiplication. Analysts mark 2026 as a transition year, with BrainChip's second-generation Akida and Intel's Loihi 2 moving from research access into product integration, and carmakers including Mercedes-Benz exploring in-vehicle neuromorphic compute. The SpiNNaker2 chip, from Manchester and Dresden, delivers a tenfold increase in neural simulation capacity per watt over its predecessor. At Conscium, we’re investing in neuromorphics and I’m consistently blown away by the potential of this technology.
Organoid intelligence: computing with living neurons. Cortical Labs announced the CL1 in March 2025 at $35,000 and shipped it that summer as the first code-deployable biological computer, with buyers able to deploy code directly to the neurons. FinalSpark is building a similar platform with 160,000 neurons across sixteen organoids, accessible over the internet and marketed as a low-energy substrate for AI. Living tissue is slow, fragile and ethically unresolved. But biological synapses adjust their weights metabolically, supporting lifelong adaptation without recalculating gradients, which is something neuromorphic cores must simulate in firmware. Hold that thought for Frontier three.
Whole-brain emulation: which is now moving from thought experiment to engineering. In 2024 a computational model of the entire fruit-fly brain, built from the FlyWire connectome, predicted motor behaviour with 95 per cent accuracy but had no body; in March 2026 Eon Systems closed the loop by coupling that emulated brain to a physics-simulated fly, producing multiple distinct behaviours driven by the circuit dynamics themselves. The virtual fly walks, grooms and feeds, and those behaviours emerged from the wiring rather than from gradient descent or reward functions; the company is targeting a mouse brain within two years, with human scale decades away.
Meanwhile the mainstream labs attack efficiency from the other end. Over a twelve-month period, and while delivering higher-quality responses, Google cut the median energy consumption per Gemini text prompt by a factor of 33 and the carbon footprint by 44. Two curves, biological substrates getting more capable and digital substrates getting cheaper, are converging on the same target.
Signals: energy per unit of capability (not per token) falling by an order of magnitude every couple of years; a model trained on a human-scale corpus matching today's frontier on reasoning; a complete mouse connectome within three years; and a shared software stack for neuromorphic chips.
Frontier 2: Deliberation
Cheap cognition makes it affordable to run many instances of a mind at once, and the second Frontier is what happens when you do. Ecosystems of agents that propose, critique, replicate and rank produce results no single agent could achieve. In human scholarly work, this is how the graduate becomes a master's student - not just by knowing more, but by checking its own work through structured critique and disagreement.
The frontier labs already build this way. DeepMind's AI Co-Scientist uses a multi-agent society to propose and collaboratively critique hypotheses in biomedical research. Edison Scientific's Kosmos combines a multi-agent architecture with a structured world model to run autonomous, data-driven discovery across domains. And in March this year the field got its first serious peer-reviewed marker: Nature published Sakana AI's The AI Scientist, a system that creates research ideas, writes code, runs experiments, analyses data, writes the manuscript and performs its own peer review, and whose output passed the first round of review at a workshop of a top-tier machine-learning conference with a 70 per cent acceptance rate. The authors were admirably candid - none of the three manuscripts met the standard of the main conference, and one of them likened the system to a first-year PhD student, with occasionally surprising ideas outnumbered by mediocre ones.
The bigger change is that the labs use these collectives to improve the substrate their next models train on. DeepMind's AlphaEvolve combines Gemini models with automated evaluators and evolutionary search; for example one of its optimisations made a matrix-multiplication kernel 23 per cent faster, cutting Gemini's training time by about 1 per cent.
Signals: a main-track paper at a top venue whose critical step was produced by an agent collective; agent teams that reliably outperform their strongest single member, which is a testable claim rather than a slogan; and enterprise workflows in which agents adjudicate one another's outputs before a human sees them.
Frontier 3: Learning in situ
Today’s models don’t learn. They were trained; the model was then locked. Give it your organisation's documents and it looks like it's learning, but the weights are untouched and the knowledge evaporates when the context does. Deployed agents are, as one recent paper puts it, like newly hired employees who never get on-the-job training, repeating the same mistakes in unfamiliar settings. The third Frontier is the model that changes because of what it does.
