Two of the smartest people I’ve followed in the AI world Sit down recently Check the developments in this field.
One is François Chollet, a widely used creator Keras Library and the author Arc-Agi Benchmarkit tests whether AI reaches “general” or widespread human intelligence. Chollet’s reputation is the reputation of AI bears, eager to reduce the maximum promotion of technology’s destination and over-objective predictions. But in the discussion, Chollet said his timeline has become shorter recently. Researchers have made great progress in what he believes is the main obstacle to achieving artificial universal intelligence, such as the weaknesses of model recalling and applying what they learned before.
Sign up here Explore the major and complex problems facing the world, and the most effective ways to solve them. Sent twice a week.
CHOLLET’s Interlocutor – Dwarkesh Patelits podcast has become the most important place to track AI scientists’ thoughts – reacting to their own reports, moving in the opposite direction. Humans are good at Continuous learning Or “at work,” Patel becomes even more pessimistic, and AI models can quickly acquire the skill.
“[Humans are] Learn from their failures. Patel notes that they are improving small improvements and efficiency. There seems to be no easy way to insert this critical function into these models. ”
All of this says two very smart people who know the field very well, and anyone else, can come to completely reasonable but contradictory conclusions about the pace of AI progress.
In this case, someone like me must be more knowledgeable than Chollet or Patel should figure out who is right?
Forecasting war, three years
One of the most promising ways I’ve seen is to resolve – or at least rule – these differences come from a name called Prediction Institute.
In the summer of 2022, the institute began the so-called There is risk persuasiveness (xpt abbreviation). XPT is Intentionally “Creating high-quality predictions of the risks facing humanity in the next century.” To do this, researchers (including Pennsylvania psychologists and Prediction pioneer Philip Tetlock And Friday’s chief Josh Rosenberg surveyed the subject matter experts of the study, who studied at least conceivable human survival (such as AI) that could endanger the summer of 2022.
But they also askedSuper Confucius“A group of people identified by Tetlock and others who are very accurate in predicting past events. SuperforeCasterGroup is not composed of experts who are threatening humanity, but generalists from various professions with solid predictive records.
In every risk, including AI, The huge gap between experts and generalist forecasters in specific areas. Experts than generalists say the risks they are studying could lead to extinction or mass deaths in humanity. Even if the researchers asked the two groups to conduct structured discussions, the gap still exists Why They disagree.
The two have fundamentally different worldviews. As far as AI is concerned, subject matter experts believe that burden of proof should be a skepticism to illustrate why a super-intelligent digital species Won’t It’s dangerous. The generalist believes that the burden of proof should be on experts to explain why technology that doesn’t even exist may kill us all.
So far, so tricky. Fortunately for American observers, each group is not only asked to estimate the long-term risks of the next century, which cannot be confirmed as soon as possible, but also in near-future events. Their task is to predict short, medium and long-term AI progress pace.
exist New paperAuthors – Tetlock, Rosenberg, Simas Kučinskas, Rebecca Ceppas de Castro, Zach Jacobs and Ezra Karger – return and evaluate how the two groups have performed in the three years since the summer of 2022.
In theory, this can tell us which group to believe in. If the concerned AI experts are better proven in predicting what will happen between 2022-2025, maybe this shows that they have a better understanding of the long-term future of technology, and therefore, we should give them a greater warning.
Alas, use Ralph Fiennes“That’s so simple!” As it turns out, the results of three years have left us without more awareness of who believes.
Both AI experts and SuperforeCaster systematically underestimate the pace of AI progress. Of the four benchmarks, the most advanced models in summer 2025 actually perform better than what Superholes or AI experts predict (although the latter is closer). For example, SuperforeCasters believes that AI will win gold at the 2035 International Mathematics Olympics. Experts believe 2030. It happened this summer.
“Overall, the average probability of Super Confucius is only the result of these four AI benchmarks, while the average probability of domain experts is only 9.7%,” the report concluded.
This makes the domain experts look better. They say Slightly What actually happens will have a higher chance of happening – but when they deal with the numbers in all the questions, the authors concluded that there is no statistically significant difference in overall accuracy between domain experts and super broadcasters. More importantly, there is no correlation between someone’s prediction in 2025 and the level of danger they believe AI or other risks. Prediction is still difficult, especially about future predictions, and especially About the future of AI.
The only trick to work reliably is to summarize everyone’s predictions – combine all predictions and make median predictions more accurate than any one person or group. We may not know which of these comforters is smart, but the crowd is still wise.
Maybe I should see this result. Ezra Karger, the economist and co-author of the original XPT paper and this new book, told me The first paper is released in 2023 That’s “There’s not much disagreement between people who disagree with long-standing issues over the next 10 years.” That is, they already know that the predictions of AI and people who don’t worry about are very similar.
Therefore, it is not surprising that one group is not better than the other during the forecast period 2022-2025. The real difference is not about the near future of AI, but about the danger it poses in the medium and long term, which is inherently difficult to judge and more speculative.
Maybe there is some valuable information that both groups underestimate how fast AI progress is: maybe it’s a sign that we all underestimate the technology, and that it will keep improving faster than expected. Again, the forecast for 2022 was made before Chatgpt, which was released in November of that year. Do you remember who predicted that AI chatbots would become ubiquitous at work and school before the app was launched? We’re not Know Has AI made a huge leap in capabilities in 2022-2025? Does this tell us anything about whether the technology may not slow down, which in turn will be the key to predicting its long-term threat?
Read the latest Friday report, I ended up in a similar place My former colleague Kelsey Piper last year. Piper pointed out that the failure to infer future trends, especially exponential trends, has led people to seriously misguided the past. The fact that relatively few Americans gained common in January 2020 does not mean that Kuvid is not a threat. This means the country is at the beginning of an exponential growth curve. Similar failures will lead people to underestimate the advancement of AI and with any potential risks.
Meanwhile, in most cases, exponential growth cannot last forever. It is maximized at some point. It is worth noting that Moore’s Law widely predicts the growth of microprocessor density To be precise, decades – but Moore’s law is known in part because it is unusual for the trend of technology created by humans to follow such a clean pattern.
“I’m increasingly convinced that when you think about these questions, there is no need to replace digging weeds,” Piper concluded. “Although there are some questions we can answer from the first principle, [AI progress] Not one of them. ”
I worry she is right – and worse, just respect for experts is not enough, and when experts disagree with each other on details and broad trajectory. We really don’t have a good choice to try to learn as much as possible in individuals while failing, waiting and seeing. This is not a satisfactory conclusion for the newsletter, nor is it a comforting answer to one of the most important questions facing humanity, but it is the best I can do.
