Start / LessWrong (30+ Karma) / The case for agi by 2030 by benjamin_todd

LessWrong (30+ Karma)

“The case for AGI by 2030” by Benjamin_Todd

60 min • 10 april 2025

I spent a couple of weeks writing this new introduction to AI timelines. Posting here in case useful to share and for feedback. The aim is to be up-to-date, more balanced than Situational Awareness, but still relatively accessible to new people.

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

OpenAI's Sam Altman: Shifted from saying in November “the rate of progress continues” to declaring in January “we are now confident we know how to build AGI”
Anthropic's Dario Amodei: Stated in January “I’m more confident than I’ve ever been that we’re close to powerful capabilities… in the next 2-3 years”
Google DeepMind's Demis Hassabis: Changed from “as soon as 10 years” in autumn to “probably three to five years away” by January.

What explains the shift? Is it just hype? Or could we really have Artificial General [...]

---

Outline:

(04:12) In a nutshell

(05:54) I. What's driven recent AI progress? And will it continue?

(06:00) The deep learning era

(08:33) What's coming up

(09:35) 1. Scaling pretraining to create base models with basic intelligence

(09:42) Pretraining compute

(13:14) Algorithmic efficiency

(15:08) How much further can pretraining scale?

(16:58) 2. Post training of reasoning models with reinforcement learning

(22:50) How far can scaling reasoning models continue?

(26:09) 3. Increasing how long models think

(29:04) 4. The next stage: building better agents

(34:59) How far can the trend of improving agents continue?

(37:15) II. How good will AI become by 2030?

(37:20) The four drivers projected forwards

(39:01) Trend extrapolation of AI capabilities

(40:19) What jobs would these systems be able to help with?

(41:11) Software engineering

(42:34) Scientific research

(43:45) AI research

(44:57) What's the case against impressive AI progress by 2030?

(49:39) When do the 'experts' expect AGI to arrive?

(51:04) III. Why the next 5 years are crucial

(52:07) Bottlenecks around 2030

(55:49) Two potential futures for AI

(57:52) Conclusion

(59:08) Use your career to tackle this issue

(59:32) Further reading

The original text contained 47 footnotes which were omitted from this narration.

---

First published:
April 9th, 2025

Source:
https://www.lesswrong.com/posts/NkwHxQ67MMXNqRnsR/the-case-for-agi-by-2030

---

Narrated by TYPE III AUDIO.

---

Images from the article:

On <span>this set</span> of software engineering & computer use tasks, in 2020 AI was only able to do tasks that would typically take a human expert a couple of seconds. By 2024, that had risen to almost an hour. If the trend continues, by 2028 it’ll reach several weeks.

Midjourney’s attempts at depicting “an otter on a plane using wifi” in 2022.

Since the start of the deep learning era, the amount of computational power (measured with ‘FLOP’) used to train leading AI models has increased more than four times each year.

LLM performance on a challenging benchmark (BIG-Bench Hard) improves as training compute increases 1000x.

The computing power of the best chips has grown about 35% per year since the beginnings of the industry, known as Moore’s Law. However, the computing power applied to AI has been growing far faster, at over 4-times per year.

AI models require 10 times less compute to reach the same accuracy in recognising images every two years (based on the ImageNet benchmark).

A comparison of GPT-4 and GPT-3.5’s percentile scores against human test takers on standardised exams.

<span>Epoch AI estimates</span> the revenues of frontier AI companies have been growing over 3x per year.

An example of the kinds of PhD-level scientific problems on the new GPQA Diamond benchmark. I did a masters-level course in theoretical physics at university, and I have no clue.

AI models couldn’t answer these difficult scientific reasoning questions in 2023 better than chance, but by the end of 2024, they could beat PhDs in the field.

On the same day that o3 demonstrated remarkable performance on extremely difficult math problems, The Wall Street Journal was reporting about delays to GPT-5 on its homepage.

Accuracy on coding problems increases as the amount of time the model has to ‘think’ scales up.

To get an idea of how this looks, see this demo of the coding agent Devin.

When given two hours to complete difficult AI research engineering problems, models outperform humans. Given more than two hours, humans still considerably outperform AI models, with the advantage increasing as the time budget gets larger. Source: Wijk, Hjalmar, et al. RE-Bench: <span>Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts.</span>

Kwa, Thomas, et al. “Measuring AI Ability to Complete Long Tasks.”<span> arxiv.org/abs/2503.14499.</span>

ChatGPT screenshot showing question and answer about smartphone movement on table.

Graph showing AI task completion times from 2020-2025, with exponential growth.

Chat interface showing a physics-related question about object movement on table.

Table comparing AI progress drivers between 2019-2023 and 2024-2028 time periods.

Mathematical equation solution showing an AI model reconsidering its approach mid-calculation.

The image shows a mathematical problem solving process with an interruption marked in red text where the AI system pauses to reconsider its solution strategy for solving a complex equation involving square roots.

When OpenAI launched, it was the fastest growing startup of all time in terms of revenue. Since then, several other AI companies have taken the record, most recently Cursor (a coding agent). Docusign, a typical successful SaaS startup before the AI wave, is shown on the chart as a comparison. <span>Source.</span>

In four years, the mean estimate on Metaculus for when AGI will be developed has plummeted from 50 years to five years. There are problems with the definition used, but the graph reflects a broader pattern of declining estimates.

Graph showing probability of AGI development from 2025-2040, peaking 2027.

Table showing AI benchmark performance predictions from 2022 through 2026 for various tests.

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Kategorier

Filosofi Poddar Samhälle och kultur Teknologi

Förekommer på

00:00 -00:00