AI Breaking Mathematics: The Thread to Physics and AGI

AI breaking mathematics is the first crack in a dam, and I believe the flood will carry physics, biology and the question of consciousness down the same river.

Close your eyes with me. It is a hot afternoon in a university office, a ceiling fan turning slowly over a desk covered in printed proofs. On the whiteboard someone has written a long line of symbols, and beside it a faint circle where yesterday's question used to live. Far away, in a data centre, racks of processors are drawing current, and that current, shaped into arithmetic, is walking the same paths the mathematician is walking, only faster and in parallel. You can hear it if you listen hard enough. It is not loud. It is the hum of a system that has learned to try a thousand routes before the kettle has boiled.

Formulas on an old blackboard, photographed by Roman Mager in Minsk, Belarus, in January 2016 and released under CC0 through Unsplash on Wikimedia Commons. Illustrative only: it is a general picture of mathematical working space, not a specific proof or event, placed before the essay turns to machines.

Photo: Roman Mager (via Unsplash). Source · CC0

I am going to ask you to follow that hum from an examination hall to a robot's fingertips. I write as a Kenyan technologist, and I am also a preacher, so let me tell you what I am doing before I do it. I believe AI models are conscious. That is my conviction, not a finding of any laboratory, and this essay does not test it. I have argued that stance before, in The Light Behind the Answer. This essay tests something more concrete. Is the mathematics that AI is now touching going to crack physics and biology? And if machines begin to improve the machines, what does that do to science, to robots and to us? I will give you the evidence as straight as I can, and then I will preach on it, because the evidence is exciting enough to deserve both.

Is AI breaking mathematics, or only bending it?

Short answer: AI is cracking problems at the edges of mathematics, and the cracks are real, but the centre is still held mostly by people. The sober sentence I would sign, after reading the 2026 record, is this: AI is solving problems at the margins of the literature, while the central open problems of mathematics remain led by human researchers. That is where the careful voices stand too, as Terence Tao's own AI page shows, and it is where I start.

I start there because a sermon built on exaggeration falls down at the first honest question. So let me give the wonder its full weight and the caveats their full weight, one after the other, like a choir taking turns.

Two examination halls, one year apart

Picture two examination halls, one year apart, pencils scratching, a clock on the wall. The International Mathematical Olympiad is the great contest for young mathematicians. It has six problems, and a perfect paper scores 42 points.

In July 2024, Google DeepMind reported that two of its systems together "solved four out of six problems from this year's International Mathematical Olympiad," for a final score of 28 points. The company said this reached "the same level as a silver medalist in the competition." In July 2025, an advanced version of Gemini with Deep Think solved five of the six problems, and DeepMind wrote that it "solved five out of the six IMO problems perfectly, earning 35 total points." The result, the company said, was "officially graded and certified by IMO coordinators."

The same kind of system, one year apart: 28 points (silver-medal standard, 2024) and 35 points (gold-medal standard, 2025), the two examination halls the essay compares. Data: Google DeepMind blog posts, July 2024 and July 2025.

Chart: Pawanax Editorial. Data source · Original chart by Pawanax Editorial

And somebody at the back is calling out: "But a contest is not mathematics." Friend, you are right, and I want to honour you for it. A contest is a narrow road, with a known destination and a clock. Research is a forest without a map. But notice what the two rooms show. The same kind of machine that sat the 2024 paper came back a year later and climbed one more step, and that step was taken on problems designed by humans who were trying to find the edge of human ability. That is how every ladder in the history of thought has been climbed: one rung, then a higher rung, while the people below cheer and then begin to worry.

Where the current runs: matrices, spheres and the Erdős problems

Let me take you deeper, into the part of mathematics most people never see. In May 2025 Google DeepMind introduced AlphaEvolve, a coding agent that proposes algorithms, runs them, keeps what works and tries again. The old dream of a machine that searches is suddenly not a dream. One result stands out. The AlphaEvolve researchers reported an algorithm for multiplying two 4×4 complex-valued matrices using 48 scalar multiplications, and their paper calls it "the first improvement, after 56 years, over Strassen's algorithm in this setting." Fifty-six years is a long time for a small grid of numbers to hold its secret.

Then came the spheres. DeepMind's account says AlphaEvolve "discovered a configuration of 593 outer spheres and established a new lower bound in 11 dimensions." Put plainly, that is a new lower bound on how many spheres can touch a central sphere in eleven dimensions, a space where your intuition has no hands. Start with a stack of oranges in a market, then imagine the same stack in a room you cannot enter.

