Major AI companies like Google and OpenAI have suggested that AI advancement is slowing down, with OpenAI co-founder Ilya Sutskever saying the 2010s were the "age of scaling" and now the industry is in the "age of wonder and discovery" again.
However, Anthropic CEO Dario Amodei sees a path to AGI (artificial general intelligence) by 2026 or 2027, a much shorter timeline than the slowdown narrative.
Reasons for Differing Interpretations
The slowdown may be related to diminishing returns from scaling up training data and models. The quality of data and how it's used for inference may matter more now.
But the ability to scale up inference compute, as seen with OpenAI's GPT-3 and upcoming models, could allow for continued rapid progress.
Amodei believes that even with potential hurdles like data shortages or geopolitical events, AGI could still arrive in the next 2-3 years.
OpenAI is reportedly working to build a coalition with the U.S. government to rival China's AI capabilities, potentially involving the military and nuclear expertise.
This reflects the high stakes and global competition around advanced AI development.
The Importance of Inference Compute
While the initial training of large language models may be showing diminishing returns, the ability to scale up inference compute could unlock further breakthroughs.
Techniques like OpenAI's "constitutional AI" aim to make models more reliable and controllable during inference.
AI Naming Woes
Amodei acknowledged in an interview that AI companies have struggled with naming their models, suggesting there is room for improvement.
The Future of AI
Despite the debate over the pace of progress, the consensus seems to be that major AI advancements are still on the horizon, whether in 2-3 years or a bit further out.
Part 1/3:
AI Advancement: Slowing Down or Accelerating?
The Debate Over AI Progress
Major AI companies like Google and OpenAI have suggested that AI advancement is slowing down, with OpenAI co-founder Ilya Sutskever saying the 2010s were the "age of scaling" and now the industry is in the "age of wonder and discovery" again.
However, Anthropic CEO Dario Amodei sees a path to AGI (artificial general intelligence) by 2026 or 2027, a much shorter timeline than the slowdown narrative.
Reasons for Differing Interpretations
The slowdown may be related to diminishing returns from scaling up training data and models. The quality of data and how it's used for inference may matter more now.
But the ability to scale up inference compute, as seen with OpenAI's GPT-3 and upcoming models, could allow for continued rapid progress.
Amodei believes that even with potential hurdles like data shortages or geopolitical events, AGI could still arrive in the next 2-3 years.
[...]
Part 2/3:
The Push for an AI Alliance
OpenAI is reportedly working to build a coalition with the U.S. government to rival China's AI capabilities, potentially involving the military and nuclear expertise.
This reflects the high stakes and global competition around advanced AI development.
The Importance of Inference Compute
While the initial training of large language models may be showing diminishing returns, the ability to scale up inference compute could unlock further breakthroughs.
Techniques like OpenAI's "constitutional AI" aim to make models more reliable and controllable during inference.
AI Naming Woes
The Future of AI
[...]
Part 3/3: