Several Breakthroughs Away

Nish · July 16, 2026

Futures   Essay
⏱️ 17 min read

The closest I came to the future in 2021 was a waiting list I never joined. I was new to the field then, and a colleague told me about a model you could reach through an API that would, he said, answer almost any question you put to it. There was a queue for access; he may have signed up, but I told myself it would take a while and never got around to it. I did skim the announcement post, and I remember the thought that crossed my mind before I filed it away: once the masses find out about this, it is going to be huge. Then I lost interest and went looking at other things. That model was GPT-3, and what stays with me now is not the model but the shrug: I saw a corner of what was coming, labelled it correctly, and still felt nothing move. It would be nearly two years before I understood that the ground had shifted anyway, and that the thing which shifted first was not the machine but what a small number of people had let themselves believe was possible.

I build with these systems for a living now, so I am writing from the inside rather than the balcony, and I want to be careful in a way the topic makes hard. It is very easy right now to oversell what today’s models can do. They are genuinely good and getting better fast, and that is a different claim from the one that matters. The claim that matters is quieter and, to me, more unsettling: it no longer feels far-fetched that a handful of further breakthroughs would put us in a genuinely high-stakes situation, one our institutions are not built to absorb. I don’t think we are there. I think we can see it from here.

The day it got real

The dots joined themselves back up in the winter of 2022, when ChatGPT put a descendant of the model I had shrugged at into everyone’s hands. Suddenly the thing I had predicted would be huge was huge, and I had still managed to be surprised by it. Then the ground properly moved a few months later, in the spring of 2023, watching the GPT-4 demos. It was not any single trick. It was the accumulation, the sense of a system that could hold a problem in mind and work it, and the realisation that the curve I had been ignoring for two years was still bending upward and showed no sign of stopping. Since then the pattern has only sharpened: models that reason through problems instead of pattern-matching to them, agents that can carry a task across many steps, and, most recently, the first embodied systems that can act in the physical world on their own, still clumsy, still narrow, but no longer science fiction. Self-driving cars that actually drive. Robots that fold laundry from having watched people fold laundry.

Here is the thing I keep coming back to, though. The most important event of the last few years was not a model release. It was a change of belief. Somewhere between 2020 and now, a critical mass of serious, technical, unsentimental people stopped treating human-level machine intelligence as a someday-maybe and started treating it as an engineering target with a plausible completion date. Demis Hassabis, who has spent his life on this, now writes that we are standing in the foothills of the singularity and that we have found a way to make sand think.1 Whatever you make of the timeline, notice what that belief does. It moves capital. It moves talent. It moves the ambition of every lab in the field. Belief is not a passive forecast of the future; when enough capable people hold it, it becomes one of the forces that builds the future. The perception changed, and the perception is itself an accelerant.

That is the part I find genuinely new. Not that the machines got good, though they did, but that the ceiling in people’s heads came off.

The thing with a body

It helps to be precise about the destination, because “AGI” has become a word that means whatever the reader fears or hopes. A chatbot that can pass any exam is remarkable, and it is not the thing that reorders an economy. The thing that reorders an economy is intelligence you can give a body and then copy.

Picture the target plainly: a system that is embodied, so it can act in the physical world; that learns continually from its own experience rather than being frozen at the moment training stopped; that can act autonomously across long horizons without a human holding its hand; and, crucially, that can be replicated. Once you have trained one competent worker of this kind, you do not hire and train the next one over months. You copy it. The marginal cost of another skilled agent collapses toward the cost of the hardware it runs on. That is the moment labour, physical and cognitive alike, stops being the scarce input it has been for all of human history. Everything downstream of that moment is a different world.

I do not think we are a single breakthrough from that system. I think we are several, spread across different disciplines that would need to converge: robotics robust enough to handle the mess of the real world, learning that continues on the job without forgetting what it knew, models that generalise from far less data than they now need, and autonomy that stays coherent over hours rather than seconds. I traced why these are genuinely hard, and how far the field actually is, in a separate post on how robots learn; the short version is that the best robot-training datasets have only recently reached the lived experience of a single five-year-old, and the honest uncertainty among serious labs is measured not in years but in which decade. That is my real position, and it is deliberately more conservative than the loudest voices. Hassabis expects something like AGI within a few short years. Leopold Aschenbrenner has argued, on the strength of the trendlines, for something close to superintelligence by the end of this decade.2 I think they may well be right about the mind and early about the body, and it is the body, the replicable physical worker, that carries most of the consequences I care about. My honest estimate is wide: plausibly fifteen to thirty years for the convergence that matters, and I hold it loosely, because everyone who has offered a confident date on this technology has so far been humbled by it.