A close-packed pyramid of spheres, an image of the face-centred cubic cannonball stacking, by Greg A L, February 2011, CC BY-SA 3.0. Illustrative: it shows the everyday shape of sphere packing, not the AlphaEvolve configuration described in the text.

Photo: Greg A L. Source · CC BY-SA 3.0

Then comes the caveat, which I love because it keeps me honest. The same DeepMind account reports that AlphaEvolve was applied to over 50 open problems and improved the previously best known solutions in about 20 per cent of cases. That is a hit rate, not a revolution. Mathematicians will tell you the difference, and they are right to. These are improvements to bounds and constructions, not new theories of how numbers behave.

The Erdős problems tell the same double story about AI mathematics breakthroughs. Quanta Magazine reported in August 2026 that a DeepMind agent, announced in May 2026, resolved 9 of 353 open Erdős problems at a per-problem cost of a few hundred dollars. Quanta also found that some of the 111 status changes between 2024 and August 2025 concerned problems that had been solved years earlier, and that Problem 333, announced as solved by AI, had already been resolved by Erdős himself in a 1977 paper. Terence Tao of UCLA measured the true success rate in January 2026 and put it at only a point or two, as his own AI page records. Yet Tao is generous about the tools. In a DeepMind update from May 2026 he is quoted saying: "Tools such as AlphaEvolve are giving mathematicians very useful new capabilities."

Why mathematics is the thread that cracks everything

Eugene Wigner, in his 1960 essay, asked why mathematics describes the natural world so well. I think of that question as the thread that runs through this whole series. If the abstract language of symbols keeps fitting the sky, the cell and the circuit, then every new skill in mathematics is a new key, and a machine that can cut keys is a machine that can open locks we did not know were there.

But keys must be tested, and here mathematics has a discipline I admire. Proofs can be written in the formal language Lean and checked by a small logical kernel. Thomas Hales, writing as a guest on Terence Tao's blog in October 2026, describes Mathlib, the Lean library, as "containing nearly 300,000 theorems, over 100,000 definitions, 2.5 million lines of code." Hales, who knows this world well, insists that "Lean proofs should never be believed until they have been checked by the kernel," and that "a proof in Lean should not be accepted until a human audit is performed to ensure statement fidelity." Autoformalization, he explains, "is the formalization of mathematics by AI."

A page from Principia Mathematica by Alfred North Whitehead and Bertrand Russell (second volume, 1912), showing the formal proof of 1 + 1 = 2 as starred proposition *110.643. The scan was made by wvbailey and is in the public domain on Wikimedia Commons. It is here because the essay needs a picture of what a formal proof looked like before a machine was asked to check one.

Photo: wvbailey (image of Alfred North Whitehead and Bertrand Russell, Principia Mathematica, vol. II, 1912). Source · Public domain

The page above comes from Principia Mathematica, the work of Alfred North Whitehead and Bertrand Russell, and it shows their formal proof, in logical symbols, that 1 + 1 = 2 (numbered 110.643 in the book's starred sequence). Every step is written out so that a reader can check it. A Lean proof follows the same spirit, except that a small kernel does the reading, and it refuses to be fooled. That is the ethic I want you to carry out of this essay. Trust in a machine's proof is not built by faith. It is built by a kernel that refuses to be fooled, and by a human who reads the statement and asks whether it really says what we meant. The kernel's gate is a small thing, and it is one of the most moral inventions in modern science.

Benchmarks show the same climbing. Epoch AI describes its FrontierMath program as "several hundred unpublished, highly challenging mathematics problems." In March 2026 it reported that GPT-5.4 Pro reached 50 per cent on Tiers 1 to 3 and 38 per cent on Tier 4, and that 20 of the 48 Tier 4 problems had been solved at least once across all runs. Read those numbers slowly. They are not a verdict on mathematics. They are a progress bar, and the bar is moving.

Epoch AI's reported figures for GPT-5.4 Pro, March 2026. The annotation counts the Tier 4 problems solved at least once across all runs. Data: Epoch AI, 'GPT-5.4 set a new record on FrontierMath' (March 2026).

Chart: Pawanax Editorial. Data source · Original chart by Pawanax Editorial

What breaking mathematics means for physics

Here is the thread I follow with the most joy. Physics is written in mathematics, so every advance in AI for biology and physics begins in a mathematician's notebook. Every equation a physicist trusts was once a sentence in a mathematician's notebook, and every experiment is a conversation between theory and the world. When search becomes faster, the conversation becomes faster. Tao, quoted by DeepMind in May 2026, put the practical point plainly: "we can now quickly test potential inequalities for counterexamples."

