Glimpsing the Coming of AGI

Altman's three observations, AI capability dimensions - language, reasoning & tool use - and an updated AGI Timeline. AGI in 2027!

Figure 1. The cosmic eye, glimpsing the AGI future.

AGI comes into View

Systems that start to point to AGI* are coming into view, and so we think it’s important to understand the moment we are in. – Sam Altman

Can you feel the AGI?

The recent AI reasoning model releases of DeepSeek-R1 and o3-mini have raised expectations and accelerated AGI timelines. The latest AI models are achieving human-level performance a range of benchmarks, saturating some benchmarks and forcing us to utilize harder benchmarks to evaluate AI.

The latest AI reasoning models, including DeepSeek R1, are achieving expert level performance on GPQA Diamond, with scores above 70% on this test of PhD level science questions. The o3 model reached an ELO rating of 2700 on Codeforces, scoring at a level of the 175^th best programmer in the world.

Figure 2. AI model performance on GPQA has in the past two years gone from random guessing to expert human level.

As AI saturates old benchmarks and even tackles the latest and hardest ones, such as ARC-AGI and Humanity’s Last Exam, we are running out of challenges to declare that AI isn’t yet AGI, human-level general intelligence.

Sam’s Three Observations

Sam Altman’s latest blog post Three Observations discusses the coming of AGI, giving his high-level sense of when and what we can expect from AI, as it reaches AGI and beyond. His message is to expect a massive increase in human capabilities thanks to AI, soon:

In a decade, perhaps everyone on earth will be capable of accomplishing more than the most impactful person can today.

His three observations confirm the pace of AI improvement and the value it brings:

1. The intelligence of an AI model roughly equals the log of the resources used to train and run it.

2. The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use.

3. The socioeconomic value of linearly increasing intelligence is super-exponential in nature. A consequence of this is that we see no reason for exponentially increasing investment to stop in the near future.

The first observation is nothing more than a restatement of well-know scaling laws, that AI model capabilities increase by the log of the compute, data, and/or parameters. With the rise of RL models, there’s even another scaling dimension: Inference-time compute, which can drive further AI reasoning performance gains.

Logarithmic scaling might imply slow gains, but we have scaled computer effort, data inputs, and AI model sizes exponentially, increasing them by many orders of magnitude over the last decade. It’s enough to enable rapid continued advance in AI capabilities.

Figure 3. The compute (FLOPs) applied to train AI models is increasing 4.6x per year, or an order of magnitude every 18 months or so. Chart courtesy of Epoch AI.

The second observation is the 10x drop in cost of AI per year. This is a remarkable feature of the AI revolution and something we first noted in AI Changes Everything almost two years ago. While better AI hardware (Moore’s Law) is a factor, it’s mostly due to constant innovation in AI algorithms and training. AI model efficiency improvement drives AI cost reductions.

Lastly, he claims the “value of linearly increasing intelligence is super-exponential.” Certainly, more intelligence unlocks AI’s ability to do higher value activities. For example, writing a whole computer program correctly is far more valuable than writing one line of code. However, the term ‘super-exponential’ seems like an attempt to over-specify a qualitative intuition.

Defining AGI

Any discussion of when and how we will get to AGI needs to nail down its definition. In his blog post, Sam asserted:

AGI is a weakly defined term, but generally speaking we mean it to be a system that can tackle increasingly complex problems, at human level, in many fields.

This general definition is highly subjective. What problems? What fields? What defines “human level”? One challenge in defining AGI is defining human abilities and tasks.

I have previously written about a fairly robust definition of AGI, called the Heinlein Test, that attempts to go beyond the Turing Test and more robustly define the general skills of a useful human. It is so robust (“set a bone … build a wall”) that we’ll have to wait for AI embodied in robots to meet that test.

Figure. Robert Heinlein on what a human being should be able to do. When we get AI to do all that, we’ll have met our match.

In Sam Altman’s world, it’s enough for the AI to work in the virtual world. He says:

Still, imagine it as a real-but-relatively-junior virtual coworker. Now imagine 1,000 of them. Or 1 million of them.

Achieving AGI is getting an AI virtual worker that does tasks for you similar to someone on Upwork. Come to think of it, I had a virtual team that frustrated me by doing tasks upon request but not getting at what I really wanted; I never figured out the right prompt engineering for that team.

The Unique Elements of Human Capability

“In some sense, AGI is just another tool in this ever-taller scaffolding of human progress we are building together.” – Sam Altman

Another crack at understanding AGI is to consider what is special about humans. What makes humans unique and different from all other animals? Three things come to mind:

• Language and communication

• Abstract reasoning

• Tool development and use

Building on top of these three was a fourth and perhaps the most unique factor: Complex social groups.¹ Communicating through language and using reasoning, early man developed the ability to work with others to hunt, teach skills to others, and in the process developing larger and more complex and capable social groups.

All four of these factors have evolved into more complex systems and technologies: Better communication via writing and the printing press; better reasoning developed through logic, mathematics, and science; tools evolved from axes to the plough to engines and computers; and social groups evolved from hunter-gatherer bands into tribes, then to complex societies and advanced civilizations.

