Are We Heading Towards AGI and Mass Unemployment Soon, or Not Yet?

If you look at what AI companies' CEOs are saying, we are already at AGI (whatever that means) and on a straight path to artificial superintelligence very soon. After all, AI models are already at the top human level at some tasks and becoming increasingly superhuman at them. However, if you closely inspect the state of the art of current frontier models, it becomes clear that, while they are becoming superhuman in some domains (those with clear objectives that are easy to verify), they are still lagging behind in many others.

Currently, even the latest frontier AI models can still be described as having islands of superhuman capabilities in a sea of failures, aka "jagged intelligence". Yes, the sea level rises over time, and islands become mountains, but the disparity remains. These AI models can already perform tasks only top humans can, yet they still stumble over things people consider trivial. Why is that?

Part of it has to do with how these models are trained. Simply put, they are placed in special training environments that automatically assess their results, where they are given one goal: to get the best score they can. Then the wiring of these models' brains changes to whatever gets the highest score. These environments must assess results automatically because AI still needs millions of examples during training to develop a novel skill it didn't have before. It's relatively easy to do with math and coding: you specify precise tasks (solve a math problem or write a piece of software), and then you construct an environment that assesses these tasks. As the diversity of tasks and environments increases, models become better and better at completing these and similar tasks.

Here lies the first major problem of AI research right now. No human can provide millions of examples of every economically valuable task. This doesn't mean AI can't be useful, since writing code alone has myriad economic uses, but it does mean it can't learn all economically significant tasks humans do within the current paradigm. It can only learn tasks that have a clear goal and can be assessed automatically.

The next major problem has to do with the messy nature of many economically useful tasks humans face. A senior developer doesn't receive a clear description of a bug and the end goal, along with a list of tests the code should pass after the bug is fixed. Most of the time, it's something like, "Hey, our users are complaining about some weird problem, go sort it out." AI models are very good at clearly specified tasks that can be easily decomposed into subtasks, each with a clear, verifiable objective. They are much worse at messy real-world tasks, where the objective isn't clear at first and can change multiple times along the way. This isn't just my experience. There's actual scientific research measuring the gap between AI performance on so-called "messy" tasks and much clearer ones. And the real world doesn't look like most AI benchmarks with clear objectives. It is often messy, and it's often unclear what exactly you want to achieve.

Finally, there is a reliability problem. According to METR's measurements, current state-of-the-art models can finish around 50% of tasks that take a human expert 16 hours or more (METR's task suite can't reliably measure beyond that), yet they fail at such tasks the other 50% of the time. If we raise the bar to 80% reliability, that's just about 3 hours of work. If some application requires even higher reliability, such as 99% or 99.9%, AI will only be able to reliably finish very small tasks.

The clearest example of this problem is the APEX-Accounting benchmark, where nine frontier AI models were given 160 real-world accounting tasks. On average, the best model passed just over half, or 56.4% of these tasks. However, the best Pass^8 result (meaning the model must complete the same task correctly eight times in a row) was only 2.6%. For 97.4% of tasks, even the most consistent model made a mistake at least once in eight attempts. If you knew your accountant was that unreliable, you would not hesitate to hire another one. That said, if a human can easily spot a mistake, a 56.4% hit rate is pretty good for a tool, and it can make an accountant more productive. Still, you wouldn't want an AI model to be your accountant without human supervision. At least, not yet.

To summarize, current AI models are becoming extremely good at clearly specified, verifiable tasks, but they are still much worse as parts of reliable autonomous processes. Despite impressive gains in capabilities across verifiable domains, reliability and performance on messy tasks outside those domains remain major problems. It remains to be seen whether AI companies will be able to solve these issues quickly or whether they will remain a major hurdle stalling further AI progress.

Best,
The EasyAlgos Team