This is the hardest of the five and, I'd argue, the real bottleneck. Catastrophic forgetting, where learning the new thing erases the old, has resisted three decades of work. But the architectures are moving. Google's Titans introduces a learned long-term memory module that incorporates historical context at inference time, shifting some learning from offline weight updates into an online memory process. Its Nested Learning work goes further, with a sequence model that learns its own update algorithm and a continuum memory system, combined in an architecture called Hope that shows resilience to catastrophic forgetting. An ICLR 2026 oral presented In-Place Test-Time Training, which turns part of an existing model into fast weights that update during inference as a drop-in enhancement. These directions are converging, alongside Silver and Sutton's framing of an era of experience in which agents learn from a continuous, never-ending stream. And yet the production reality is sobering: frontier-scale weight updates remain offline in 2026 because every update needs evaluation, safety review and rollback, so what ships under the banner of continual learning is mostly fast adaptation through context, retrieval, and adapters on a cadence.
That last sentence is where this Frontier connects to my work at Conscium. A model that changes in deployment is a model whose verification expires the moment it learns something. The tooling to certify a system that rewires itself does not yet exist at scale, and I'd argue that, not the algorithms, is the gating factor. Learning in situ is what turns a master's student into a PhD: a generalist becoming an expert through practice. It is also what makes the automation of expert white-collar work real rather than rhetorical, because expertise is precisely what you cannot download from a context window.
Signals: a deployed model measurably better at your organisation's specific work in month six than in month one, with no new release in between; published benchmarks for learning-during-use, as distinct from few-shot prompting; and a regulator or auditor publishing a standard for re-verifying a model that has changed itself.
Frontier 4: Persistence
I've added this Frontier because I don't think you can reach postdoc level without it. A postdoc isn't a faster PhD. A postdoc has a research programme: an accumulated, structured memory of what has been tried, what failed and what matters, carried across projects and years. Learning in situ gets you a specialist. Persistence gets you a scientist with a history.
Two lines of work matter here. The first is memory as architecture rather than retrieval: episodic and semantic stores that consolidate, the way sleep consolidates ours. The Kosmos and Hope work above are early versions. The second is world models, internal simulations of how the world behaves that an agent can plan against. Roughly $6 billion flowed into six or seven world-model companies in the first quarter of 2026 alone, because physical AI needs embodied data that does not exist online in usable form, and world models offer a way to generate it synthetically at scale. The same machinery that lets a robot rehearse a task lets a research agent rehearse a hypothesis.
Persistence has a governance consequence. A system with a history has a disposition. What it does when nobody has told it what to do becomes a property of its accumulated experience rather than its prompt, which is why I've been arguing for dispositional verification alongside functional and non-functional testing.
Signals: an agent running a single research programme over months, with each result building on its own prior results; longitudinal evaluations measuring whether an agent's judgement improves with tenure; and the first serious dispute over whether a long-running agent is still the system that was certified.
Frontier 5: Learning to learn
The professor doesn't apply a method, she invents it. The fifth Frontier is systems that improve their own learning strategies, and it closes the loop with the first, because the brain is the existence proof: a twenty-watt device that learned how to learn.
This Frontier has already started, in bounded form. The Darwin Gödel Machine rewrites its own code and validates each change empirically, and in doing so raised its own SWE-bench performance from 20 to 50 per cent. In February OpenAI reported that GPT-5.3-Codex was instrumental in creating itself, helping debug training, manage deployment and analyse evaluations; Anthropic says the majority of its code is now written by Claude Code; and in all these cases humans still direct and verify the work. OpenAI's stated plan is intern-level AI research agents by September 2026 and fully functional ones by 2028. September 2026 is now. Hold them to it. Google's own 2026 research agenda lists self-improving AI and automated AI research as a target area, describing a goal rather than a completed system, and there is no universally accepted test that would let a lab declare it has reached recursive self-improvement.
None of this is self-improvement in the science-fiction sense. But the direction is unmistakable, and it is where neuromorphics, organoids and emulation stop being a side-quest about energy and become central. On-chip plasticity, metabolically self-adjusting synapses and connectome-derived dynamics are all attempts to build a substrate whose learning rule is itself learnable. That is the professor level, and it is being approached from silicon and from biology at once.