AGI in Human Terms

I mention this because it serves as a template for AI development towards AGI. Consider that AI development has parallels with these three fundamental factors of human development:

• Language: AI understanding and generation of language has advanced with the rise of powerful LLMs that can now translate across many languages, understand complex prompts, and generate well-crafted poems and essays, and answer detailed and complex queries correctly.

• Abstract reasoning: Since early AI days, a key focus was on getting AI to reason. Now, Alpha Geometry 2 just showed that AI have become as good as the best humans at reasoning, and the releases of o1, o3-mini, and DeepSeek-R1 give us a glimpse of continued scaling-based improvement in this area.

• Tool development and use: AI-based robots are not yet able to use tools like screwdrivers and drills like humans, but AI systems are learning to connect to virtual tool capabilities, with Anthropic’s Computer Use, OpenAI’s Operator and many other AI agents that use tools.

In short, to get to AGI, AI systems need to do three things as well as humans: Communicate with language, reason, and use tools. This serves both as a definition of AGI and a roadmap of how to get there.

Three turns – in language, reasoning, and tool use - to AGI

What’s missing to get to AGI? In terms of the features and capabilities needed to get to AGI, the three features of language, reasoning, and tool use cover much.

Two years ago, I speculated that it would take “3 turns of the crank” to get from GPT-4-level AI to AGI. I had envisioned it mainly in terms of scaling LLMs while understanding that scaling isn’t all we need.

It is indeed 3 turns of the crank to get to AGI; those turns expand the 3 dimensions of capabilities:

• Language: Moving from text-only chatbots to multi-modal (audio, video, image, and text) AI systems that can understand and communicate across all modalities and multiple languages.

• Reasoning: Moving from basic “System 1” linear thinking LLMs to System-2 capable AI reasoning models opens up more complex reasoning, planning, and decision-making abilities.

• Tool use: AI models being able to call upon external tools helps them find knowledge and data (via search and knowledge retrieval tools), analyze (via code writing and execution), and check work (via verifiers).

We have already turned the crank on language capabilities with the rise of multi-modal AI models. There is further to go, but the latest multi-modal multi-lingual Gemini 2.0 points the way to full multi-model AI.

The root of human ability to do all things – communicate, invent tools, and more – has been reasoning. For AI, reasoning is the ultimate driver of high-level AI performance. The rise of AI reasoning models is the most critical ‘turn of the crank’ that has opened up the path to AGI. Further scaling beyond the o3 model will enable extreme abilities to reason over tough problems.

For humans, tools led to technologies that became force and ability multipliers. For AI, tool use will be a similar multiplier. AI will solve many difficult problems not directly but by coding then executing solvers for planning, logic, or data analysis. Tool use will also extend AI model controls to interact with the world, creating impact far beyond a chatbot.

Clear path to AGI

This perspective on AI capability (language, reasoning, tool use) explains why Sam is speaking of OpenAI having a ‘clear path’ to AGI. OpenAI’s 3 latest releases each take a major step forward in these areas:

• Deep Research is about completing far more complex research and reporting tasks, advancing reasoning, language communication, and AI agent planning. By leveraging web search, it shows the power of tool use as well.

• As the most powerful AI reasoning model yet released, o3-mini is a clear advance in AI reasoning.

• OpenAI’s Operator, based on their Computer-using Agent, is still an early prototype, but it gives a glimpse of what AI can do on tasks such as shopping, research, accounting, etc. by having AI control actions through a browser.

It’s not just that they have already achieved better AI reasoning, they have a path to further scale improvements in it. This is in parallel with possible further scaling improvements in LLMs via pre-training.

Conclusion - AGI Timelines

To get to AGI, AI needs to advance on the three major dimensions of AI capability - language, reasoning, tool use – towards human capabilities.

The latest AI model releases show progress on all fronts. AI models are saturating some benchmarks and achieving human-level performance on others. While there may be limits in scaling pre-training of AI models due to limits on quality data, we have another dimension – test-time compute scaling – for improving AI performance.

The road to AGI is open:

• One turn of the crank to build AI models that excel at tool use and virtual system control.

• Further scaling of AI reasoning (RL effort), 1-2 orders-of-magnitude increase beyond o1.

• Further scaling of pre-training, 1-2 orders-of-magnitude compute effort beyond GPT-4.

AI progress is rapid and is not slowing down, as more AI labs pursue development and progress on various fronts is made in parallel. The progress in the past two years from GPT-4 to multi-modal frontier AI models with reasoning capabilities is at least half-way to where AI models need to be for AGI.

Given all this, it’s a reasonable conclusion that we will reach AGI in the next two years, by 2027.

AI progress won’t stop there; it will accelerate into ASI. The end game? Intelligence abundance.

Anyone in 2035 should be able to marshal the intellectual capacity equivalent to everyone in 2025; everyone should have access to unlimited genius to direct however they can imagine. – Sam Altman

Footnotes

1. One can argue that humans are not unique in these traits: Whales and apes communicate with each other; apes can reason; various animals will use tools. However, animals only possess very primitive capabilities on these fronts. Human capability is both quantitatively and qualitatively different.