Why AGI could be here by 2028

In recent months, the CEOs of leading AI companies have grown increasingly confident about rapid progress:

• OpenAI’s CEO shifted from saying in November 2024 he expects “the rate of progress we’ve made over the last three years continues” to declaring just one month later “we are now confident we know how to build AGI as we have traditionally understood it.”
• Anthropic’s CEO stated in January 2025: “I’m more confident than I’ve ever been that we’re close to powerful capabilities…A country of geniuses in a data centre…that’s what I think we’re quite likely to get in the next 2-3 years.”
• Even Google DeepMind’s more cautious CEO shifted from saying “as soon as 10 years” in the autumn, to “I think we’re probably three to five years away” by January.

What explains the shift? Could they be right, or is this just hype? Could we really have Artificial General Intelligence (AGI) by 2028?

In this article, I look at what’s driven recent progress, estimate how far those drivers can continue, and explain why they’re likely to continue for at least four more years.

In particular, while in 2024 progress in LLM chatbots seemed to slow, a new approach started to work: teaching the models to reason using reinforcement learning. In just a year, this let them surpass human PhDs at answering difficult scientific reasoning questions and achieve expert-level performance on one-hour coding tasks.

Extrapolating the current rate of progress suggests that, by 2028, we could reach AI models with beyond-human reasoning abilities, expert-level knowledge in every domain, and that can autonomously complete multi-week projects. Progress would likely continue from there.

No longer mere chatbots, these ‘agent’ models might soon satisfy many definitions of AGI — roughly, AI systems that match human performance at most knowledge work.¹

This means that, while the company leaders are probably overoptimistic, there’s enough evidence to take their position very seriously.

Where we draw the ‘AGI’ line is ultimately arbitrary. More importantly, these models could help to accelerate AI research and chip production itself, unlocking vastly greater numbers of more capable ‘AI workers’. In turn, sufficient automation could trigger explosive growth and 100 years of scientific progress in 10 — a transition society isn’t prepared for.

While this might sound outlandish, it’s within the range of possibilities many other experts think is possible. This article aims to give you a primer on what you need to know to understand why.

I’ve been writing about AGI since 2014. Back then, AGI arriving within five years seemed very unlikely. Today, the situation seems dramatically different. We can see the outlines of how it could work and who will build it.

In fact, the next five years seem unusually crucial. The basic drivers of AI progress — computational power and algorithmic advances — cannot continue increasing at current rates much beyond 2030. That means we either reach AI systems capable of triggering an acceleration soon, or progress will most likely slow significantly.

Either way, the next five years are when we’ll find out. Let’s see why.

I. What’s driven recent AI progress? And will it continue?

The deep learning era

In 2022, Yann LeCun, the chief AI scientist at Meta and a Turing Award winner, said:

“I take an object, I put it on the table, and I push the table. It’s completely obvious to you that the object will be pushed with the table…There’s no text in the world I believe that explains this. If you train a machine as powerful as could be…your GPT-5000, it’s never gonna learn about this.”

And, of course, if you plug this question into GPT-4 it has no idea how to answer:

Just kidding. Within a year of LeCun’s statement, here’s GPT-4.

And this isn’t the only example of experts being wrongfooted.

Before 2011, AI was famously dead.

But that totally changed when conceptual insights from the 1970s and 1980s combined with massive amounts of data and computing power to produce the deep learning paradigm.

Since then, we’ve repeatedly seen AI systems going from total incompetence to greater-than-human performance in many tasks within a couple of years.

For example, in 2022, if you asked Midjourney to draw “an otter on a plane using wifi,” this was the result:

Two years later, you could get this with Veo 2:

In 2019, GPT-2 could just about stay on topic for a couple of paragraphs. And that was considered remarkable progress.

Critics like LeCun were quick to point out that GPT-2 couldn’t reason, show common sense, exhibit understanding of the physical world, and so on. But many of these limitations were overcome within a couple of years.

Over and over again, it’s been dangerous to bet against deep learning. Today, even LeCun says he expects AGI in “several years.”²

The limitations of current systems aren’t what to focus on anyway. The more interesting question is: where this might be heading? What explains the leap from GPT-2 to GPT-4, and will we see another?

What’s coming up

At the broadest level, AI progress has been driven by:

• More computational power
• Better algorithms

Both are improving rapidly.

More specifically, we can break recent progress down into four key drivers:

1. Scaling pretraining to create a base model with basic intelligence
2. Using reinforcement learning to teach the base model to reason
3. Increasing test-time compute to increase how long the model thinks about each question
4. Building agent scaffolding so the model can complete complex tasks

In the rest of this section, I’ll explain how each of these works and try to project them forward. Delve (ahem) in, and you’ll understand the basics of how AI is being improved.

