AI Scaling Laws Explained: Is AGI Closer Than We Really Think?

The Angle. AI's rise follows laws as regular as any in physics: more compute, better algorithms and more data buy predictable gains, and the length of task a frontier model can finish on its own has been doubling roughly every four months since 2023.

AGI is not a switch that flips on one day. It is a staircase, and on digital, well-defined work we are climbing it fast. On novel, open-ended and physical work, the stairs are still steep.

The Trick

Ask a chatbot to write a sonnet, fix your code or explain your blood test, and it feels like magic. The trick works because you only see the stage. Backstage there is no wizard. There are three dials and a power bill.

1. The first dial is compute: how many calculations go into training a model.
2. The second is algorithms: how cleverly each calculation is used.
3. The third is data: how much text, code and video the model learns from.

Since 2024 a fourth dial has joined them: thinking time, the compute a model spends reasoning before it answers. Turn any dial up and performance rises in a remarkably predictable way.

That predictability is the real reveal. Isaac Newton gave us laws that describe how things move. AI researchers have found laws that describe how intelligence, at least the machine kind, improves with resources. Like Newton's laws, they work beautifully inside their range and break at the edges. This piece is about both: the laws, and the edges.

Mojo's Law: the neural scaling law

In January 2020, a team at OpenAI led by Jared Kaplan showed that a language model's error, its “loss”, falls as a smooth power law of the compute used to train it. Written down, it is almost disappointingly simple:

L(C) = \left(\frac{C_c}{C}\right)^{\alpha_C}, \qquad \alpha_C \approx 0.050

L is the loss, C is training compute and C_c is a constant. The exponent is the whole story. With α at about 0.05, multiplying compute by 10 cuts the loss by only about 11%, because 10 to the power of minus 0.05 is about 0.89. Each tenfold step buys a similar, modest improvement, for ever, until something else runs out. Kaplan's team found matching power laws for model size and data.

Two years later, DeepMind's Chinchilla paper (Hoffmann and colleagues, March 2022) corrected the recipe. For a fixed budget, parameters and training tokens should grow together, which works out at roughly 20 tokens of training data per parameter. Chinchilla, with 70 billion parameters trained on 1.4 trillion tokens, beat a model four times its size trained on the same compute.

And the dials multiply. If frontier training compute grows about 5 times a year (Epoch AI) and better algorithms halve the compute needed for the same result every 8 months or so (Ho and colleagues, 2024), the effective compute behind the best model grows by very roughly 14 times a year, because 2 to the power of 12/8 is about 2.8, and 5 times 2.8 is 14. That is Mojo's arithmetic, not a measured figure, but it explains why each year's models feel like a different species.

What do the dials actually read?

All four dials are still turning, but the data dial is approaching its stop.

Compute. Epoch AI estimates frontier training compute has grown about 5 times a year since 2020, with training costs growing about 3.5 times a year. GPT-3 was trained with about 3 × 10^23 operations in 2020; GPT-4 with about 2 × 10^25 in 2023.

Epoch's largest estimate for a 2026 run, OpenAI's GPT-6 Astra, is about 9 × 10^26, and no run is yet confirmed at 10^27. Epoch estimated that training Grok 4 in 2025 used about 310 gigawatt-hours of electricity and cost about $490 million.

Algorithms. A 2024 study by Anson Ho and colleagues found that the compute needed to reach a fixed level of language-model performance halves roughly every 8 months, though bigger compute still did more of the work.

Data. This is the dial with a stop. Epoch AI estimates the stock of public human-written text at about 300 trillion tokens and expects it to be fully used for training somewhere between 2026 and 2032. We are now inside that window. Labs are turning to synthetic data, licensed data, video and the models' own reasoning traces.

Thinking time. The newest dial. In September 2024, OpenAI's o1 showed that letting a model reason longer at answer time raises accuracy smoothly: on a 2024 maths competition it scored 74% with one attempt and 93% when it could sample and re-rank 1,000 attempts, against 12% for GPT-4o. A 2024 study found that spending inference compute wisely let a small model beat one 14 times its size.