But fifteen to thirty years is not comfort. It is roughly one career. It is well within the lifetime of most people reading this. The right response to “not imminent” is not to relax; it is to notice that we have a little time, and to ask what we would do with it if we were serious.

One staircase

So let me lay out the shape of the thing, because I have come to think the good future and the bad future are not two different roads. They are the same staircase, walked twice.

The capabilities arrive in a rough order, easiest economic unlock first. First we amplify: intelligence makes the people who have it dramatically more productive. Then we cure: the same intelligence, pointed at biology and medicine, starts solving problems that were previously out of reach. Then, furthest away and hardest, we power the world: we loosen the physical constraints, energy and materials, that cap what civilisation can build at all. Each step is the same underlying technology maturing one more notch. And each step can land as a gift or as a shock, depending on one variable that has nothing to do with the technology itself.

A schematic chart titled 'One staircase, two ways to walk it' with the subtitle 'The height is roughly given. The gap is a choice.' An ember staircase labelled capability rises in three steps across an x-axis marked with three unlocks in order: Amplify, Cure, Power the world. A teal line labelled readiness rises beneath it, almost catching up along each flat step and then falling behind again when the next step jumps; an annotation with an arrow at the first jump reads 'a breakthrough lands, the gap reopens.' The wedge between the two lines, labelled 'the readiness gap' with a double-headed arrow, is filled in translucent red that fades where the gap is small and deepens where it widens toward the top-right. A dashed teal line labelled 'if we keep pace' hugs just under the staircase. A note at the thin early wedge reads 'gap stays small: absorbed as progress'; a note at the wide late wedge reads 'gap yawns open: arrives as a shock.'
The essay's engine, schematic. Capability (ember) rises in steps as breakthroughs land; readiness (teal) nearly catches up along each step, then the next unlock reopens the gap wider than before. The red wedge is the readiness gap, drawn darker where it is larger: it fades wherever readiness closes in and flares each time a breakthrough lands. The dashed line is the pace we could choose to keep. Which of the two walks we get is decided in the wedge, not on the staircase.

That variable is readiness: how prepared our economy, our institutions, and our norms are for each new capability at the moment it lands. When readiness keeps pace, an unlock is absorbed and feels like progress. When readiness lags, the same unlock arrives as a rupture. The technology sets the height of the staircase. We set the size of the gap.

The good climb

Walk the staircase upward with readiness keeping pace, and the future it produces is genuinely worth wanting. This is the part I refuse to be too cool to say out loud.

The first step is already under our feet. Intelligence is amplifying the people who wield it, most visibly in my own corner of the world, software. The descendants of the model I shrugged at in 2021 now write a large share of my code, and my working day has quietly reorganised itself around deciding what to trust rather than typing; Andrej Karpathy described flipping in a matter of weeks from mostly writing code to mostly reviewing it, and I watched the same flip happen at my own desk.3 I wrote about what that shift demands of us in a post on working with coding agents. Extend the pattern outward and the good version is not mass redundancy but mass leverage: a scientist, a doctor, a teacher, a founder, each with a tireless expert collaborator, human judgement multiplied rather than retired.

The second step is close enough to see clearly. Point this capability at biology and the payoff is not incremental. AlphaFold predicted the three-dimensional structure of nearly every protein known to science, over two hundred million of them, and released them to the world; problems that would have consumed careers were handed to researchers for free.4 That is a preview, not a peak. Faster drug discovery, molecules designed rather than stumbled upon, cures for diseases that have shadowed us for as long as we have kept records: this is the nearest of the truly civilisational wins, and it is nearer than the one after it. And when it lands it will not be felt as a statistic. It will be felt one waiting room at a time, a diagnosis that used to be a sentence handed back instead as a treatment plan.