In section two I’ll use this to forecast future AI progress, and finally explain why the next five years are especially crucial.

1. Scaling pretraining to create base models with basic intelligence

Pretraining compute

People often imagine that AI progress requires huge intellectual breakthroughs, but a lot of it is more like engineering. Just do (a lot) more of the same, and the models get better.

In the leap from GPT-2 to GPT-4, the biggest driver of progress was just applying dramatically more computational power to the same techniques, especially to ‘pretraining.’

Modern AI works by using artificial neural nets, involving billions of interconnected parameters organised into layers. During pretraining (a misleading name, which simply indicates it’s the first type of training), here’s what happens:

1. Data is fed into the network (such as an image of a cat).
2. The values of the parameters convert that data into a predicted output (like a description: ‘this is a cat’).
3. The accuracy of those outputs is graded vs. reference data.
4. The model parameters are adjusted in a way that’s expected to increase accuracy.
5. This is repeated over and over, with trillions of pieces of data.

This method has been used to train all kinds of AI, but it’s been most useful when used to predict language. The data is text on the internet, and LLMs are trained to predict gaps in the text.

More computational power for training (i.e. ‘training compute’) means you can use more parameters, which lets the models learn more sophisticated and abstract patterns in the data. It also means you can use more data.

Since we entered the deep learning era, the number of calculations used to train AI models has been growing at a staggering rate — more than 4x per year.

This was driven by spending more money and using more efficient chips.³

Historically, each time training compute has increased 10x, there’s been a steady gain in performance across many tasks and benchmarks.

For example, as training compute has grown a thousandfold, AI models have steadily improved at answering diverse questions—from commonsense reasoning to understanding social situations and physics. This is demonstrated on the ‘BIG-Bench Hard’ benchmark, which features diverse questions specifically chosen to challenge LLMs:

Likewise, OpenAI created a coding model that could solve simple problems, then used 100,000 times more compute to train an improved version. As compute increased, the model correctly answered progressively more difficult questions.⁴

These test problems weren’t in the original training data, so this wasn’t merely better search through memorised problems.

This relationship between training compute and performance is called a ‘scaling law.’⁵

Papers about these laws had been published by 2020. To those following this research, GPT-4 wasn’t a surprise — it was just a continuation of a trend.

Algorithmic efficiency

Training compute has not only increased, but researchers have found far more efficient ways to use it.

Every two years, the compute needed to get the same performance across a wide range of models has decreased tenfold.

These gains also usually make the models cheaper to run. DeepSeek-V3 was promoted as a revolutionary efficiency breakthrough, but it was roughly on trend: released two years after GPT-4, it’s about 10 times more efficient.⁶

Algorithmic efficiency means that, not only is four times as much compute used on training each year, but that compute also goes three times further. The two multiply together to produce a 12 times increase in ‘effective’ compute each year.

That means the chips that were used to train GPT-4 in three months could have been used to train a model with the performance of GPT-2 about 300,000 times over.⁷

This increase in effective compute took us from a model that could just about string some paragraphs together to GPT-4 being able to do things like:

• Beat most high schoolers at college entrance exams
• Converse in natural language — in the long-forgotten past this was considered a mark of true intelligence, a la the Turing test
• Solve the Winograd schemas — a test of commonsense reasoning that in the 2010s was regarded as requiring true understanding⁸
• Create art that most people can’t distinguish from the human-produced stuff⁹

How much further can pretraining scale?

If current trends continue, then by around 2028, someone will have trained a model with 300,000 times more effective compute than GPT-4.¹⁰

That’s the same increase we saw from from GPT-2 to GPT-4, so if spent on pretraining, we could call that hypothetical model ‘GPT-6.’¹¹

After a pause in 2024, companies are already close to GPT-5-sized models, which forecasters expect to be released in 2025.

But can this trend continue all the way to GPT-6?

The CEO of Anthropic, Dario Amodei, projects GPT-6-sized models will cost about $10bn to train.¹² That’s still affordable for companies like Google, Microsoft, or Meta, which earn $50–100bn in profits annually.¹³

In fact, these companies are already building data centres big enough for such training runs¹⁴ — and that was before the $100bn+ Stargate project was announced.

Frontier AI models are also already generating over $10bn of revenue,¹⁵ and revenue has been more than quadrupling each year, so AI revenue alone will soon be enough to pay for a $10bn training run.

I’ll discuss the bottlenecks more later but the most plausible one is training data. However, the best analysis I’ve found suggests that there will be enough data to carry out a GPT-6 scale training run by 2028.

