Why we might have superintelligence sooner than most think
If we reach superintelligence before we solve the alignment problem, we face a risk of extinction . So having an estimated range of when we could have superintelligence is essential to making sure we don’t get caught off guard. If our predictions are too far off, we may not be able to prepare in time.
We reached human-level performance in many domains in 2023
On Metaculus, the community prediction for (weak) AGI was 2057 just three years ago, and now it’s
Now, let’s dive into the definition of AGI used in that survey:
- Score >90% in the Winograd Schema Challenge
- Score >75% in SAT scores
- Pass a Turing test
- Finish Montezuma’s revenge
GPT-4 Scores 94.4% on the Winograd Schema Challenge , and 93% on the SAT reading exam, 89% on the SAT math exam . It hasn’t passed the Turing test, but probably not because of a lack of capabilities. It’s because GPT-4 has been fine-tuned to not mislead people. It’s not good for business if your AI is telling people it’s actually a person. That only leaves Montezuma’s Revenge. It is not unthinkable that it can be finished by a clever setup of GPT-4, using something like AutoGPT to analyze the screen and generate the correct inputs. In May 2023, GPT-4 was able to write code to get diamond gear in Minecraft . In short: GPT-4 got 2/4 criteria with certainty, with the other two in reach.
We’re there, folks. We already have (weak) AGI. It did not take us 35 years, it took us three. We were off by a factor of 10.
Why most underestimate the progress of AI
There are many reasons why people underestimate the progress of AI.
- It’s hard to keep up. Almost daily we see new breakthroughs in AI. It’s almost impossible to keep up with the pace of progress. You’re not alone if you feel like you’re falling behind.
- We keep moving the goalpost. In the 90s, people thought the holy grail of AI was something that could play chess. When AI beat Kasparov, its next challenge was Go. Now, we have machines that score in the 99.9th percentile in IQ tests , can translate 26 languages and win photography contests , yet we’re still asking questions like “When will AI reach human level?“. It already surpasses us in many areas, but we always focus on the increasingly small number of things we can still do better.
- We like to think that we’re special. Humans like to feel that we are special. If an AI can do what we can do, we’re not special anymore. This is a hard pill to swallow, and the brain has many defense mechanisms to avoid this .
- We’re really bad at exponential growth. We tend to structurally and predictably underestimate how exponential growth cumulates over time. This has been shown in scientific studies .
Luckily there are still some things that an AI can’t do yet. It cannot hack better than the best hackers , and it cannot do AI research as well as the best AI researchers. When we reach either of these thresholds, we will be in a new regime of increased risk.
So when will we reach the point when an AI can do all these things at a superhuman level? When will have a superintelligence?
The Ilya threshold
I think the crucial point that we should consider, is the point at which an AI is more capable of doing AI research than someone like Ilya Sutskever (chief scientist at OpenAI). An AI that can make meaningful contributions to AI algorithms and architectures is likely to be able to improve itself. Let’s call this point of potential self-improvement the Ilya threshold. When it reaches this, an AI might improve itself because it was explicitly instructed to do so, or because being smarter is a useful sub-goal for other goals (AIs are already creating their own sub-goals ). These iterations might take weeks (training GPT-3 took 34 days), but it is also possible that some type of runtime improvement is implemented that makes significant progress in a matter of minutes: an Intelligence Explosion .
So how far off are we from the Ilya threshold? It’s fundamentally difficult to predict when certain capabilities emerge as LLMs scale, but so far we’ve seen many capabilities emerge that were previously thought to be far off. GPT-4 is already an impressive programmer, and combined with AutoGPT it can do autonomous research on the internet . Being able to autonomously do AI research and making meaningful improvements to a codebase does not seem impossible in the near future. There are multiple potential paths to Ilya’s level of capabilities:
- A bigger model. This is likely to be a combination of more data, more parameters, more compute. So far scaling has proven very successful. The training costs are becoming astronomical ($100 million for GPT-4), but there’s a lot at stake, and there are many billions of dollars being invested in either staying ahead of the curve or catching up. New hardware is being developed that makes training larger models more feasible. A 10x size increase on its own might be enough to get us past the Ilya threshold. However, Sam Altman has stated that there are serious diminishing returns to scaling and that we might be close to the limit of what is achievable by scaling current architectures.
- Runtime improvements. Tools like MemoryGPT and AutoGPT have shown that simply putting an existing LLM in a new context unlocks radically new types of capabilities. Some have argued that LLMs right now have only access to something akin to “System 1” type thinking (fast, intuitive), and not “System 2” (slow, critical). It might be possible that a runtime improvement would unlock such capabilities.
So we have at least these three paths to reaching the Ilya threshold. We have no guarantee that any of these, or all of these, would bring us past the Ilya threshold, but it seems probable. It’s hard to quantify these, but there are now countless people working on all three of these - far more than just a few months ago. We have no idea how to align such an AI (even OpenAI admits this ), and the consequences of having a misaligned superintelligence are likely to be catastrophic .
We could have a superintelligence in months. A 1% risk is unacceptably large. We can only conclude that we need to slow down AI development right now. Read our proposal .