Latest / In the Loop AI / OpenAI’s Unreleased Model Solved Math: Is Biology Next?

Transcript

OpenAI’s Unreleased Model Solved Math: Is Biology Next? · In the Loop AI

0:00
Ejaaz: What's up guys, I just got to Japan, I am the most jet-lagged I have ever felt
0:06
Ejaaz: in my life, and I'm sat on my hotel room floor because I need to record this video.
0:12
Ejaaz: OpenAI just dropped one of their biggest releases and it's not an AI model,
0:17
Ejaaz: it's the announcement that they solved around 90 of the toughest problems in mathematics ever.
0:23
Ejaaz: And I know what you're thinking, you're wondering why on earth should I or you
0:27
Ejaaz: care about maths, That's something I haven't touched since high school or college.
0:31
Ejaaz: And the point of this video isn't that they solve these math problems.
0:35
Ejaaz: This will slowly and very quickly roll over into every particular field that
0:40
Ejaaz: you could think of that you probably studied to do your job today.
0:44
Ejaaz: It'll roll over into chemistry, into engineering, into physics,
0:48
Ejaaz: into whatever requires reasoning. And it's all happening much quicker than anyone can even imagine.
0:54
Ejaaz: So in this episode, I'm going to get into what exactly it's solved.
0:57
Ejaaz: But also more importantly, what happens next and what this means for you.
1:02
Ejaaz: All of that and more on today's episode of In The Loop, brought to you by Qualcomm.
1:09
Ejaaz: Let me just set some quick context. Three years ago, AI could not do math at
1:15
Ejaaz: all. Two plus two didn't equal four.
1:18
Ejaaz: It equaled 11. Strawberry had five R's in it.
1:22
Ejaaz: LLMs just weren't smart enough. And I remember commentary from Terence Tao,
1:27
Ejaaz: who is one of the leading mathematicians in the entire world.
1:30
Ejaaz: And he said, this is a reasonable attempt by AI, but I do not see how reasoning
1:37
Ejaaz: can eventually lead to solving maths.
1:40
Ejaaz: Reasoning is basically how you or I or any other human reasons about a problem.
1:44
Ejaaz: We don't just take it for face value, we apply different contexts to it.
1:48
Ejaaz: Back then, AIs could not do that, but they were at the start of their journey
1:52
Ejaaz: of figuring out this whole thing.
1:54
Ejaaz: Now, fast forward to six weeks ago, and OpenAI, who are back then and still
1:59
Ejaaz: are the experts in reasoning for training AI models, took that to its fullest
2:03
Ejaaz: extent when they announced that they had an internal model that was only 11 days old,
2:07
Ejaaz: had solved one of the Millennium Prize problems.
2:11
Ejaaz: Now, Millennium Prize problems, there's seven of them, and they're all in the
2:14
Ejaaz: field of mathematics. If anyone
2:17
Ejaaz: can go out there and solve any of these problems, you get $1 million.
2:20
Ejaaz: Now, for 200 years, these problems have existed. People have made progress,
2:24
Ejaaz: but the truth is, there are just too few mathematicians and experts in the world
2:28
Ejaaz: to try and figure this out, we are limited by human cognition and capacity.
2:32
Ejaaz: OpenAI came along with zero experience in maths, with zero help from math researchers,
2:38
Ejaaz: and they just decided, hmm, I wonder if I just use this powerful model that's
2:42
Ejaaz: more powerful than GPT-6 Astra and just let it go.
2:46
Ejaaz: Until it was able to solve the problems. There was infinite compute,
2:49
Ejaaz: infinite money, just get it done.
2:52
Ejaaz: And it solved it in 88 hours, 10,000 agents. It was one of the most costly experiments
2:56
Ejaaz: that they've done, probably that anyone's ever done for a research problem.
2:59
Ejaaz: It cost over $10 million.
3:01
Ejaaz: And there was a few interesting things about that. One, this was the first instance
3:05
Ejaaz: that an AI model had demonstrated that it wasn't just mimicking human intelligence.