And even if this isn’t the case, it’s no longer crucial — the AI companies have discovered ways to circumvent the data bottleneck.

2. Post training of reasoning models with reinforcement learning

People often say “ChatGPT is just predicting the next word.” But that’s never been quite true.

Raw prediction of words from the internet produces outputs that are regularly crazy (as you might expect, given that it’s the internet).

GPT only became truly useful with the addition of reinforcement learning from human feedback (RLHF):

1. Outputs from the ‘base model’ are shown to human raters.
2. The raters are asked to judge which are most useful.
3. The model is adjusted to produce more outputs like the helpful ones (‘reinforcement’).

A model that has undergone RLHF isn’t just ‘predicting the next token,’ it’s been trained to predict what human raters find most helpful.

You can think of the initial LLM as providing a foundation of conceptual structure. RLHF is essential for directing that structure towards a particular useful end.

RHLF is one form of ‘post training,’ named because it happens after pretraining (though both are simply types of training).

There are many other kinds of post training enhancements, including things as simple as letting the model access a calculator or the internet. But there’s one that’s especially crucial right now: reinforcement learning to train the models to reason.

This idea is that instead of training the model to do what humans find helpful, it’s trained to correctly answer problems. Here’s the process:

1. Show the model a problem with a verifiable answer, like a math puzzle.
2. Ask it to produce a chain of reasoning to solve the problem (‘chain of thought’).¹⁶
3. If the answer is correct, adjust the model to be more like that (‘reinforcement’).¹⁷
4. Repeat.

This process teaches the LLM to construct long chains of (correct) reasoning about logical problems.

Before 2023, this didn’t seem to work. If each step of reasoning is too unreliable, then the chains quickly go wrong. And if you can’t get close to the answer, then you can’t give it any reinforcement.

But in 2024, as many were saying AI progress had stalled, this new paradigm started to take off.

Consider the GPQA Diamond benchmark — a set of scientific questions designed so that people with PhDs in the field can mostly answer them, but non-experts can’t, even with 30 minutes of access to Google. It contains questions like this:

In 2023, GPT-4 performed only slightly better than random guessing on this benchmark. It could handle the reasoning required for high school-level science problems, but couldn’t manage PhD-level reasoning.

However, in October 2024, OpenAI took the GPT-4o base model and used reinforcement learning to create o1.¹⁸

It achieved 70% accuracy — making it about equal to PhDs in each field at answering these questions.

It’s no longer tenable to claim these models are just regurgitating their training data — neither the answers nor the chains of reasoning required to produce them exist on the internet.

Most people aren’t answering PhD-level science questions in their daily life, so they simply haven’t noticed recent progress. They still think of LLMs as basic chatbots.

But o1 was just the start. At the beginning of a new paradigm, it’s possible to get gains especially quickly.

Just three months after o1, OpenAI released results from o3. It’s the second version, named ‘o3’ because ‘o2’ is a telecom company. (But please don’t ask me to explain any other part of OpenAI’s model-naming practices.)

o3 is probably o1 but with even more reinforcement learning (and another change I’ll explain shortly).

It surpassed human expert-level performance on GPQA:

Reinforcement should be most useful for problems that have verifiable answers, such as in science, math, and coding.¹⁹ o3 performs much better in all of these areas than its base model.

Most benchmarks of math questions have now been saturated — leading models can get basically every question right.

In response, Epoch AI created Frontier Math — a benchmark of insanely hard mathematical problems. The easiest 25% are similar to Olympiad-level problems. The most difficult 25% are, according to Fields Medalist Terrance Tao, “extremely challenging,” and would typically need an expert in that branch of mathematics to solve them.

Previous models, including GPT-o1, could hardly solve any of these questions.²⁰ In December 2024, OpenAI claimed that GPT-o3 could solve 25%.²¹

These results went entirely unreported in the media. On the very day of the o3 results announcement, The Wall Street Journal was running this story:

This misses the crucial point that GPT-5 is no longer necessary — a new paradigm has started, which can make even faster gains than before.

How far can scaling reasoning models continue?

In January, DeepSeek replicated many of o1’s results. Their paper revealed that even basically the simplest version of the process works, suggesting there’s a huge amount more to try.

DeepSeek-R1 also reveals its entire chain of reasoning to the user, demonstrating its sophistication and surprisingly human quality: it’ll reflect on its answers, backtrack when wrong, consider multiple hypotheses, have insights, and more.

All of this behaviour emerges out of simple reinforcement learning. OpenAI researcher Sabastian Bubeck observed:

“No tactic was given to the model. Everything is emergent. Everything is learned through reinforcement learning. This is insane.”

