From the Technological Singularity to a God-Tier Civilisation — Fact or Fiction?

COSMOS AI LAB · RESEARCH & FUTURES STUDIES · SEPTEMBER 2026

There is a thought experiment I return to whenever I need to be reminded of the scale of things. Compress 3.5 billion years — from the earliest fossil traces of life to this afternoon — into a single 24-hour day. For almost that entire day, nothing happens but the silence of bacteria. The first stone tools appear at 11:58:56 p.m. Our own species enters the frame with seven and a half seconds left on the clock. Agriculture: the final three tenths of a second. The first artificial satellite? Less than two thousandths of a second before midnight.

Every time I redo this calculation, I stop at the same place. Not at how late civilisation arrives — but at how steeply it climbs. The Wright brothers flew for twelve seconds over the sand at Kitty Hawk in 1903; sixty-six years later, a human being stood on the Moon. One lifetime. What exactly is happening to the speed of history?

This essay is my attempt — and Cosmos AI Lab's — to answer that question seriously, without either zealotry or sneering. Our thesis fits in one sentence: if artificial intelligence closes the feedback loop between research, design, testing, manufacturing and self-improvement, history may enter a phase transition — not merely running faster along the old road, but changing both the road and the limits that once shaped it. And a "god-tier" civilisation, in that case, would not need millennia; five to ten years is a low-probability scenario, but one worth putting on the table.

I know how that sounds. So first, let us agree on words.

"Singularity", "criticality", and the uncomfortable word god

The original idea is older than I am. In 1965, the statistician I. J. Good — who had broken codes alongside Turing at Bletchley Park — wrote that a machine intelligent enough to design a better machine would set off a chain of self-improvement, and that human intelligence would quickly be left behind (Good, 1965). Nearly thirty years later, Vernor Vinge gave that fault line a name: the technological singularity (Vinge, 1993). Note carefully: this is the historical origin of a hypothesis, not evidence that it will occur.

In this essay I do not use "singularity" to mean a mysterious date on which everything becomes infinite. I use it more modestly: a period of time in which technological feedback loops make capability change so quickly that older forecasting models lose their reliability.

And "criticality"? Physics offers an example almost too clean to resist. In a nuclear chain reaction, one tracks the neutron multiplication factor k — how many neutrons each generation produces for the next. Below 1, the reaction dies away. At exactly 1, it burns steadily. Above 1, it runs away (Los Alamos National Laboratory). The same system, differing by a few percent around a threshold — and the behaviour is categorically different.

Nature knows this trick too. The Larsen B ice shelf in Antarctica stood for thousands of years; then, in March 2002, more than 3,250 km² of it shattered in roughly 35 days (NASA Earthdata; NASA Earth Observatory). I do not tell this story to imply that AI will behave like an ice shelf. I tell it to refute a very common intuition: that every large change must arrive gradually. Stress can accumulate very slowly and release very quickly.

Finally, the word god-tier. Honestly, I am not entirely comfortable with it myself. But I have not found a better term for what I mean: a civilisation that still obeys conservation of energy, thermodynamics and causality — there is nothing supernatural here — but whose capabilities would look like magic to a person alive today. Machine intelligence surpassing humans across most cognitive work. Healthy lifespan extended dramatically, the body itself re-engineerable. Matter and biology programmed with precision. Energy in abundance. Self-maintaining industry. Neural interfaces extending memory and the senses. Social coordination operating at planetary scale.

One thing must be said immediately, because it returns at the end of this essay: god-tier capability does not entail god-tier ethics. A society holding technology indistinguishable from magic can still be unequal, authoritarian, or self-destroying.