The third step is the furthest, and I want to be honest that it is the least certain. This is the dream of abundance: intelligence turned on energy and materials until, as Hassabis puts it, resources are no longer the limiting factor for human progress.5 I believe it is possible. I also think it sits behind the others, because it depends on breakthroughs in physics, materials science, and engineering that intelligence can accelerate but not conjure. You cannot think your way past the laws of thermodynamics; you have to build. Abundance is the top of the staircase precisely because it is where thinking finally has to become making.

Notice the ordering, because it matters for everything that follows. The good future does not arrive all at once. It arrives in waves, and the early waves are already lapping at the shore.

The same climb, unready

Now walk the identical staircase with readiness lagging, and watch each gift curdle into its shock. Nothing about the technology changed. Only our preparation did.

I keep reaching for the same analogy, imperfect but instructive: the early months of the COVID pandemic. The virus was not, by historical standards, the worst thing that could have happened. What made it a catastrophe in so many places was that it arrived faster than institutions could bend, and the gap between the speed of the threat and the speed of the response was where the damage lived. That gap is the thing I fear here, not a machine with red eyes.

Take the amplification step, the one already under us, and remove readiness. The same tools that give a scientist leverage quietly dissolve the entry-level rungs of whole professions, the junior roles where people used to learn their trade. Picture the graduate who did everything right, arriving to find the bottom rung gone. Dario Amodei has warned that AI could eliminate something like half of entry-level white-collar jobs within a few years; I think that specific figure is contestable and I hope it is wrong, but the shape of the worry is right.6 The danger is not that the work becomes impossible. It is that the transition is fast and broad, hitting people not by profession but by whatever their machine can now do, faster than retraining or policy or a new social contract can catch them. An amplification we are ready for is leverage. The same amplification we are not ready for is a labour shock with no precedent for its speed.

Take the autonomy step. A system that is embodied, always on, and learning continuously from its own experience is exactly the system that is hardest to keep meaningful watch over, because it is changing while it runs and acting while you sleep. Hassabis names this directly: as systems become more agentic and begin to improve themselves, we will need robust safeguards to keep them under control, and some of the problems will only become clear as we get closer.7 This is not the cartoon of a malevolent AI. It is the sober engineering worry that we build something capable and autonomous before we build the ability to reliably understand and correct it, and that the gap between the two is where control slips.

And take the whole staircase at once and ask who owns it. If the decisive input to the economy becomes intelligence you can copy onto hardware, then whoever controls the compute and the robots controls a share of power that no previous concentration of wealth could match, because it no longer needs anyone else’s labour to sustain itself. A gap in readiness here is not economic. It is the possibility of an asymmetry you cannot vote your way out of.

I am not predicting any one of these. The bad future is not a single event; it is a spectrum, and which points on it we actually hit depends on which breakthroughs land, in what order, and whether our readiness had closed the gap by the time they did. That is the whole point of the staircase. The steps are roughly the same either way. The wedge between capability and readiness is what turns a step into a stumble.

A fork, not a forecast

So I find myself unable to write either of the two essays that are easy to write. I cannot write the one where technology saves us, because it plainly can hurt us. I cannot write the one where it dooms us, because the outcome is not the technology’s to decide. What I am left with is less satisfying and, I think, more true: this is a fork, not a forecast.

The case rests on one observation: the capability staircase is, at this point, coming more or less regardless of any single actor, pulled forward by the belief and the competition I described at the start. What is genuinely undetermined is us: whether readiness rises to meet each step. Hassabis makes the same point from inside the most optimistic camp, and I think it is the wisest thing in his essay. Progress on the frontier, he writes, is outpacing our understanding of it; nobody knows for sure what happens from here, and even the experts disagree. When uncertainty is that wide and the stakes are that high, the honest posture is neither hope nor doom. It is preparation. He calls it cautious optimism, and reaches for concrete machinery to close the gap, a standards body to test frontier models before they ship.8

I sit a little apart from all of the voices I have named, and it is worth saying where. I am slower than Hassabis and Aschenbrenner on the part that matters most to me, the replicable body, and I think the timelines pointed at the mind alone understate how much hard physical engineering still stands between here and the world-changing version. I am more hopeful than the darkest forecasts, because I do not believe the ending is written. And I am less amused than the deflationary skeptics, the ones who have watched every wave of hype crest and recede and conclude this one will too, because this time I cannot make the reasons it is overblown add up.9 What unites my disagreements is a single conviction: the interesting question is not how fast the staircase rises. It is whether we close the gap while we still have a little time to.

