What will the world look like when AI transforms our lives? Epoch AI’s research team is focused on the future. The team’s mission is to understand the future of AI.
There is disagreement between them about the timeline for AGI. The roots of the disagreements between two Epoch AI researchers who have different timelines for AGI are examined by both. Ege and Matthew discuss their views and the reasons why they differ by two to three factors for important transformative milestones.
They discuss:
The median timeframes of their specific milestones, such as sustained GDP growth above 5%, highlight the differences between conservative and optimistic AI forecasts.
The AI transformation that will be primarily driven by superhuman researchers (“a country of geniuses”), or the widespread automation of cognitive and physical everyday tasks.
Moravec’s Paradox Today: How AI is still challenged by practical skills such as agency and commonsense despite advances in reasoning.
It is the interplay between hardware scaling, algorithms, data availability (especially agentic tasks) and the persistence of transfer learning.
The pitfalls of conventional academic AI and the reasons why it might be wrong.
A World with AGI : Moving away from the totalizing narratives of “a single AGI” and “utopia or doom”, to include economic forces, agents decentralized, and co-evolution AI and society.
This will likely not be the case. While we can expect to see an increase in output and productivity per worker due to increased returns on scale, this will most likely not be from R&D. AI’s primary impact on the economy will come from its ability to automate large amounts of labor.
AI benchmarks don’t reflect economic impact.
The goal was rarely to measure the impact of AI on real world tasks, as this would have required substantial resources. Benchmarks are mainly optimized to compare the capabilities of different models for tasks which are “just reachable”. Researchers are more motivated to create benchmarks for AI systems that reflect the real economic impact of AI. As AI becomes increasingly powerful and is used by more industries, they have a stronger incentive to do so.
AI Progress will soon accelerate.
GPT-4, released in March 2023, is notable because it represents a scale-up of 10x over previous models. We haven’t seen another scaling up of this magnitude since then: All currently available frontiers models (except Grok 3) were trained with a budget that was similar or lower than GPT-4. Dario Amodei CEO of Anthropic confirmed recently that Claude 3.5 Sonnet’s training cost was close to tens-of-millions-of dollars. Based on the trends of GPU performance and price, this would mean that it was trained with around two times the computing power that GPT-4 required.
Frontier Labs’ next-generation models, Grok 3, will break this barrier. They represent a scale-up of more than one order of magnitude over GPT-4 and possibly two orders of magnitude in terms of reasoning RL. According to past scaling experience, this should lead to an increase in performance at least equal to the leap from GPT-3.5 up to GPT-4.
Expectations for the New Models
The macro-scale [EpochAI] Expect the new models will be trained using around 100K H100s. This may include 8-bit quantization and a training computation of 3e26 FLOP.
Note: xAI will expand its datacenter to 400k GPUs (200K B200, 200K H100/H200) or approximately 12X compute the 100K H100. It would have been 3.6e27 FLOP. It will take 2-3 months to complete. By 2025, or in Q1 of 2026 xAI should have about 1M graphics cards. These will likely be Dojo 2, B200 or B300.
When the 100K H100s model is first released it will have a parameter count that’s an order of magnitude larger than GPT-4o. This means we can expect to see API tokens costing 2-3x more and a 2x slower decoding time for short contexts. However, these issues should improve by the end of the year as inference clusters switch to newer hardware.
Overall, the industry will grow at an average rate of 2% per year. The median Epoch AI forecast of OpenAI revenue in 2025 is $12B, which aligns with OpenAI’s forecasts. Around $4.5B comes from the final quarter. The growth in revenue will be driven primarily by improving agency and maintaining consistency over long contexts. We’ve already seen some indications of this with products such as Deep Research and Operator. They appear more important than complex reasoning gains, but are less easily measured today.
The impact of a 10x scale up on the model’s quality is roughly equivalent to 17 months algorithmic progression if we use the highest end of estimates for algorithmic advancement rates. There will also be another 10 months’ worth of algorithmic advancements until 2025. This means that the progress made until 2025 is equivalent to more than two years’ algorithmic development. More than we have seen in the last few years since GPT-4, when compute expenses were relatively flat.
The progress in pretrained model performance is greater than what would be expected from simply increasing computing power. It occurs at a rate equivalent to doubling computation power every five to fourteen months.
Timestamps
Watch the Preview at 0:00.
0:01:08 – Contrasting AGI Timelines
0:08 – Updating beliefs as capabilities advance
0:17.07 – Moravec’s Paradox and The Agency Challenge
0:32.40 – AGI missing capabilities
0:47:43 – Beating benchmarks vs Being Useful
AI excelling at some tasks while struggling with others
1:07:33 Economic Impact of AI and the Internet
Datacenters: Widespread automation beats genius
The Storytelling Effect on AI Expectations
The Impact of AGI on Culture
2:10:46 – Beyond Utopia or Extinction
AI Impact on Wages and Labour
2:27.49 – Why Better Preservation of Information Speeds Up Change
2:39.32 – The Markets and Cultural Priorities
The Challenges of Defining what We Want To Preserve
3:06;47 – Attitudes to Risk in AI Decision Making
3:12:50 – Historical Lessons on AI Coexistence
The Warning Sign at 3:21.45
3:30:20 Revisiting core assumptions in AI Alignment
3:49.46 – Simple models in complex domains
Brian Wang, a Futurist and Science Blogger with over 1,000,000 monthly readers is one of the most popular Futurists. Nextbigfuture.com, his blog is the #1 Science News Blog. The blog covers a wide range of disruptive technologies and trends, including Space, Robotics and Artificial Intelligence. It also includes Medicine, Antiaging Biotechnology and Nanotechnology.
He is known for his ability to identify cutting-edge technologies. He currently serves as a co-founder of a company and a fundraiser for early stage companies with high potential. He is Head of Research for Allocations for Deep Technology Investments and an Angel investor at Space Angels.
He is a frequent corporate speaker. In addition, he’s been a TEDx Speaker, a Singularity University Speaker, and a guest on numerous radio interviews and podcasts. He accepts public speaking engagements and advisory roles.