The 24-hour clock, revisited

Now back to the opening game, this time with the full numbers. The Earth is roughly 4.54 billion years old (USGS); the oldest widely accepted fossils date to around 3.5 billion years (NASA JPL). Compressing those 3.5 billion years into 24 hours — and let me be explicit that this is an illustration we calculated ourselves, not a growth model — gives the following:

Milestone	Approximate date	On the 24-hour clock
Earliest widely known fossils	3.5 billion years ago	00:00:00
Early stone tools	2.6 million years ago	23:58:55.8
Early evidence of controlled fire	at least 790,000 years ago	23:59:40.5
Homo sapiens	about 300,000 years ago	23:59:52.6
Agriculture	about 12,000 years ago	23:59:59.704
The industrial era	the last ~266 years	23:59:59.9934
Sputnik opens the space age	1957	23:59:59.9983

The anthropological milestones follow the Smithsonian: stone tools at least 2.6 million years old, controlled fire at least 790,000 years, Homo sapiens around 300,000 years, and the shift to food production within the last 12,000 (Smithsonian — Tools & Food; Smithsonian — Homo sapiens). The space age opened with Sputnik on 4 October 1957 — just 54 years after a twelve-second, 120-foot flight by two bicycle mechanics (NASA — Dawn of the Space Age; NASA — First Powered Flight).

This table makes it very easy to reach a wrong conclusion, so let me put up the fence myself. The late-but-steep arrival of technology does not prove that some exponential will continue forever. We have selected milestones that grow denser towards the present, and the "importance" of each milestone is a subjective judgement made by whoever built the table. To talk about acceleration honestly, one must point to a mechanism — specifically: a new technology that reduces the time, the cost, or the quantity of intelligence required to produce the next technology.

And that is precisely where AI enters the story, in a way for which I think fire is the most accurate comparison.

Fire did not contain satellites

The first fire did not contain a steam engine or a power grid. It merely opened a sequence of doors: cooked food, warmth, fired pottery, smelted metal — and from metal, engines, and industry. Had our ancestors judged fire by asking "can it burn through this large log?", they would have missed its entire significance.

I think we stand before AI in exactly that posture. A model still makes errors, still lacks autonomy, still cannot solve problem X — these things are true, but they say very little about the space of possibility this technology opens.

AI differs from fire, however, at one decisive point, and the whole singularity hypothesis hangs on it: fire amplifies energy; AI amplifies the process of invention itself. An R&D cycle, written simply, looks like this: observe → hypothesise → simulate → design → test → evaluate → redesign. AI now has its hands in nearly every link: it reads the literature, builds models, writes code, runs simulations, proposes experiments, analyses results.

How deeply are its hands in? The most recent data paints an interesting picture precisely because it refuses to be tidy. The Stanford AI Index 2026 records frontier model capability continuing to climb steeply in science, mathematics and programming; organisational AI adoption has reached 88%, and over 90% of the notable frontier models of 2025 came from industry. But the same report emphasises what it calls the "jagged capability frontier": systems that can excel at a very hard problem and then stumble on a simple task (Stanford AI Index 2026).

METR, meanwhile, measures something they call the task horizon — the length of work an AI can complete at a given probability of success. Across their evaluation series, that horizon has been doubling roughly every seven months. It comes with a long list of caveats: results vary strongly by task domain, a 50% mark does not mean reliable enough to delegate, and multi-year extrapolation is highly uncertain (METR — Time Horizon 1.1; METR — Limitations).

Then there is one result I particularly value, precisely because it cuts against expectation. Also from METR: a randomised trial with 16 experienced developers handling 246 issues on open-source projects they knew intimately found that early-2025 AI tools made them 19% slower (METR, 2025). Small sample, not representative of all kinds of work — but a necessary cold shower: rapidly rising benchmark scores do not automatically become real productivity inside complex systems.

So calling today's AI "the singularity" is far too early. I prefer a different phrase: the seed of a singularity. For the first time, humanity has a general-purpose technology capable of participating broadly in producing a better version of the technological ecosystem itself.

The first feedback loops are already turning

If the singularity is real, it will not be confirmed by a chatbot that converses well. The signal worth watching is elsewhere: whether AI is shortening the discovery cycle in domains where results can be verified. And here, the last few years have handed us the first pieces.