The compute for the reinforcement learning stage of training DeepSeek-R1 likely only cost about $1m.

If it keeps working, OpenAI, Anthropic, and Google could now spend $1bn on the same process, approximately a 1000x scale up of compute.²²

One reason it’s possible to scale up this much is that the models generate their own data.

This might sound circular, and the idea that synthetic data causes ‘model collapse‘ has been widely discussed.

But there’s nothing circular in this case. You can ask GPT-o1 to solve 100,000 math problems, then take only the cases where it got the right answer, and use them to train the next model.

Because the solutions can be quickly verified, you’ve generated more examples of genuinely good reasoning.

In fact, this data is much higher quality than what you’ll find on the internet because it contains the whole chain of reasoning and is known to be correct (not something the internet is famous for).²³

This potentially creates a flywheel:

1. Have your model solve a bunch of problems.
2. Use the solutions to train the next model.²⁴
3. The next model can solve even harder problems.
4. That generates even more solutions.
5. And so on.

If the models can already perform PhD-level reasoning, the next stage would be researcher-level reasoning, and then generating novel insights.

This likely explains the unusually optimistic statements from AI company leaders. Sam Altman’s shift in opinion coincides exactly with the o3 release in December 2024.

Although most powerful in verifiable domains, the reasoning skills developed will probably generalise at least a bit. We’ve already seen o1 improve at legal reasoning, for instance.²⁵

In other domains like business strategy or writing, it’s harder to clearly judge success, so the process takes longer, but we should expect it to work to some degree. How well this works is a crucial question going forward.

3. Increasing how long models think

If you could only think about a problem for a minute, you probably wouldn’t get far.

If you could think for a month, you’ll make a lot more progress — even though your raw intelligence isn’t higher.

LLMs used to be unable to think about a problem for more than about a minute before mistakes compounded or they drifted off topic, which really limited what they could do.

But as models have become more reliable at reasoning, they’ve become better at thinking for longer.

OpenAI showed that you can have o1 think 100 times longer than normal and get linear increases in accuracy on coding problems.

This is called using ‘test time compute’ – compute spent when the model is being run rather than trained.

If GPT-4o could usefully think for about one minute, GPT-o1 and DeepSeek-R1 seem like they can think for the equivalent of about an hour.²⁶

As reasoning models get more reliable, they will be able to think for longer and longer.

At current rates, we’ll soon have models that can think for a month — and then a year.

(It’s particularly intriguing to consider what happens if they can think indefinitely—given sufficient compute, and assuming progress is possible in principle, they could continuously improve their answers to any question.)

Using more test time compute can be used to solve problems via brute force. One technique is to try to solve a problem 10, 100, or 1000 times, and to pick the solution with the most ‘votes’. This is probably another way o3 was able to beat o1.²⁷

The immediate practical upshot of all this is you can pay more to get more advanced capabilities earlier.

Quantitatively, in 2026, I expect you’ll be able to pay 100,000 times more to get performance that would have previously only been accessible in 2028.²⁸

Most users won’t be willing to do this, but if you have a crucial engineering, scientific, or business problem, even $1m is a bargain.

In particular, AI researchers may be able to use this technique to create another flywheel for AI research. It’s a process called iterated distillation and amplification, which you can read about here. Here’s roughly how it would work:

1. Have your model think for longer to get better answers (‘amplification’).
2. Use those answers to train a new model. That model can now produce almost the same answers immediately without needing to think for longer (‘distillation’).
3. Now have the new model think for longer. It’ll be able to generate even better answers than the original.
4. Repeat.

This process is essentially how DeepMind made AlphaZero superhuman at Go within a couple of days, without any human data.

4. The next stage: building better agents

GPT-4 resembles a coworker on their first day who is smart and knowledgeable, but who only answers a question or two before leaving the company.

Unsurprisingly, that’s also only a bit useful.

But the AI companies are now turning chatbots into agents.

An AI ‘agent’ is capable of doing a long chain of tasks in pursuit of a goal.

For example, if you want to build an app, rather than asking the model for help with each step, you simply say, “Build an app that does X.” It then asks clarifying questions, builds a prototype, tests and fixes bugs, and delivers a finished product — much like a human software engineer.

Agents work by taking a reasoning model and giving it a memory and access to tools (a ‘scaffolding’):

1. You tell the reasoning module a goal, and it makes a plan to achieve it.
2. Based on that, it uses the tools to take some actions.
3. The results are fed back into the memory module.
4. The reasoning module updates the plan.
5. The loop continues until the goal is achieved (or determined not possible).

AI agents already work a bit.