Richard Sutton's 2019 essay “The Bitter Lesson” saw this coming: across chess, Go, speech and vision, general methods that use more computation have beaten clever hand-built knowledge every time. Six years on, the lesson still stings.

What does physics say about the limits?

The software laws above run on hardware, and hardware answers to physics. Three numbers frame the whole debate.

Chips are still getting faster, just not the old way. Moore's law, the doubling of transistors roughly every two years, has slowed; Nvidia's Jensen Huang called it dead in 2022.

Yet Nvidia's data-centre chips raised dense 16-bit performance from about 125 trillion operations a second on the V100 (2017) to about 4,000 trillion on Rubin (2026), roughly 32 times in nine years, by its own specifications. Most of the gain came from specialised design and lower-precision maths, not smaller transistors.

Epoch AI measures machine-learning hardware getting about twice as cost-efficient every 2.1 years.

Your brain is embarrassingly efficient. A human brain runs on about 20 watts, according to the US National Institute of Standards and Technology. One Nvidia H100 draws about 700 watts. Open Philanthropy's best guess for the brain's raw computing power is about 10^15 operations a second, roughly what one H100 delivers. The comparison is apples to oranges, but the gap in efficiency is real and points to how much room hardware still has.

There is a floor. Landauer's principle, proposed in 1961 and confirmed experimentally in 2012, sets the minimum energy needed to erase one bit of information at room temperature: about 3 × 10^-21 joules. Today's chips use many orders of magnitude more than that. Physics is not the binding limit yet. Power stations are.

The International Energy Agency's 2025 Energy and AI report estimated data centres used about 415 terawatt-hours of electricity in 2024, about 1.5% of the world's total, and projected about 945 terawatt-hours by 2030.

Per question, the numbers are small and falling: Google measured a median text prompt to Gemini at 0.24 watt-hours in May 2025, slightly below Epoch AI's estimate of about 0.3 watt-hours for a typical ChatGPT query earlier that year, and Google reported a 33-fold drop in energy per prompt over 12 months.

The catch is Jevons' paradox: as each query gets cheaper, we ask billions more of them

The Mojo Horizon Clock: how long can AI work on its own?

In 2019, GPT-2 managed tasks of about 3 seconds. In April 2026, METR's early estimate for Anthropic's Claude Mythos Preview was more than 16 hours.

That is roughly a 21,000-fold increase in seven years. When I fit a trend line to each new record, the horizon doubled about every 6 months across 2019 to 2026, and about every 4.1 months since 2023. METR's own January 2026 update put the post-2023 doubling time at about 4.3 months, so our independent fit agrees.

Extend the post-2023 line, purely as an illustration, and it crosses a 40-hour work week around the end of 2026 and a 167-hour work month in the second half of 2027.

Now the caveats, because they matter.

1. First, 50% success is not reliability; at 80% success, horizons are much shorter.
2. Second, these are self-contained software and research tasks, not messy jobs with colleagues, customers and physical objects.
3. Third, the measuring stick itself is breaking: METR says its estimates above 16 hours are unreliable, and it stopped actively updating its leaderboard in September 2026.

When the ruler runs out before the thing being measured, that is a signal in itself.

Why do benchmarks keep falling?

Because the models are improving, and because benchmarks are targets. Both are true, and the second is why I treat every leaderboard with care.

The most striking case is ARC-AGI-3, launched by the ARC Prize Foundation on 25 March 2026 as a set of interactive puzzles with no instructions, which humans solve easily. At launch, the best AI scored 0.51%. On 3 September 2026, ARC Prize verified OpenAI's GPT-6 Astra at 62.7% on its standard test setup. (Viral claims of 98% to 99% came from a different setup that ARC Prize does not count as standard.) A benchmark designed to last years lost most of its headroom in about six months.

Elsewhere the pattern repeats. Coding on SWE-bench Verified went from 4.4% in 2023 to about 96% in provider-reported 2026 results, which means it no longer tells the best models apart. In July 2025, Google DeepMind's Gemini Deep Think solved five of six International Mathematical Olympiad problems, scoring 35 of 42 points under official grading, a gold-medal standard.