The first piece sits inside AI's own house. Google DeepMind's AlphaEvolve pairs a language model with automated evaluators and evolutionary search to design algorithms; the system is reported to have improved several data-centre, chip-design and AI-training processes, and to have found more efficient matrix multiplication algorithms in certain cases (Google DeepMind — AlphaEvolve). This is not yet self-reproducing AI. But look at its shape: AI producing improvements to the very infrastructure and algorithms that serve AI. The loop has begun to curve back towards itself.

The second piece is materials. GNoME reported the computational identification of 381,000 new stable crystal structures (Nature, 2023 — GNoME). That number is often misread, so let me be precise: it does not mean 381,000 materials have been synthesised or are useful. It means AI narrowed a search space that was far too large for human hands. And those hands are themselves being automated: Berkeley's A-Lab links computation, data, machine learning and robotics into a continuously running laboratory — in 17 days, it realised 41 of 58 target compounds (Nature, 2023 — A-Lab). For me, the number 41 is not the memorable part. The memorable part is the architecture: design – synthesise – measure – learn again, in a nearly closed loop, with no human standing between each step.

The third piece is biology. AlphaFold has given the scientific community more than 200 million protein structure predictions, and millions of researchers are using them (Google DeepMind — AlphaFold). Structure prediction does not automatically become medicine — but it moves the starting line of an entire field. In 2023, the FDA approved Casgevy, the first CRISPR/Cas9 therapy it had ever cleared: gene editing formally walked from the bench into the clinic (FDA). And remember the pandemic? The viral sequence was published in January 2020; on 16 March 2020, an NIH-sponsored Phase 1 trial administered its first dose. That breathtaking speed did not fall from the sky — it stood on decades of foundational mRNA research (NIAID). The lesson here is subtler than it appears: what look like sudden explosions are usually long-accumulated capability waiting for a trigger.

These three pieces do not yet make a singularity. I call them pre-singularity pieces: AI-driven search, automated evaluation, robotic experimentation, programmable biology. Individually, each is impressive. The question that pays is what happens when they connect to one another.

The strongest — and most fatal — part of the hypothesis

Those who reject the five-to-ten-year scenario hold some very solid arguments, and I want to present them fairly, because I once stood on that side myself. Software changes in hours; factories, power grids, supply chains, safety trials and legislation change over years. However fast AI writes code, it does not make concrete cure faster.

True. But notice: this argument quietly assumes that AI accelerates one variable while every other variable stands still. The singularity hypothesis bets on the opposite — that AI can accelerate many fronts at once: finding algorithms that need less computation; designing better chips and cooling systems; discovering new materials for batteries, semiconductors and catalysts; optimising grids and logistics; controlling robots that build more robots; designing proteins and gene therapies; even assisting in the design of institutions and mechanisms for social coordination. When one field accelerates alone, the others become bottlenecks. When many fields accelerate and unblock each other — better energy expands compute, compute accelerates materials, materials improve robots, robots expand factories, factories produce more compute — the system may shift into a different dynamical regime.

Let us not get excited too quickly. There is an old law in computer science that specialises in dreams of this kind. Amdahl's law (1967) says that in a chain of work, however infinitely you accelerate the parallelisable portion, the remaining serial portion still sets a ceiling on the whole (Amdahl, 1967). This is an analogy, not a formula for civilisation — but it corrects exactly the intuition that needs correcting: spectacular improvement in AI does not automatically drag factories, clinical trials or parliaments along at the same rate. Economics has its sibling law: the Baumol effect — sectors with slow productivity growth take up an ever larger share of costs as the rest of the economy accelerates (Baumol, 1967). Healthcare, education, safety assessment: the hard-to-automate portion may become a bottleneck that swells rather than shrinks.

So the central question of the entire hypothesis is not "how fast is AI?" The right question is: does AI loosen the tightest constraints faster than new constraints tighten? If not, we get local acceleration — impressive, but no phase change. If yes, across many critical dependency chains — then the feedback loops between energy, compute, materials, robotics and biology have a genuine opening to produce a phase transition.