SWE-bench Verified is a benchmark of real-world software engineering problems from GitHub that typically take about an hour to complete.

GPT-4 basically can’t do these problems because they involve using multiple applications.

However, when put into a simple agent scaffolding:²⁹

• GPT-4 can solve about 20%.
• Claude Sonnet 3.5 could solve 50%.
• And GPT-o3 reportedly could solve over 70%.

This means o3 is basically as good as professional software engineers at completing these discrete tasks.

On competition coding problems, it would have ranked about top 200 in the world.

Here’s how these coding agents look in action:

Now consider perhaps the world’s most important benchmark: METR’s set of difficult AI research engineering problems (‘RE Bench’).

These include problems, like fine-tuning models or predicting experimental results, that engineers tackle to improve cutting-edge AI systems. They were designed to be genuinely difficult problems that closely approximate actual AI research.

A simple agent built on GPT-o1 and Claude 3.5 Sonnet is better than human experts when given two hours.

This performance exceeded the expectations of many forecasters (and o3 hasn’t been tested yet).³⁰

AI performance increases more slowly than human performance, so human experts still surpass the AIs at around the four hour mark.

But the AI models are catching up fast.

GPT-4o was only able to do tasks which took humans about 30 minutes.³¹

METR made a broader benchmark of tasks categorised by time horizon. GPT-2 was only able to do tasks that took humans a few seconds; GPT-4 managed a few minutes; and the latest reasoning models could do tasks that took humans just under an hour.

If this trend continues to the end of 2028, AI will be able to do AI research & software engineering tasks that take several weeks as well as many human experts.

This would make them much more like a ‘digital worker’, who can be managed like a human to autonomously complete difficult projects.

AI models are also increasingly understanding their context — correctly answering questions about their own architecture, past outputs, and whether they’re being trained or deployed — another precondition for agency..

On a lighter note, while Claude 3.7 is still terrible at playing Pokemon, it’s much better than 3.5, and just a year ago, Claude 3 couldn’t play at all.

AI models today are very ‘intelligent’ at answering questions, but they can’t yet replace human workers because real jobs aren’t just a list of discrete one hour tasks –– real jobs involve figuring out what to do coordinating with a team; long, novel projects with a lot of context, etc.

However, these trends suggest this is likely to change in the next few years.

How much can agents improve?

OpenAI dubbed 2025 the “year of agents.”

• While AI agent scaffolding is still primitive, it’s a top priority for the leading labs, which should lead to more progress.
• Gains will also come from hooking up the agent scaffolding to ever more powerful reasoning models — giving the agent a better ‘planning brain.’
• Those in turn will be based on base models that have been trained on a lot more video data, which might make the agents much better at perception — a major bottleneck currently.

Once agents start working a bit, that unlocks more progress:

• Set an agent a task, like making a purchase or writing a popular tweet. Then if it succeeds, use reinforcement learning to make it more likely to succeed next time.
• In addition, each successfully completed task can be used as training data for the next generation of agents.

The world is an unending source of data, which lets the agents naturally develop a causal model of the world.³²

Any of these measures could significantly increase reliability, and as we’ve seen several times in this article, reliability improvements can suddenly unlock new capabilities:

• Even a simple task like finding and booking a hotel that meets your preferences requires tens of steps. With a 90% chance of completing each step correctly, there’s only a 10% chance of completing 20 steps correctly.

• However with 99% reliability per step, the overall chance of success leaps from 10% to 80% — the difference between not useful to very useful.

Fully functional agents would have 10–100x more applications than chatbots. Companies could start to ‘hire’ huge numbers of AI workers overseen by a small number of humans.

II. How good will AI become by 2030?

The four drivers projected forwards

Let’s recap everything we’ve covered so far. Looking ahead at the next two years, all four drivers of AI progress seem set to continue and build on each other:

1. A base model trained with 500x more effective compute than GPT-4 will be released (‘GPT-5’).
2. That model could be trained to reason with up to 100x more compute than o1 (‘o5’).
3. It’ll be able to think for the equivalent of a month per task when needed.
4. It’ll be hooked up to an improved agent scaffolding and further reinforced to be more agentic.

And that won’t be the end. The leading companies are on track to carry out $10bn training runs by 2028. This would be enough to pretrain a GPT-6-sized base model and do 100x more reinforcement learning (or some other combination).³³

In addition, advances like reasoning models appear roughly every 1–2 years, so we should project at least one more discovery like this in the next four years. And there’s some chance we might see a more fundamental advance more akin to deep learning itself.

Putting all this together, people who picture the future as ‘slightly better chatbots’ are making a mistake. Absent a major disruption,³⁶ progress is not going to plateau here.