The fork, handed to you

Five years ago a colleague pointed me at the future, and I could not even be bothered to join the waiting list. I had read enough to call it huge, and I still felt nothing move. I know better now than to trust that stillness, because the ground was already shifting underneath it. That is the honest reason I cannot tell you which future we get. I have watched myself be wrong about the pace of this once already, in the safe direction, and the lesson I took is not a new prediction but a loss of faith in confident ones.

What I can tell you is that the staircase is real, that it is being built a rung at a time by people who now believe it can be finished, and that the distance between the good version and the bad version is not a distance in the technology at all. It is the gap in the picture above: the space between what we can suddenly do and what we are actually ready for. That gap is not fixed. It is the one part of this whole story that is unambiguously ours to change, and the window in which changing it is cheap is open now and will not stay open.

I am not going to close by telling you what to do; I distrust essays that pretend the choice is theirs to hand down. The future here is genuinely not yet written.10 So take the fork, and the one thing I have become sure of. We are several breakthroughs away from a world that will demand far more readiness than we have. The breakthroughs no longer look far-fetched. And the only question still open is the one that was always ours. Which way do we walk it?

Sources & further reading

  1. Demis Hassabis, “A Framework for Frontier AI and the Dawning of a New Age” (2026). His full framing is that AGI is “more akin to the discovery of electricity or fire” than to the internet, with an impact “perhaps 10x of the Industrial Revolution at 10x the speed.” ↩

  2. Leopold Aschenbrenner, “Situational Awareness: The Decade Ahead” (2024), argues from compute and algorithmic trendlines (“counting the OOMs”) toward AGI around 2027 and superintelligence shortly after. ↩

  3. Recounted in my post on working with coding agents: Karpathy described going from roughly 80% hand-written to 80% agent-written code in a matter of weeks, calling it the biggest change to his workflow in two decades. ↩

  4. DeepMind’s AlphaFold predicted structures for nearly all cataloged proteins, over 200 million, and released them openly via the AlphaFold Protein Structure Database. The 2024 Nobel Prize in Chemistry recognised the work. ↩

  5. Hassabis: “We could even reach a point where resources are no longer the limiting factor for human progress, leading to an amazing new era of abundance.” I treat this as the furthest and least certain step precisely because it is gated on physical, not just cognitive, breakthroughs. ↩

  6. Dario Amodei, “The Adolescence of Technology”, warns of disruption to a large share of entry-level white-collar work on a timescale of a few years. I quote it as his claim, not mine; the specific fraction is contested, but the shape of the concern does not depend on the exact number. ↩

  7. Hassabis writes that we will need “robust safeguards to maintain control of increasingly agentic, recursively self-improving systems,” and that some issues “will only become clearer over time.” This is the mainstream-lab version of the control worry, distinct from more speculative doom scenarios. ↩

  8. His proposal is a FINRA-style public-private standards body that tests frontier models before release, initially voluntarily and later as a condition of deployment, designed to be ratcheted up if the situation demands. Whatever the specifics, the instinct is the right one: build the readiness deliberately rather than hope it accretes. ↩

  9. George Hotz’s blog is the sharpest version of the deflationary read, and a useful discipline against hype. I take the skepticism seriously and still cannot make its conclusion hold for the embodied, replicable version of the technology. ↩

  10. The phrasing is deliberately borrowed from the close of Hassabis’s essay, “the future is not yet written,” because it is the one point on which the optimist and I agree without reservation. ↩

Citation Information

If you find this content useful, please cite this work as:

Bhana, Nish. "Several Breakthroughs Away". Nish Blog (July 2026). https://www.nishbhana.com/Several-Breakthroughs-Away/

Or use the BibTeX citation:

@article{bhana2026severalbreakthroughs,
  title   = {Several Breakthroughs Away},
  author  = {Bhana, Nish},
  journal = {nishbhana.com},
  year    = {2026},
  month   = {July},
  url     = {https://www.nishbhana.com/Several-Breakthroughs-Away/}
}

x.com, Facebook