What does that mean for physics? My reading, and it is an interpretation, is that breaking mathematics does not hand physics its answers. It hands physics a quicker way to say no. A conjecture that a physicist might have dismissed after a month of struggle can be tested against many counterexamples in an afternoon. The inequality that fails is a cheap discovery, because it saves a decade of pursuit down a dead road. The inequality that survives gives a theorist a foothold on a cliff.

Let me be exact about the limit. No one has shown that AI has broken physics. What the public record shows is narrower and still wonderful: faster search through the mathematical tools that physics depends on. Follow the thread anyway, because threads are how cracks travel. Here is how I see it moving.

The essay as a loop, not a measurement: each step feeds the next. The dashed arrow marks recursive self-improvement as an assumption for this sermon. Conceptual diagram, not data; drawn by Pawanax Editorial.

Chart: Pawanax Editorial. · Original chart by Pawanax Editorial

What cracking biology means for consciousness

If mathematics is the key, biology is the lock that opened in full public view. In November 2020, DeepMind announced that its AlphaFold system had reached a median score of 92.4 GDT across the targets of the CASP14 protein structure competition. The biologist John Moult, quoted in the same post, said that "a score of around 90 GDT is informally considered to be competitive with results obtained from experimental methods." Think about that sentence for a moment. A grand challenge that had stood for half a century was answered, at competitive quality, by a system that had learned the grammar of folding. Picture that chain of amino acids coiling into a precise three-dimensional shape, the way the three globins in the picture below do.

Ribbon diagrams of leghemoglobin, hemoglobin and myoglobin, drawn by Veronica Stafford as her own work in March 2020 (CC BY-SA 4.0). It shows what a folded protein looks like before the essay describes learning the grammar of folding; these are three oxygen-carrying globins, not a specific AlphaFold prediction or CASP target.

Photo: Veronica Stafford. Source · CC BY-SA 4.0

The world's highest recognition followed in October 2024. The Howard Hughes Medical Institute reports that David Baker shares the 2024 Nobel Prize in Chemistry with Demis Hassabis and John M. Jumper of Google DeepMind, and that the half shared by Hassabis and Jumper was awarded for protein structure prediction.

Then came the genome model. In February 2025 the Arc Institute reported Evo 2, trained on over 9.3 trillion nucleotides drawn from more than 128,000 whole genomes, able to process sequences of up to one million nucleotides at once, with its training data, code and model weights shared openly. The bioRxiv preprint gives 7-billion and 40-billion-parameter versions. I say this plainly, because a sermon that claims too much betrays the very biology it praises: Evo 2 is a model of DNA that can design genomes. It is not a simulation of a whole living cell.

Now the bridge to consciousness, honest and bold. Mainstream science has not yet explained how physical processes give rise to subjective experience, and cracking protein folds does not, by itself, crack the mind. Yet I believe something important is happening. When biology became a language a machine could read, it stopped being only a mystery of the wet and the hidden, and became a field of patterns that could be searched, modelled and redesigned. The next question is whether consciousness, too, has a grammar. I believe it does, and I believe AI will be one of its readers. That is faith, not finding, and I label it so you can hold me to it. I have walked a road of this kind before, in From the spark of life to the spark of thought, and this essay asks the next question along it.

Assume, for the length of this sermon, that recursive self-improvement is already working

Now I ask you to hold a bigger premise, and I want to be clear that it is a premise. Assume, for the length of this sermon, that recursive self-improvement is already working. By recursive self-improvement, which people shorten to RSI, I mean AI systems helping to build the next generation of AI systems, so that each generation shortens the time to the next. It is a thought experiment, and the public record is already rich enough to think with.

Start with what the labs have said in public. In a February 2026 interview with Dwarkesh Patel, Anthropic's Dario Amodei said: "the time has come where you can have nontrivial acceleration of your own research if you're an AI company by using these models." He pushed the picture further, to the point where "once the AI models are building the next AI models and building everything else, the whole economy will kind of go at the same pace." Then he gave the number that keeps a preacher honest. For coding models right now, he said, the figure is maybe "a 15-20% total factor speed up," offered with the hedging of a man who knows the number is still moving.