3:10
Ejaaz: Up until then, people had been thinking, okay, AI models are smart,
3:14
Ejaaz: but they're just mimicking what it's learned on its data set.
3:16
Ejaaz: It's a next token predictor. What
3:18
Ejaaz: it means is it's trained on all human knowledge that we put out there.
3:21
Ejaaz: It's limited by the cap or the ceiling of how intelligent we already are.
3:26
Ejaaz: It solving Navier-Stokes was the first demonstration that it could go above
3:31
Ejaaz: that level, that it knew things that we as humans couldn't understand.
3:35
Ejaaz: And when Navier-Stokes got resolved, people weren't as excited as you might
3:39
Ejaaz: expect them to be, particularly mathematicians.
3:42
Ejaaz: They looked at this result and they weren't happy. They were perplexed.
3:46
Ejaaz: They were perplexed because they couldn't make sense of the solution.
3:50
Ejaaz: Now, fast forward to today and we have news that their new internal model has
3:56
Ejaaz: solved not just one, not two, but 90 out of 500 of the world's most hardest open math problems.
4:03
Ejaaz: That's 20% of all the hardest math problems that have existed for decades at this point.
4:08
Ejaaz: But there's a few differences between this and Navier-Stokes.
4:12
Ejaaz: It solved it in a much quicker way. On average, each problem was solved in three hours.
4:18
Ejaaz: And it also got it done at a much cheaper cost, costing only tens of thousands of dollars.
4:23
Ejaaz: So you can imagine or extrapolate in the future that these models will get faster
4:27
Ejaaz: or quicker and even cheaper to the point where the cost of solving problems
4:32
Ejaaz: that humans have spent decades trying to figure out becomes undeniable.
4:37
Ejaaz: You should just use AI and you don't need humans anymore.
4:40
Ejaaz: But I was critical. I saw this announcement from OpenAI here and I saw that
4:44
Ejaaz: they had dumped a massive GitHub, which was interesting. They open sourced the
4:47
Ejaaz: entire thing. So all the formal verification and proofs were laid out for anyone
4:51
Ejaaz: and everyone to critique.
4:52
Ejaaz: And it got me thinking, okay, it's solved a bunch of problems,
4:55
Ejaaz: but how is this useful for you or I?
4:58
Ejaaz: Like, how should I think about this being applied to the real world?
5:02
Ejaaz: And our friend Ole actually had a great breakdown in his tweet.
5:06
Ejaaz: He ran through a few of them, but being very high level here,
5:08
Ejaaz: because I'm not a mathematician, one of these problems is called Vlasov Maxwell.
5:12
Ejaaz: And if it's truly solved, this research can then result in a way to create virtually
5:19
Ejaaz: limitless clean energy to power civilization.
5:22
Ejaaz: Another example is the inverse elasticity problem that, if it's solved,
5:26
Ejaaz: can scan and map tissue stiffness, basically allowing you to evaluate and medicate
5:32
Ejaaz: and help individuals in the medical field.
5:34
Ejaaz: There's also another one called matrix multiplication. If you've heard that
5:37
Ejaaz: term before, it's because we've probably talked about it on this show.
5:41
Ejaaz: It's the foundation of how AI models run and how they run and compute all the
5:45
Ejaaz: important data to give you an amazing result. It's also how GPUs and any other
5:50
Ejaaz: modern graphics card processor works for videos, gaming, or whatever that might be.
5:54
Ejaaz: It's a very important fundamental mechanism of computing.
5:57
Ejaaz: It solved one of the most incredibly hard problems that that industry was facing.
6:02
Ejaaz: And if it's able to figure out, guess what? You get cheaper AI and bigger scientific
6:06
Ejaaz: simulations that you could use to map or design proteins that could result or
6:10
Ejaaz: end in the cure for cancer, for example.
6:13
Ejaaz: So the idea here is this isn't just a list of random niche, arbitrary mathematical problems.
6:19
Ejaaz: These are things that have stumped scientists, builders, engineers,
6:22
Ejaaz: and creators in the real world from innovating. It has prevented humans from
6:27
Ejaaz: innovating and building something that is marvelous and novel that would help
6:31
Ejaaz: world and humanity as it is today.