The multi-trillion dollar question is how advanced AI will get.

Trend extrapolation of AI capabilities

One way to get a more precise answer is to extrapolate progress on benchmarks measuring AI capabilities.

Since all the drivers of progress are continuing at similar rates to the past, we can roughly extrapolate the recent rate of progress.³⁷

Here’s a summary of all the benchmarks we’ve discussed (plus a couple of others) and where we might expect them to be in 2026:

This implies that in two years we should expect AI systems that:

• Have expert-level knowledge of every field
• Can answer math and science questions as well as many professional researchers
• Are better than humans at coding
• Have general reasoning skills better than almost all humans
• Can autonomously complete many multi-day and maybe multi-week tasks on a computer
• And are still rapidly improving

The next leap might take us into beyond-human-level problem solving — the ability to answer as-yet-unsolved scientific questions independently.

What jobs would these systems be able to help with?

Many bottlenecks hinder real-world AI agent deployment, even for those that can use computers. These include regulation, reluctance to let AIs make decisions, insufficient reliability, institutional inertia, and lack of physical presence.⁴¹

Initially, powerful systems will also be expensive, and their deployment will be limited by available compute, so they will be directed only at the most valuable tasks.

This means most of the economy will probably continue pretty much as normal for a while.

You’ll still consult human doctors (even if they use AI tools), get coffee from human baristas, and hire human plumbers.

However, there are a few crucial areas where, despite these bottlenecks, these systems could be rapidly deployed with significant consequences.

Software engineering

This is where AI is most useful today. Google has said about 25% of their new code is written by AIs. Y Combinator startups say it’s 95%, and that they’re growing several times faster than before.

If coding becomes 10x cheaper, we’ll use far more of it. Maybe fairly soon, we’ll see billion-dollar software startups with a small number of human employees and hundreds of AI agents. Several AI startups have already become the fastest-growing companies of all time.

This could produce hundreds of billions of dollars of economic value pretty quickly — sufficient to fund continued AI scaling.

Scientific research

The creators of AlphaFold already won the Nobel Prize for designing an AI that solves protein folding.

A recent study found that an AI tool made top materials science researchers 80% faster at finding novel materials, and I expect many more results like this once scientists have adapted AI to solve specific problems, for instance by training on genetic or cosmological data.

Future models might be able to have genuinely novel insights simply by someone asking them. But, even if not, a lot of science is amenable to brute force. In particular, in any domain that’s mainly virtual but has verifiable answers — such as mathematics, economic modeling, theoretical physics, or computer science — research could be accelerated by generating thousands of ideas and then verifying which ones work.

Even an experimental field like biology is also bottlenecked by things like programming and data analysis, constraints that could be substantially alleviated.

A single invention like nuclear weapons can change the course of history, so the impact of any speed up here could be dramatic.

AI research

A field that’s especially amenable to acceleration is AI research itself. Besides being fully virtual, it’s the field that AI researchers understand best, have huge incentives to automate, and face no barriers to deploying AI.

Initially, this will look like researchers using ‘intern-level’ AI agents to unblock them on specific tasks or software engineering capacity (which is a major bottleneck), or even help brainstorm ideas.

Later, it could look like having the models read all the literature, generate thousands of ideas to improve the algorithms, and automatically test them in small-scale experiments.

An AI model has already produced an AI research paper that was accepted to a conference.

Given all this, it’s plausible we’ll have AI agents doing AI research before people have figured out all the kinks that enable AI to do most remote work jobs.

Broad economic application of AI is therefore not necessarily a good way to gauge AI progress — it may follow explosively after AI capabilities have already advanced substantially.

What’s the case against impressive AI progress by 2030?

Here’s the strongest case against in my mind.

First, concede that AI will likely become superhuman at clearly defined, discrete tasks, which means we’ll see continued rapid progress on benchmarks.

But argue they’ll remain poor at ill-defined, high-context, and long-time-horizon tasks.

That’s because these kinds of tasks don’t have clearly verifiable answers, and so they can’t be trained with reinforcement learning, and they’re poorly represented in the training data.

Second, argue that most knowledge jobs consist significantly of these long-horizon, messy, high-context tasks.

For example, software engineers spend a lot of their time figuring out what to build, coordinating with others, and understanding massive code bases rather than knocking off a list of well-defined tasks. Even if their productivity at coding increases 10x, if coding is only 50% of their work, their overall productivity only roughly doubles.

In this scenario, we’ll have extremely smart and knowledgeable AI assistants, and perhaps an acceleration in some limited virtual domains (perhaps like mathematics research), but they’ll remain tools, and humans will remain the main economic bottleneck.