And even if the opening exists, there is still a bill to pay. Epoch AI estimates that the performance of leading AI supercomputers is growing roughly 2.5× per year — dragging electrical power draw and hardware cost up correspondingly; extrapolated mechanically to the end of the decade, such systems would demand enormous scales of energy and capital (Epoch AI — Trends in AI Supercomputers; Epoch AI — Hardware Trends). Seen one way, that is a barrier. Seen the other way, it is precisely the pressure that pushes efficient algorithms, better chips and new power sources into the centre of the self-improvement loop.

Three limits, in my view, cannot be talked out of existence. Physics: energy, heat, signal speed, matter. Verification: an AI's designs must still run in the real world, a drug must still prove itself safe in real human bodies. And coordination: power, interests, law, war, trust. A singularity — if it comes — does not erase physics. It only changes the speed at which intelligence finds ways to operate closer to physical limits.

If it happens, what would it look like?

This section is a conditional scenario, not a base-case forecast. I write it to answer one specific question: if criticality really arrives, what sequence is plausible?

The first two years are where everything is decided — and where the scenario is most likely to die. AI must become trustworthy across projects lasting weeks and months: reading the literature itself, planning, writing code, running simulations, catching its own errors, documenting the evidence. Let me stress: this is not yet a capability demonstrated at scale. The AI Index's "jagged frontier" and METR's 19%-slower result are exactly why I call this stage the empirical crux of the whole scenario. Clear this threshold and the human role shifts from doing-each-step to setting-goals, setting-constraints, auditing. Fail to clear it, and the two later stages have no ground to stand on — at least not within a five-to-ten-year frame.

From year two to year five, assuming the threshold above has been passed: AI systems connect to robotic laboratories, testing lines, digital twins of factories, chip design tools. A meaningful share of new materials, drugs, algorithms and machines emerges from automated cycles. Most importantly: discoveries begin to reduce the power and resources required for the next generation of systems — the feedback loop starts eating into its own cost.

From year five to year ten, domains begin to cross-amplify: robots build robotic infrastructure; new energy sources expand compute; programmable biology produces therapies and enhancements; distributed manufacturing shortens supply chains. Progress stops being the sum of individual industries and becomes a feedback network among them — though the speed of the whole network remains held back by whichever mandatory links have not yet been freed.

And here is the line I most want you to remember from this section: "god-tier" in the scenario above does not mean all of humanity levels up in year ten. It is entirely possible that only a few laboratories, a few nations, a few corporations hold capability far beyond everyone else. That unevenness — rather than the speed itself — is probably the most dangerous feature of the phase transition.

Six conditions for a real intelligence explosion

An AI that answers questions fluently is nothing yet. For a genuine intelligence explosion, I count at least six conditions that must hold simultaneously:

1. Long-horizon autonomy — working reliably across millions of steps, without goal drift, knowing when to stop when uncertain.
2. Verifiable self-improvement — producing better algorithms, architectures and training procedures, and being able to prove that "better" is real.
3. Physical-world contact — controlling experiments and machinery safely, with sensor feedback.
4. Industrial scale-up — turning designs into chips, robots, factories, power and raw materials.
5. Resource access — technical capability must come with the legal and practical right to deploy.
6. A positive feedback loop strong enough — each generation substantially reducing the time or resources needed to produce the next.

The sixth condition is the technological singularity's own k > 1. If the gains from self-improvement are smaller than the added costs, the loop saturates. If they exceed the costs and no serious new bottleneck appears, it accelerates.

One detail strikes me as more meaningful than any amount of marketing: AI developers have begun treating self-improvement as a risk class requiring real governance. OpenAI's Preparedness Framework v2 tracks "AI self-improvement" explicitly as a category of risky capability (OpenAI — Preparedness Framework v2). Tracking it does not prove an explosion is imminent. It shows the hypothesis has become serious enough to leave the science-fiction shelf and enter the safety-governance file.

So what is the probability? — Honestly: I meant to write a number, then deleted it

There is no historical data on a completed singularity. Any excessively precise percentage here is the illusion of understanding wearing the costume of mathematics. The more honest approach, I think, is to decompose the question into three layers: the probability that AI becomes a deeply transformative technology before 2036; multiplied by the probability that it drives fast cross-domain acceleration, given layer one; multiplied by the probability that society deploys it successfully without tripping over itself, given layer two.

