We acheived AGI yesterday and not a single thread about it?

noticer of things wrote:

OpenAI's latest model completely shattered all possible AI benchmarking tests, exceeding PhD-level / expert human scores across virtually all domains.

Not only that, it was able to solve 25% of problems from a math data set consisting of the most difficult, highly theoretical math problems in the world that only a handful of people (literally, like 5-10 people) are even capable of solving / proving. It did this without previous knowledge of the questions or answers. These tests were done independently, by the ARC prize foundation.

This is a massive, massive step change in performance from existing frontier AI models, which are only 3-6 months old.

Most people are simply not grasping how significant this - AI (specifically ChatGPT O3) is now able to solve some the most difficult math problems known to man. It is pushing very close to boundaries of human knowledge and doing so independently.

If you're wondering why you haven't "seen" any remarkable manifestations of this in the real world, consider a metaphor: We've invented the jet engine before we invented the airplane. This technology is so powerful, so advanced, we are still grappling with how to use it. But we'll figure it out quickly and when we do, it's off to the races.

noticer of things wrote:

OpenAI's latest model completely shattered all possible AI benchmarking tests, exceeding PhD-level / expert human scores across virtually all domains.

Not only that, it was able to solve 25% of problems from a math data set consisting of the most difficult, highly theoretical math problems in the world that only a handful of people (literally, like 5-10 people) are even capable of solving / proving. It did this without previous knowledge of the questions or answers. These tests were done independently, by the ARC prize foundation.

This is a massive, massive step change in performance from existing frontier AI models, which are only 3-6 months old.

Most people are simply not grasping how significant this - AI (specifically ChatGPT O3) is now able to solve some the most difficult math problems known to man. It is pushing very close to boundaries of human knowledge and doing so independently.

If you're wondering why you haven't "seen" any remarkable manifestations of this in the real world, consider a metaphor: We've invented the jet engine before we invented the airplane. This technology is so powerful, so advanced, we are still grappling with how to use it. But we'll figure it out quickly and when we do, it's off to the races.

lob wrote:

noticer of things wrote:

OpenAI's latest model completely shattered all possible AI benchmarking tests, exceeding PhD-level / expert human scores across virtually all domains.

Not only that, it was able to solve 25% of problems from a math data set consisting of the most difficult, highly theoretical math problems in the world that only a handful of people (literally, like 5-10 people) are even capable of solving / proving. It did this without previous knowledge of the questions or answers. These tests were done independently, by the ARC prize foundation.

This is a massive, massive step change in performance from existing frontier AI models, which are only 3-6 months old.

Most people are simply not grasping how significant this - AI (specifically ChatGPT O3) is now able to solve some the most difficult math problems known to man. It is pushing very close to boundaries of human knowledge and doing so independently.

If you're wondering why you haven't "seen" any remarkable manifestations of this in the real world, consider a metaphor: We've invented the jet engine before we invented the airplane. This technology is so powerful, so advanced, we are still grappling with how to use it. But we'll figure it out quickly and when we do, it's off to the races.

People were saying similar things about o1 a few months ago. Terry Tao said something along the the lines of using it was like interacting with a weak but not incompetent grad student.

Meanwhile, even with hints I can't get o1 to answer a majority of the homework questions I give my second year undergrads. Sure, it'll sometimes give a correct answer, but it has no ability to recognize when its answer is nonsense, no ability to ask for hints even when encouraged to do so, and seemingly very little ability to correct its own mistakes (for example, when pointing out an error in its reasoning, it will often accept that it's wrong but just rephrase its old solution when asked to try again). It also has complete confidence in its answers. When I prompt it to grade its own confidence in its answers and that a confidently wrong answer will be judged more harshly than no answer at all it still gives itself full marks on all of its self evaluations, including on questions where its answers were complete nonsense. It also has no ability to correct the mistakes of the user: if you ask it to solve a problem but include a typo that renders the problem itself non-sense it'll still happily provide a 'proof'.

These models will be worse than useless in most work environments until these sorts of overconfidence issues are corrected. I will be flabbergasted if the new o3 has made non-trivial progress towards this.

This isn't to say I think this technology is useless in general. But I do think that making it useful is a hard engineering task that will take lots of specific training for a given use case which won't be immediately applicable to other tasks. For example, projects like alphafold are along these lines and really cool. I also think openai will not heavily pursue such applications in the near future, and instead make way more money crushing benchmarks and claiming they're a couple more years of exponential growth away from completely changing the world.

noticer of things wrote:

lob wrote:

People were saying similar things about o1 a few months ago. Terry Tao said something along the the lines of using it was like interacting with a weak but not incompetent grad student.