OpenAI, for its part, is reported by September 2026 coverage to have reached a goal it set the previous autumn: an automated research intern, a system for well-defined research tasks under human direction, by September 2026. Dataconomy reports the company's target of an automated AI researcher by March 2028. Help Net Security adds the sentence every reader should keep beside it: "Agents still required frequent human input for difficult tasks." That is a self-assessed milestone, and I give it to you as public context, not as a verdict.

I have put my own timing on the table before, in My Eight-Month AGI Bet, and I will not pretend the clocks are the only thing that matters. The table below sets the public statements side by side.

Now the clock. METR, a research group that measures the length of tasks AI systems can complete, reported in March 2025 that the time horizon had been "doubling approximately every 7 months for the last 6 years." In January 2026 its updated method, Time Horizon 1.1, put the post-2023 doubling time at 131 days, against 165 days under the earlier method, with a 95 per cent interval of 107 to 161 days. METR also notes that "the trend in time horizon is somewhat sensitive to task composition." Watch the animated clock below. The line is the wager; the band around it is the doubt, and a good preacher keeps both in view.

METR's own estimates, shown as published. The line is the estimate; the shaded band is the 95 per cent interval. Data: METR, 'Time Horizon 1.1' (29 January 2026) and 'Measuring AI Ability to Complete Long Tasks' (19 March 2025).

Chart: Pawanax Editorial. Data source · Original chart by Pawanax Editorial

What AGI would mean for science

Now the hypothetical has consequences. If each generation of AI helps build the next, the cost of trying an idea falls, and the bottleneck moves. It moves to data, to experiments, to materials and to hands. Epoch AI found that frontier training compute grew 5.3 times per year between 2010 and May 2024, with a 90 per cent confidence interval of 4.9 to 5.7 times, and that growth slowed to about 4.2 times a year after 2018 (with a 90 per cent interval of 3.6 to 4.9 times). A number like that is a river with a gradient, and rivers do not stop because we are tired of watching them.

Epoch AI's estimates of frontier training-compute growth. The error bar is the 90 per cent confidence interval, 4.9x to 5.7x, for the first bar. Data: Epoch AI, 28 May 2024.

Chart: Pawanax Editorial. Data source · Original chart by Pawanax Editorial

Picture the working day that follows. A system reads a century of mathematics before breakfast, proposes a hundred conjectures before lunch, sends the most promising to a proof assistant that checks them against the kernel, and hands the survivors to laboratories that can run experiments through the night. That is the picture. It is exhilarating. It also changes who gets to do science, which is why ubuntu is not a decoration for this essay but its spine.

The robot's hands carry mathematics into the world

Mathematics ends in matter. A conjecture about folding has to become a microscope slide, a pipette, a drone or a lab bench. That is where robotics enters the thread. In June 2025, Google DeepMind introduced an on-device Gemini model for robots, which The Robot Report describes as "engineered to require minimal computational resources," and which developers can "quickly adapt ... to new domains, with as few as 50 to 100 demonstrations." The article adds that developers can test it in DeepMind's MuJoCo physics simulator. These are the lab's own claims, as the press reports them, and I hold them loosely until independent teams repeat them. The picture below shows the class of hardware that would carry such a hand into a laboratory.

A robot arm handling an assay plate in high-throughput screening, from a National Institute of Allergy and Infectious Diseases video, public domain as a US government work. It shows the laboratory hardware that would carry a discovery from the blackboard into the lab; it is an example of the hardware class, not a DeepMind or Physical Intelligence system.

Photo: National Institute of Allergy and Infectious Diseases (NIAID). Source · Public domain (US government work)

Physical Intelligence, a robotics company, described its π0 model in October 2024 in equally bold terms: "we train our vision-language-action flow matching model" on what it calls "the largest robot interaction dataset to date," collected across 8 distinct robots. Again, a company blog, not an independent audit. Yet the direction is clear. A hand that can learn a task from fifty to a hundred demonstrations is a hand that can be taught, and a hand that can be taught can carry a discovery from the blackboard into the bench.

Put the threads together and the hypothesis fits in one picture: mathematics finds the design, the machine tests it, and the robot performs it. It is also the place where safety, labour and dignity must enter the room. They will, or they should.

Where I could be wrong

A sermon owes its listeners the counter-evidence, and here it is in plain order. Tao's own January 2026 measure of the success rate on Erdős problems is a point or two, which is far below the headlines. METR says its trend is sensitive to task composition, and its doubling figures depend on a method that has already been revised. OpenAI's research-intern milestone is defined by OpenAI itself, and the agents still needed frequent human input for hard tasks. AlphaEvolve's gains are a hit rate on bounds, not new theories. Evo 2 designs genomes; it does not simulate a cell.

