AI Won’t Plateau — if We Give It Time To Think | Noam Brown | TED - Video Insight
AI Won’t Plateau — if We Give It Time To Think | Noam Brown | TED - Video Insight
TED
Fullscreen
HumanAmplify Chrome Extension Get our Chrome Extension


The evolution of AI over the last five years highlights the importance of scaling both data and thoughtful processing time in enhancing performance.

In the past five years, the field of artificial intelligence (AI) has witnessed unprecedented progress, predominantly driven by an increase in scale. Notably, while foundational algorithms like the transformer architecture introduced in 2017 remain in use, the scale of data and computational resources allocated to training these models has expanded dramatically. For instance, the cost of training a model like GPT-2 in 2019 was approximately $5,000, whereas contemporary models reach costs of up to hundreds of millions. This escalating investment raises concerns about a potential plateau in AI progress. However, the speaker asserts that AI development will not stagnate, signaling an expectation that advancements will continue at an even faster rate due to emerging paradigms in model training and performance enhancement. The speaker reflects on their experience as a PhD student working on developing AI capable of playing poker. Initially, the prevailing notion in the research community was to improve performance by merely scaling the models, which involved training increasingly larger AIs on vast datasets. Despite these efforts, early poker AIs struggled against expert human players. A pivotal realization occurred when the speaker discovered that implementing a brief 20-second thinking period significantly enhanced the bot's decision-making capability, achieving performance gains equivalent to a dramatic increase in model size and training duration. Following this revelation, the team restructured their approach to AI design, leading to remarkable success against human counterparts, highlighting the importance of incorporating


Content rate: A

The content provides a deep analysis of AI development trends backed by personal experiences and empirical evidence from strategic experiments. The discussion incorporates both historical context and forward-looking perspectives that challenge common concerns about stagnation in the field, making it highly informative and thought-provoking.

AI research poker models training

Claims:

Claim: The cost of training AI models has drastically increased over five years.

Evidence: The training cost of GPT-2 was approximately $5,000 in 2019, whereas today's models cost hundreds of millions of dollars.

Counter evidence: Concerns exist regarding the sustainability of these costs in the long term, and some experts posit that models may eventually reach limits due to financial constraints.

Claim rating: 8 / 10

Claim: Extending thinking time significantly boosts AI performance.

Evidence: A poker AI demonstrated that spending just 20 seconds on decisions resulted in a performance boost equivalent to scaling the model size by 100,000 times.

Counter evidence: Some argue that focusing solely on longer thinking times could hinder real-time applications, such as instantaneous responses in chatbots.

Claim rating: 9 / 10

Claim: Advancements in AI will continue to accelerate rather than plateau.

Evidence: The speaker expresses confidence in ongoing AI development, asserting that current challenges can be addressed by scaling System 2 thinking.

Counter evidence: Skeptics argue that the financial and computational resources required may eventually halt unprecedented growth, suggesting that a plateau might be inevitable.

Claim rating: 7 / 10

Model version: 0.25 ,chatGPT:gpt-4o-mini-2024-07-18

Here's what you need to know: In recent years, the progress in artificial intelligence has primarily revolved around one key factor: scale. While advances in algorithms are important, modern frontier models still rely on the transformer architecture introduced back in 2017. Over the last five years, these models have dramatically increased in size and the amount of data they are trained on, leading to substantial improvements. However, concerns exist about the sustainability of these enormous training costs, which can run into the hundreds of millions or even billions of dollars. The speaker shares insights from their own experiences developing poker-playing AI, highlighting how initial models failed to match human expertise despite extensive data training. It was found that giving the AI more thinking time—similar to a human’s more methodical decision-making process—greatly enhanced its performance. This finding led to a redesign that prioritized longer thinking times, ultimately allowing the AI to outperform top human players significantly. This principle of scaling up thinking time is not limited to poker, as similar patterns have been observed in chess and go. The emergence of new language models, like the o1 series, demonstrates that longer thinking times can also lead to better performance in complex tasks. Although these models may take longer and cost more per query, they could be invaluable for important challenges, transforming our perceptions of AI beyond mere chatbots to powerful problem-solving tools. In conclusion, the future of AI could lie not just in scaling up training but also in enhancing the thinking capabilities of these systems, hinting at a new frontier in AI development that is already underway.