Meanwhile, even with hints I can't get o1 to answer a majority of the homework questions I give my second year undergrads. Sure, it'll sometimes give a correct answer, but it has no ability to recognize when its answer is nonsense, no ability to ask for hints even when encouraged to do so, and seemingly very little ability to correct its own mistakes (for example, when pointing out an error in its reasoning, it will often accept that it's wrong but just rephrase its old solution when asked to try again). It also has complete confidence in its answers. When I prompt it to grade its own confidence in its answers and that a confidently wrong answer will be judged more harshly than no answer at all it still gives itself full marks on all of its self evaluations, including on questions where its answers were complete nonsense. It also has no ability to correct the mistakes of the user: if you ask it to solve a problem but include a typo that renders the problem itself non-sense it'll still happily provide a 'proof'.

These models will be worse than useless in most work environments until these sorts of overconfidence issues are corrected. I will be flabbergasted if the new o3 has made non-trivial progress towards this.

This isn't to say I think this technology is useless in general. But I do think that making it useful is a hard engineering task that will take lots of specific training for a given use case which won't be immediately applicable to other tasks. For example, projects like alphafold are along these lines and really cool. I also think openai will not heavily pursue such applications in the near future, and instead make way more money crushing benchmarks and claiming they're a couple more years of exponential growth away from completely changing the world.

Can you give us an example of a question that o1 gets consistently wrong but your undergrads can solve? I'm skeptical.

lob wrote:

noticer of things wrote:

Can you give us an example of a question that o1 gets consistently wrong but your undergrads can solve? I'm skeptical.

I have not been able to get it to give a reasonable proof of Schur's theorem (a common problem almost surely in its training data): Suppose that the natural numbers are partitioned into finitely many sets $C_1 \cup C_2 \cup ... \cup C_r$. Show that there is some $i \in \{1,...,r\}$ and natural numbers $x,y$ such that $\{x,y,x+y\} \subset C_i.$

When asked, o1 knows that this is Schur's theorem and knows that it is a corollary of Ramsey's theorem, but when asked to give the proof from scratch usually gives a strange case analysis that doesn't get close to a valid answer. It's also once given an answer that was pretty close (but still wrong) when I told it to use Ramsey's theorem in its proof (it answered along the lines of "color pair (x,y) by the color of x+y and then find a monochromatic triangle. When I told it that this was very nearly correct it could not find let alone correct the error).

I also couldn't get it to prove the infinite Ramsey's theorem itself (which, again, is almost surely in its training data). The prompt I used was: "Let X be the set of all unordered, distinct pairs of natural numbers. Show that in any partition of X into finitely many disjoint sets there
must be distinct natural numbers a,b,c such that {a,b}, {b,c}, and {a,c} all belong to the same part of our partition."

Usually it does the first part of the usual inductive proof correctly but then starts doing strange cases instead of repeating the same idea in the inductive step.

After these failures, I also gave it the following problem, which is an extremely easy pure reasoning problem and not even something I would give my students as a warm up problem. It also gave an incorrect answer: "Let N be the set of natural numbers, and let M be the NxN integer lattice (i.e., M consists of all ordered pairs of natural numbers (x,y), where it is possible that x=y). Suppose that I paint the elements of M red or blue. Is it true that I must either find a unit square whose corners are painted red or a unit square whose corners are painted blue?"

noticer of things wrote:

OpenAI's latest model completely shattered all possible AI benchmarking tests, exceeding PhD-level / expert human scores across virtually all domains.

Not only that, it was able to solve 25% of problems from a math data set consisting of the most difficult, highly theoretical math problems in the world that only a handful of people (literally, like 5-10 people) are even capable of solving / proving. It did this without previous knowledge of the questions or answers. These tests were done independently, by the ARC prize foundation.

This is a massive, massive step change in performance from existing frontier AI models, which are only 3-6 months old.

Most people are simply not grasping how significant this - AI (specifically ChatGPT O3) is now able to solve some the most difficult math problems known to man. It is pushing very close to boundaries of human knowledge and doing so independently.

If you're wondering why you haven't "seen" any remarkable manifestations of this in the real world, consider a metaphor: We've invented the jet engine before we invented the airplane. This technology is so powerful, so advanced, we are still grappling with how to use it. But we'll figure it out quickly and when we do, it's off to the races.