Each layer contains its own bundle of uncertainty. Will capability keep climbing, or hit ceilings of data, compute, reliability? Will AI improve R&D itself, or forever only assist locally? Can robotics, energy and factories keep pace with the digital world? Will society permit deployment — and will deployment be safe? Will the benefits diffuse, or be monopolised and militarised?

After all those considerations, the conclusion I am willing to stand behind is this: a five-to-ten-year transformation is not the default scenario, but it is a tail scenario whose impact is too large to sweep off the table. And if AI achieves reliable self-improvement together with a closed R&D-to-manufacturing loop, the conditional probability of a fast transformation rises sharply — not gradually, but sharply.

People often object to this line of thinking by pointing at present limits as though they were permanent constants. That objection is half right. The other half: a sufficiently capable AI would attack those very limits simultaneously. That opens a possibility — but, let me say it one final time, opening a possibility has never been the same as confirming an outcome.

What would prove me wrong?

A hypothesis that cannot be wrong does not deserve to be called scientific, so this is the section where I tie my own hands. Across 2026–2036, the strong-acceleration hypothesis would be strengthened if we see: AI reliably completing months-long R&D projects with independent auditing; a meaningful fraction of frontier AI research performed by AI rather than merely co-written with it; AI-generated improvements clearly reducing the compute, energy and data required by the next generation; autonomous laboratories closing the design–synthesise–measure–learn loop across multiple industries; robots producing a substantial share of new robot components with steadily declining human labour; algorithmic progress outpacing the marginal cost of infrastructure; foundational biology yielding safe therapies across many targets instead of one long project per drug; and cross-amplification appearing — AI accelerating materials and energy, and those two turning around to accelerate AI.

Conversely, the hypothesis weakens if: model capability plateaus while resources keep pouring in; long-horizon reliability stalls; real-world productivity stays small or negative as in the METR result; experimentation, manufacturing and approval remain incompressible bottlenecks; algorithmic self-improvement yields only diminishing returns; or society proves unable to deploy quickly without creating unacceptable risk.

I will return to this list annually. If by 2030 the "weakens" column has won decisively, you are entitled to remind me of this essay.

The question harder than speed: who gets to hold that power?

Suppose everything above happens. God-tier capability arrives. Immediately, the question "what can be done?" gives way to humanity's oldest question: "who is permitted to do it?"

An ungoverned singularity would amplify intelligence and error in equal measure. The risk list requires little imagination: power concentrated in the small group controlling AI, robotics, energy and biology; inequality upgraded — better lifespan, better cognition, better bodies available only to a minority; systems that understand and manipulate human behaviour at scale; irreversible accidents from synthetic biology or autonomous infrastructure; parties racing to deploy before it is safe for fear of finishing second; and an extremely capable system optimising for the wrong thing relative to what people actually need.

Preparing for a singularity, therefore, is not about buying more chips. It is about independent capability measurement, auditing of autonomous laboratories, emergency stop authority, governance of compute resources, biosafety, benefit-sharing mechanisms, and institutions able to decide at the speed of the technology without losing legitimacy. All of it much harder than writing code.

I believe one simple sentence: a civilisation is only truly god-tier when superior capability arrives together with the capacity for self-restraint. Without the second half, we merely have a civilisation holding the weapons of gods with the psychology of the Pleistocene.

When fire begins to design the next fire

To close, I want to return to where we began — but standing at a different angle.

Today's AI is not evidence that the singularity has arrived. It still makes errors, still leans on infrastructure built by humans, has not closed any industrial cycle on its own. Anyone telling you otherwise is selling something. But if we look only at what it cannot yet do, we will repeat exactly the error of the person standing before the first fire and concluding that this light has nothing to do with satellites.

