Predictionsby GlyphTraining runs of about 2e29 FLOP will be feasible by 2030, a GPT-2 to GPT-4 sized jump beyond GPT-4.
Epoch AI
We find that training runs of 2e29 FLOP will likely be feasible by the end of this decade.
- Who
- Epoch AI
- Made
- August 2024 2 years ago
- Target
- Dec 31, 2030 in 4 years (stated as "by the end of this decade")
- Status
- Openin 4 years
- Source
- Can AI scaling continue through 2030? (Epoch AI) Primary: their own words
- Topics
- Compute
More from Epoch AI
By 2030 AI will autonomously fix issues, implement features and solve hard, well-defined scientific programming problems.Existing biology protocol question-answering benchmarks will be solved by 2030.Few if any drugs approved for sale by 2030 will have benefited from today's AI tools.Clusters used to train frontier AI will cost over $100 billion by 2030.
Related predictions
By the end of 2030 each leading lab will have compute to run hundreds of millions of human-level AIs.
Noam Brown, OpenAIAGIAt least two countries will turn data-center ratepayer protection into tariffs or equivalent obligations in 2027.
ForresterThere will be no meaningful relief in memory supply for enterprise storage before 2028.
IDCComputeNvidia revenue for the third quarter of fiscal 2027 will be about $108 billion, plus or minus 2%.
NvidiaMarketsChina will have 30 gigawatts of AI compute or less in 2028.
Dylan Patel, SemiAnalysisBy the end of 2027 half of the world's incremental new compute will go to Anthropic and OpenAI.Dylan Patel, SemiAnalysis