Human AI researchers will see their productivity increase but not enough to start a positive feedback loop.

These limits, combined with problems finding a business model and the other barriers to deploying AI, will mean the models won’t create enough revenue to justify training runs over $10bn. That’ll mean progress slows massively after about 2028.⁴² Once progress slows, the profit margins on frontier models collapse, making it even harder to pay for more training.

The primary counterargument is the earlier graph from METR: models are demonstrably improving at acting over longer horizons, which requires deeper contextual understanding and handling of more abstract, complex tasks. Projecting this trend forward suggests highly autonomous models within four years.

Agency over longer horizons could be achieved via many incremental advances I’ve sketched,⁴³ but it’s also possible we’ll see a more fundamental innovation arise — the human brain itself proves such capabilities are possible.

This is perhaps the central question of AI forecasting: will the horizon over which AIs can act plateau or continue to improve?

Here are a few other ways progress could be slower:

• Pretraining will have big diminishing returns, so GPT-5 and GPT-6 won’t be much more useful than the models we have today (perhaps due to diminishing data quality).
• There could be a long distance between expert-level question answering and the ability to have novel insights (even considering the possibility of brute force problem solving).
• AI will continue to be bad at visual perception, limiting its ability to use a computer (see Moravec’s paradox).
• Benchmarks might be misleading due to issues with data contamination , and the difficulty of capturing messy tasks.
• An economic crisis, Taiwan conflict, other disaster, or massive regulatory crackdown could delay investment by several years.
• There are other unforeseen bottlenecks (cf planning fallacy).

For deeper exploration of the skeptical view, see “Are we on the brink of AGI?” by Steve Newman, “The promise of reasoning models” by Matthew Barnnett, and “A bear case: My predictions regarding AI progress,” by Thane Ruthenis.

When do the ‘experts’ expect AGI to arrive?

I’ve made some big claims. As a non-expert, it would be great if there were experts who could tell us what to think.

Unfortunately, there aren’t. There are only different groups, with different drawbacks.

I’ve reviewed the views of these different groups of experts in a separate article. In short, I argue AGI before 2030 is within the scope of what AI experts and forecasters think is plausible. Many also think it’ll take much longer. But if 30% of experts think a plane will explode, and the other 70% think it’ll be fine, as non-experts we shouldn’t conclude it’ll definitely be fine.

III. Why the next 5 years are crucial

It’s natural to assume that since we don’t know when AGI will emerge, it might arrive soon, in the 2030s, the 2040s, and so on.

Although it’s a common perspective, I’m not sure it’s right.

The core drivers of AI progress are more compute and better algorithms.

More powerful AI is most likely to be discovered when the compute and labour used to improve AIs is growing most dramatically.

Right now, the total compute available for training and running AI is growing 3x per year,⁴⁴ and the workforce is growing rapidly too.

This means that each year, the number of AI models that can be run increases three times. In addition, three times more compute can be used for training that uses better algorithms, which means they get more capable as well as more numerous.

Earlier, I argued these trends can continue until 2028. But now I’ll show it most likely runs into bottlenecks shortly thereafter.

Bottlenecks around 2030

First, money:

• Google, Microsoft, Meta etc. are spending tens of billions of dollars to build clusters that could train a GPT-6-sized model in 2028.
• Another 10x scale up would require hundreds of billions of investment. That’s more than their current annual profits and would be similar to another Apollo Program or Manhattan Project in scale.⁴⁵
• GPT-8 would require trillions. AI would need to become a top military priority or already be generating trillions of dollars of revenue (which would probably already be AGI).

Even if the money is available there will also be bottlenecks such as:

• Power: Current levels of AI chip sales, if sustained, mean that AI chips will use 4%+ of US electricity by 2028⁴⁶, but another 10x scale up would be 40%+. This is possible, but it would require building a lot of power plants.
• Chip production: Taiwan Semiconductor Manufacturing Company (TSMC) manufactures all of the world’s leading AI chips, but its most advanced capacity is still mostly used for mobile phones. That means TSMC can comfortably produce 5x more AI chips than it does now. However, reaching 50x would be a huge challenge. ⁴⁷
• ‘Latency limitations‘ could also prevent training runs as large as GPT-7.⁴⁸

So most likely, the rate of growth in compute slows around 2028–2032.

Algorithmic progress is also very rapid right now, but as each discovery gets made, the next one becomes harder and harder. Maintaining a constant rate of progress requires an exponentially growing research workforce.