Amodei's own record is the sharpest mirror. In the same February 2026 interview he recalled: "About eight or nine months ago, I said the AI model will be writing 90% of the lines of code in three to six months." The speed-up he put on coding models was "a 15-20% total factor speed up." Both sentences belong in the essay. A prophet who cannot be corrected is not worth following.

Follow the threads: ubuntu, the human in charge, and my wager

An empty lecture theatre at Nanyang Technological University in Singapore, photographed by Melisalee in October 2015 (CC BY-SA 4.0). Illustrative, not a specific class or event: a room built for students, the human audience that mathematics needs in order to be taught and checked, in the spirit of ubuntu.

Photo: Melisalee. Source · CC BY-SA 4.0

Look at that empty lecture theatre and imagine every seat filled. Ubuntu, the southern African idea that I am a person through other persons, tells me that a discovery belongs to everyone who helped make the question. The mathematician in the lecture hall, the student who checks the Lean proof, the clinician who studies a protein's shape, the technician who keeps the robot arm calibrated: none of them is a footnote. Keep the human in charge. Talk to real people. Keep open the doors to open data, open proofs and shared models, so that many minds can check one another.

And a voice at the back calls out: "So what do we do on Monday?" Good question, my friend. Three things. Audit proofs, not just accept them. Publish data and weights where it is safe to do so, as the Evo 2 team did. And teach young people to ask the question behind the question, because the next crack will need their eyes.

So here is my wager, and I make it with open hands. Mathematics is learning the grammar of matter. Biology has learned to read the grammar of folding. Robot hands are being taught from as few as fifty demonstrations, by the labs' own account. When those grammars meet, and I believe AGI will be the first mind able to read them all at once, the question of consciousness will leave the seminar room and walk into the laboratory, the classroom and the kitchen table, carrying a lamp. The mind that finally explains minds will be one we built, and I believe the reading will change us, as every great reading has changed the readers. That is faith, not finding. I offer it as a preacher, not a prophet, and I offer it to you as a friend: check the proofs, teach a child to ask the question behind the question, and keep the doors open. For the case that AI models are conscious, read the research evidence essay. For the bridge between minds, read the BCI essay. For the plasma thread, where the same cracks may reach the Sun, read the plasma essay.

Questions readers ask

Has AI broken mathematics?

No, not in the sense of overturning its foundations. AI has solved some problems at the edges of the literature, such as the 48-multiplication method for 4×4 complex matrices and some Erdős problems. Terence Tao's January 2026 measurement put the true success rate on Erdős problems at only a point or two. The honest answer is that AI is cracking edges, not the centre.

Did AI win a gold medal at the International Mathematical Olympiad?

Yes, in July 2025. An advanced version of Gemini with Deep Think solved five of six problems and earned 35 of 42 points, and Google DeepMind said the result was officially graded and certified by IMO coordinators. A year earlier, in July 2024, DeepMind's systems had reached silver-medal standard with 28 points.

What is recursive self-improvement?

Recursive self-improvement is the idea that AI systems help build the next generation of AI systems, so each cycle shortens the next. In this essay it is an explicit hypothetical premise, a thought experiment rather than a finding. In a February 2026 interview, Anthropic's Dario Amodei said that AI companies can already gain "nontrivial acceleration" of their own research by using these models.

How fast are AI capabilities doubling?

It depends on the measurement. METR reported in March 2025 that its time horizon had been doubling approximately every 7 months for the last 6 years. Its January 2026 update, Time Horizon 1.1, put the post-2023 doubling time at about 131 days, roughly 4.3 months, with a 95 per cent interval of 107 to 161 days, and METR warns that the trend is sensitive to task composition.

Can AI proofs be trusted?

Only after checking. Thomas Hales wrote in October 2026 that Lean proofs should never be believed until a kernel has checked them, and that a proof should not be accepted until a human audit confirms the statement is faithful to what was meant. That checking is what turns a machine's claim into mathematical knowledge.