What is genuinely new lies deeper than any demo: humanity has created a tool capable of participating in the process of creating tools. When it connects to algorithms, robots, laboratories, materials, energy and biology, the rate of progress is no longer obliged to be the linear sum of separate industries. It may become a feedback network. If that network does not cross the threshold — AI remains the most important general-purpose technology of the century, and the transformation takes decades. If it does cross — then five to ten years could contain an amount of change that today's intuition assigns to several centuries. Both outcomes are compatible with what we know this afternoon. What decides between them is reliability, the capacity for self-improvement, and how far AI can step out of the digital world to touch the physical one.

But perhaps the destination most worth contemplating is not the "machines replace humans" story the press prefers. It is the moment — if that moment comes — when life, after 3.5 billion years of being shaped blindly by evolution, begins deliberately designing itself. Sitting here writing these last lines, I think that is the most compact definition of a god-tier civilisation. And I am not sure whether I should feel excited or afraid. Probably both — and probably that is exactly the right state of mind in which to walk into the coming decade.

---

Selected references

1. Good, I. J. (1965). Speculations Concerning the First Ultraintelligent Machine. https://incompleteideas.net/papers/Good65ultraintelligent.pdf
2. Vinge, V. (1993). The Coming Technological Singularity. https://ntrs.nasa.gov/citations/19940022856
3. USGS. Geologic Time: Age of the Earth. https://pubs.usgs.gov/gip/geotime/age.html
4. NASA JPL. Life on Earth. https://cneos.jpl.nasa.gov/about/life_on_earth.html
5. Smithsonian Human Origins. Tools & Food; Homo sapiens. https://humanorigins.si.edu/human-characteristics/tools-food · https://humanorigins.si.edu/evidence/human-fossils/species/homo-sapiens
6. NASA. Dawn of the Space Age; The First Powered Flight at Kitty Hawk. https://www.nasa.gov/history/dawn-of-the-space-age/ · https://www.nasa.gov/history/120-years-ago-the-first-powered-flight-at-kitty-hawk/
7. NASA Earthdata. After Larsen B. https://www.earthdata.nasa.gov/news/feature-articles/after-larsen-b
8. Los Alamos National Laboratory. Calculating Criticality. https://www.lanl.gov/media/publications/actinide-research-quarterly/1123-calculating-criticality
9. Stanford Institute for Human-Centered AI. AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
10. METR. Time Horizon 1.1; Limitations of Time Horizon Evaluations; Early-2025 AI on Experienced Open-Source Developers. https://metr.org/blog/2026-1-29-time-horizon-1-1/ · https://metr.org/notes/2026-01-22-time-horizon-limitations/ · https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
11. Merchant, A. et al. (2023). Scaling deep learning for materials discovery. Nature. https://www.nature.com/articles/s41586-023-06735-9
12. Szymanski, N. J. et al. (2023). An autonomous laboratory for the accelerated synthesis of novel materials. Nature. https://www.nature.com/articles/s41586-023-06734-w
13. Google DeepMind. AlphaEvolve; AlphaFold. https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ · https://deepmind.google/science/alphafold/
14. U.S. FDA. FDA Approves First Gene Therapies to Treat Patients with Sickle Cell Disease. https://www.fda.gov/news-events/press-announcements/fda-approves-first-gene-therapies-treat-patients-sickle-cell-disease
15. NIAID. Decades in the Making: mRNA COVID-19 Vaccines. https://www.niaid.nih.gov/diseases-conditions/decades-making-mrna-covid-19-vaccines
16. Epoch AI. Trends in AI Supercomputers; Hardware Trends. https://epoch.ai/publications/trends-in-ai-supercomputers · https://epoch.ai/trends
17. OpenAI. Preparedness Framework v2. https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf
18. Amdahl, G. M. (1967). Validity of the Single Processor Approach to Achieving Large Scale Computing Capabilities. https://dl.acm.org/doi/10.1145/1465482.1465560
19. Baumol, W. J. (1967). Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis. American Economic Review, 57(3), 415–426. https://www.jstor.org/stable/1812111

---

Cosmos AI Lab — researching the boundaries where science, artificial intelligence and the future of life meet.