Signals: an AI-discovered training method or architecture adopted in a frontier run and credited as such; a main-track paper with no human author; the interval between capability generations shrinking rather than lengthening; and labs publishing self-improvement results together with the evaluators that constrained them.
The other trajectory: the body
Everything above concerns unembodied capability. The second trajectory, a machine that can do physically what an average person does, runs through the same five Frontiers, but on different data.
In July DeepMind released Gemini Robotics ER 2 to developers, with real-time video understanding, task-progress tracking and multi-robot collaboration. At CES it partnered with Boston Dynamics to put those models on the electric Atlas, with testing planned at Hyundai plants; 1X opened its NEO factory in April, says customer shipments start before the end of the year, and booked its first 10,000 units of planned capacity in five days. And yet 2026 is a validation year rather than proof of mainstream adoption, and the strongest operating evidence is still narrow industrial and logistics work with controlled tasks. Policies that hit 95 per cent success in the lab drop to around 60 per cent in the field, and most humanoids run for 90 to 120 minutes on a charge against shifts of eight to twenty hours. Bessemer calls it robotics' GPT-2.5 moment - real capabilities, emerging scaling laws, and a wide gap to the 99.9 per cent reliability production demands.
That gap is the in-situ learning problem in a body. Physical intelligence is like the unembodied kind, but with a bigger data bill.
What we should see
If this trajectory is right, the evidence will arrive as a pattern rather than a single announcement. Some of the pattern is already visible.
On 20 May 2026 OpenAI announced that an internal model had produced a counterexample to Erdős's 1946 unit-distance conjecture, the first historically significant mathematical proof to come from an AI. The result was not definitive, and human mathematicians improved on it within weeks, but it imported ideas from a distant branch of mathematics and related techniques were used on other problems within days. OpenAI's August claim that an unreleased model made ten further advances deserves the posture "interesting and unverified" until the community has checked it. DeepMind's Aletheia has produced a publication-grade paper with no human intervention and resolved four open Erdős questions. Jack Clark, at Oxford in May, predicted that AI will work with humans to make a Nobel-worthy scientific discovery within twelve months.
So watch for: results that the relevant community accepts as AI-led rather than AI-assisted; energy per unit of capability falling faster than capability rises; models that improve between releases, not just at them; agents with tenure; and verification regimes visibly struggling to keep pace, which may prove the most reliable signal of all.
Why this is good news
Each Frontier is also an opportunity. Cheap cognition means intelligence at the edge, in devices and processes where a data-centre round trip was never viable. Deliberation means fewer confident errors. Learning in situ means the systems that know your organisation best will be the ones that have worked inside it longest, a moat no one can buy. Persistence means institutional memory that doesn't retire. And learning to learn means a scientific method that scales, pointed at problems we have been too slow to solve.
The frontier of frontier AI is not a larger model. It is a learner: cheaper, more collaborative, more adaptive, more persistent and eventually more inventive than the ones we have. That is what goal-directed adaptive behaviour was always going to require, and the interesting work now is building the assurance to trust it as it climbs.
Which brings me back to the house. Darby spent a lifetime describing a future that could not be tested, and built one of the most durable belief systems of the modern era on it. I would rather build something more fragile - a trajectory with forecasts, with evidence, and with the possibility of being wrong. If the signals above don't appear, we still learn something. That’s how science works, and that’s what I think being a futurist should be - not about prophecy, but instead about using expertise to make falsifiable predictions.
*Let me clarify what I mean by this. It is not a claim about peak capability. Frontier models already outperform most professors at some tasks and underperform a hungover graduate at others. The capability frontier is jagged, as Ethan Mollick and colleagues put it, and it expands unevenly, which is rather the point of the title. The ladder measures generality of method. The intoxicated graduate knows a great deal and applies it unreliably. The master's student applies a known method competently. The PhD specialises: takes a general method and, through sustained practice, becomes expert in one narrow domain. The postdoc accumulates a research programme over years. The professor invents method rather than inheriting it. What we see today are spikes of postdoc-level, occasionally professor-level, performance on narrow problems. What the trajectory predicts is those spikes widening into general ability, and the five Frontiers below are how I think that happens.