In 2021, OpenAI had about 300 employees; today, it has about 3,000. Anthropic and DeepMind have also grown more than 3x, and new companies have entered. The number of ML papers produced per year has roughly doubled every two years.⁴⁹

It’s hard to know exactly how to define the workforce of people who are truly advancing capabilities (vs selling the product or doing other ML research). But if the workforce needs to double every 1–3 years, that can only last so long before the talent pool runs out.⁵⁰

My read is that growth can easily continue to the end of the decade but will probably start to slow in the early 2030s (unless AI has become good enough to substitute for AI researchers by then).

Algorithmic progress also depends on increasing compute, which enables more experiments. With sufficient compute, researchers can even conduct brute force searches for optimal algorithms. Thus, slowing compute growth will correspondingly slow algorithmic progress.

If compute and algorithmic efficiency increase by just 50% annually rather than 3x, a leap equivalent to the leap from GPT-3 to GPT-4 would take over 14 years instead of 2.5.

It also reduces the probability of discovering a new AI paradigm.

So there’s a race:

• Can AI models improve enough to generate enough revenue to pay for their next round of training before it’s no longer affordable?
• Can the models start to contribute to algorithmic research before we run out of human researchers thrown at the problem?

The moment of truth will be around 2028–2032.

Either progress slows, or AI itself overcomes these bottlenecks, allowing progress to continue or even accelerate.

Two potential futures for AI

If AI capable of contributing to AI research isn’t achieved before 2028–2032, the annual probability of its discovery decreases substantially.

Progress won’t suddenly halt — it’ll slow more gradually. Here are some illustrative estimates of probability of reaching AGI:

Very roughly, we can plan for two scenarios:⁵¹

1. Either we hit transformative AI before 2030: AI progress continues or even accelerates, and we probably enter a period of explosive change.
2. Or progress will slow: AI models will get much better at clearly defined tasks, but won’t be able to do the ill-defined, long horizon work required to unlock a new growth regime. We’ll see a lot of AI automation, but otherwise the world will look more like ‘normal’ the coming decades.

We’ll know a lot more about which scenario we’re in within the next few years.

I roughly think of these scenarios as 50:50 — though I can vary between 30% and 80% depending on the day.

The numbers also depend on your definition. I prefer to focus on forecasting AI that can meaningfully contribute to AI research.⁵² AGI in the sense of a model that can do almost all remote work tasks cheaper than a human may well takee longer due to a long tail of bottlenecks. On the other hand, AGI in the sense of ‘better than almost all humans at reasoning when given an hour’ seems to be basically here already.

Conclusion

So what about the opening claim of AGI by 2028?

Whatever the exact definition, significant evidence supports this possibility — we may only need to sustain current trends for a few more years.

And if not by 2028, it could easily happen by 2032.

It seems clearly overconfident to think the probability before 2032 is below 10%.

Given the massive implications and serious risks, there’s enough evidence to take this possibility extremely seriously.

Today’s situation feels like February 2020 just before COVID lockdowns: a clear trend suggested imminent, massive change, yet most people continued normal lives.

In an upcoming article, I’ll argue that AGI automating much of remote work and doubling the economy could be a conservative outcome.

If AI can do AI research, the gap between AGI and ‘superintelligence’ could be short.

This could trigger a massive research workforce expansion, potentially delivering a century’s worth of scientific progress in under a decade. Robotics, bioengineering, and space settlement could all arrive far sooner than commonly anticipated.

The next five years would be the start of one of the most pivotal periods in history.

Further reading

• The best case for rapid near-term AI progress is Chapter 1 of Situational Awareness of Leopold Aschenbrenner.

• Reinforcement learning works! a podcast by Nathan Labenz is a good explanation of reasoning models.

• Tomas Pueyo has a more accessible intro covering similar material to this article: The most important time in history is now.

• Epoch AI has a review of all the different ways of forecasting AI. All of them are consistent with AGI arriving before 2030, though some give lower probabilities. (Several of the estimates have also shortened after it was written.)

• Epoch AI also has many great datasets that underpin this post. See their key trends page for an overview. Also see their article Can scaling continue through 2030.

• An approach to AI forecasting that was popular several years ago is to estimate the compute used to train the human brain, and then estimate when leading AI models might surpass that point (tldr: we might be there around now). See Forecasting transformative AI: the ‘biological anchors’ method in a nutshell by Holden Karnofsky for an introduction.

• When do experts expect AGI? Also see Through a glass darkly by Scott Alexander, which is an exploration of what can be learned from expert forecasts on AI.

• Here are some of the best articles I’ve seen making the case against transformative AI progress in the next few years: Are we on the brink of AGI? by Steve Newman, The promise of reasoning models by Matthew Barnnett, and A bear case: My predictions regarding AI progress, by Thane Ruthenis.