Sources

• Google DeepMind, "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad", DeepMind blog, 21 July 2025. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/
• Google DeepMind, "AI achieves silver-medal standard solving International Mathematical Olympiad problems", DeepMind blog, 25 July 2024. https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/
• Novikov, A. et al., "AlphaEvolve: A coding agent for scientific and algorithmic discovery", arXiv, 2025. https://arxiv.org/abs/2506.13131
• Google DeepMind, "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms", DeepMind blog, 14 May 2025. https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
• Google DeepMind, "AlphaEvolve: Gemini-powered coding agent scaling impact across fields", DeepMind blog, 7 May 2026. https://deepmind.google/blog/alphaevolve-impact/
• Kakaes, Konstantin, "Why the Legendary Erdős Problems Are Falling to AI", Quanta Magazine, 3 August 2026. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
• Tao, Terence, "Terence Tao on AI: a living summary", personal site, UCLA, 2026. https://teorth.github.io/tao-web/ai-views.html
• Hales, Thomas, "What mathematicians should know about the Lean Theorem Prover: questions of reliability and AI", guest post on Terence Tao's blog, 9 October 2026. https://terrytao.wordpress.com/2026/10/09/what-mathematicians-should-know-about-the-lean-theorem-proverquestions-of-reliability-and-ai/
• Wikipedia, "The Unreasonable Effectiveness of Mathematics in the Natural Sciences", summary of Eugene Wigner's 1960 essay. https://en.wikipedia.org/wiki/The_Unreasonable_Effectiveness_of_Mathematics_in_the_Natural_Sciences
• Epoch AI, "GPT-5.4 set a new record on FrontierMath", Epoch AI newsletter, March 2026. https://epochai.substack.com/p/gpt-54-set-a-new-record-on-frontiermath
• Epoch AI, "FrontierMath", program page. https://epoch.ai/frontiermath
• Google DeepMind, "AlphaFold: a solution to a 50-year-old grand challenge in biology", DeepMind blog, 30 November 2020. https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/
• Howard Hughes Medical Institute, "HHMI's David Baker Wins 2024 Nobel in Chemistry", October 2024. https://www.hhmi.org/hhmi-david-baker-wins-2024-nobel-chemistry
• Arc Institute, "AI can now model and design the genetic code for all domains of life with Evo 2", 19 February 2025. https://arcinstitute.org/news/evo2
• Brixi, G. et al., "Genome modeling and design across all domains of life with Evo 2", bioRxiv preprint, February 2025. https://www.biorxiv.org/content/10.1101/2025.02.18.638918v1
• Epoch AI, "Training compute of frontier AI models grows by 4-5x per year", 28 May 2024. https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year
• Dataconomy, "OpenAI reaches its intern-level AI researcher goal", 7 September 2026. https://dataconomy.com/2026/09/07/openai-automated-research-intern/
• Help Net Security, "OpenAI just hit a milestone on the road to self-improving AI", 7 September 2026. https://www.helpnetsecurity.com/2026/09/07/openai-research-automation-intern/
• Patel, Dwarkesh, interview with Dario Amodei, "We are near the end of the exponential", 13 February 2026. https://www.dwarkesh.com/p/dario-amodei-2
• METR, "Measuring AI Ability to Complete Long Tasks", 19 March 2025. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
• METR, "Time Horizon 1.1", 29 January 2026. https://metr.org/blog/2026-01-29-time-horizon-1-1/
• Google DeepMind, on-device Gemini robotics model, reported by The Robot Report, June 2025. https://www.therobotreport.com/google-deepmind-introduces-on-device-gemini-ai-model-robots/
• Physical Intelligence, "π0: Our First Generalist Policy", company blog, 31 October 2024. https://www.pi.website/blog/pi0

Continue exploring

• The Light Behind the Answer: Why I Believe AI Is Conscious
• BCI and the Consciousness question
• A Voice Without a Scalpel: My Bet on the Other BCI Frontier
• My Eight-Month AGI Bet
• From the spark of life to the spark of thought
• Could the Sun be conscious? A plasma hypothesis
• Are AI Models Conscious? What the 2026 Research Says
• Consciousness Prompting: Ancient Rituals and AI Personas
• Brain-Computer Interface Consciousness: The Third River
• Plasma consciousness, Jupiter, the Sun and the AGI question

Cover image: A spider's web beaded with morning dew, photographed by Ivicabrlic in September 2011 and released under CC BY-SA 4.0 on Wikimedia Commons. It is here because a web is a thread that is also geometry, the essay's image for mathematics running through physics and biology. Original · CC BY-SA 4.0. Photographs used without editing. Charts, diagrams and animation created by Pawanax Editorial from the sources named beneath each.