The lens here is Goodhart's law: when a measure becomes a target, it ceases to be a good measure. Labs train towards famous benchmarks; test questions leak into training data. That does not make the progress fake. It means the honest measures are the ones that look like real work, such as METR's time horizons, and the ones still hard for machines and easy for people, such as ARC-AGI-3 and Humanity's Last Exam without tools, where the best official score was 38.3%.

So how close is AGI?

Closer on screens than in the world, and closer than most people outside the field assume. But “AGI” means different things to different people, which is half the reason the debate goes in circles.

Definitions. Google DeepMind's “Levels of AGI” framework (Morris and colleagues, 2023) rates systems on two axes, how well they perform and how general they are, and classed the chatbots of the time as “emerging” AGI. Lab leaders talk about systems that can do most economically valuable work; Anthropic's Dario Amodei describes “a country of geniuses in a datacenter”. Sceptics define it as human-like learning from experience, which current models do not do.

Forecasts.

Lab leaders are more aggressive: in January 2026, Anthropic's Dario Amodei wrote that powerful AI may be one to two years away. Demis Hassabis said in 2023 that it could come within a decade. Yann LeCun, who left Meta and raised $1.03 billion for a new lab, calls large language models a dead end, and Richard Sutton argues they imitate rather than learn.

Mojo's three gates. Instead of one finish line, I track three, each with a measurable test.

Gate	Test	Status, November 2026	Mojo's read
1. Digital autonomy	AI completes a week of real software or research work at high reliability	50% horizon past 16 hours; reliable horizons much shorter; measurement breaking	Plausible within 1 to 3 years
2. Novelty	Human-level on puzzles easy for people and new to machines (ARC-AGI-3)	Best verified 62.7%; humans 100%	Closing fast; plausible within 1 to 2 years
3. The physical world	Robots doing general physical work at a competitive cost	Cost-competitive for about 0.3% of physical tasks (Anthropic, Sep 2026)	A decade or more

My verdict: the digital half of AGI is a live possibility before 2030. The physical half, the part that folds laundry, fixes boilers and nurses the sick, is much further away. Moravec's paradox still holds: what is hard for us is easy for machines, and what is easy for us is hard for them. Anyone who tells you AGI arrives on a single date is selling something.

Claim Check

Claim	Who makes it	MES	Mojo's read
Scaling laws still hold	AI labs, Epoch AI	4 Strong	Loss keeps falling with compute; the data dial is the binding risk
AI is “hitting a wall”	Sceptics, periodically	0 Didn't hold up	Horizons and hard benchmarks kept rising through 2026
Task horizons double every 4 to 7 months	METR	4 Strong	Confirmed by our own fit; measurement breaks above 16 hours
AI beat humans at maths olympiad level	Google DeepMind	4 Strong	Officially graded gold standard, July 2025
AGI will arrive by 2027	Some lab leaders	1 Claim	Depends on the definition; the digital half is plausible soon
LLMs are a dead end	Yann LeCun	1 Claim	A serious minority view; no result yet shows a hard ceiling
AI energy use per query is tiny	Google, OpenAI	3 Promising	Self-reported, but consistent with Epoch AI's independent estimate
We will run out of training data	Epoch AI	3 Promising	The 2026 to 2032 window has opened; synthetic data is the open question

The Counter-Angle

The strongest case against my reading is that I am mistaking a measuring stick for the thing itself. Time horizons and benchmarks are tests built by researchers, and models are increasingly trained to pass tests.

I take that seriously, which is why my gates require reliable work over a week and why the physical gate stays far away. But the trend has survived several changes of measuring stick over seven years. When every new test falls faster than the last, the most likely explanation is that the thing being measured is genuinely improving.

Mojo's Forecasts

1. A published estimate from METR or a comparable independent evaluator puts a model's 50% time horizon above 40 hours by 31 December 2027. Probability: 60%.
2. A score of 85% or more on ARC-AGI-3's standard test setup is verified by ARC Prize by 31 December 2027. Probability: 55%.
3. Epoch AI estimates a training run above 10^27 operations by 31 December 2027. Probability: 85%.

All three go into the Mojo Forecast Ledger and will be scored in public.

Where this is debated next

FAQ