6:33
Ejaaz: So when you frame things this way, you start to realize that humanity's progress
6:37
Ejaaz: has always been dependent on the geniuses that they have in its civilization,
6:41
Ejaaz: in its society at the time.
6:43
Ejaaz: Who is smart enough to fix or figure out some of the hardest problems in the
6:46
Ejaaz: world so that we can progress as a human race?
6:49
Ejaaz: I loved this tweet from Will Depew, where he basically asked
6:54
Ejaaz: Fable 5.1 and GPT-6 Pro to rank all the discoveries that OpenAI's internal model
6:59
Ejaaz: had made over the last three years, just to see how important these discoveries
7:02
Ejaaz: actually were relative to that perspective that I just laid out,
7:05
Ejaaz: which is, you know, solving, you know, human innovation.
7:08
Ejaaz: And 81% of them were released on the day that OpenAI had solved these problems,
7:14
Ejaaz: meaning that a lot of these mathematical problems were fundamentally crucial.
7:17
Ejaaz: Just think about that for a second.
7:19
Ejaaz: The next Albert Einstein, the next Galileo isn't going to be a human.
7:25
Ejaaz: It's going to be a computer.
7:27
Ejaaz: It's going to be an AI because it's cheaper.
7:30
Ejaaz: It can work 24-7. And most importantly, it's now demonstrated that it's smarter
7:35
Ejaaz: than us. Framing it the way that Alex has here makes a lot of sense.
7:38
Ejaaz: He goes quasi RH. He's referencing one of the problems that it solved. Are you kidding me?
7:43
Ejaaz: If a human did this, it would be an instant Fields Medal. No questions asked.
7:47
Ejaaz: A Fields Medal is one of the highest honors you can earn in the mathematics field.
7:51
Ejaaz: PhDs, mathematicians make it one of their life goals to try and earn a Fields
7:55
Ejaaz: Medal. It is incredibly hard and a rare thing to do. It's very prestigious.
8:00
Ejaaz: And now the cost and effectiveness of that has been completely taken from mathematicians.
8:06
Ejaaz: And understandably, their reaction to this has mostly been mathematicians being
8:11
Ejaaz: very uncomfortable with the way open AI has approached this.
8:14
Ejaaz: Actually, it started off when they solved Navier Stokes six weeks ago.
8:17
Ejaaz: You might have thought mathematicians would have reacted positively to the news,
8:20
Ejaaz: but they really didn't. They were actually offended.
8:23
Ejaaz: They thought that the act of open AI using these LLMs to solve mathematical
8:27
Ejaaz: research, the act or thought of a computer doing their job that they had spent
8:32
Ejaaz: three to four years studying at a master's
8:34
Ejaaz: and bachelor's degree level and then and doing a PhD on top of this to,
8:37
Ejaaz: you know, their sole goal as a PhD student is to create net new information,
8:42
Ejaaz: can now be automated, that their prestige and their status that they've dedicated
8:46
Ejaaz: their life to has suddenly been taken away.
8:48
Ejaaz: And these are people that are performing at the top level of their field,
8:52
Ejaaz: suddenly automated overnight.
8:54
Ejaaz: Was a very uncomfortable feeling for them. And I totally get it.
8:57
Ejaaz: But at the same time, I'm grappling with this thinking, okay, well.
9:01
Ejaaz: The whole point about mathematicians doing their thing, chemists doing their
9:05
Ejaaz: things, Nobel laureates doing their thing, is so that we can push humanity forwards,
9:09
Ejaaz: that we can have a major invention or discovery
9:12
Ejaaz: that could make a lot of money, generate a lot of benefit for humans,
9:15
Ejaaz: and really help push us along.
9:17
Ejaaz: So why shouldn't we be happy about this AI model doing this for us?
9:20
Ejaaz: And I kind of tried to lay out my point in this tweet, where math PhDs basically
9:24
Ejaaz: dedicate their lives to trying to contribute in as small or as large a way as they can.
9:30
Ejaaz: But often, because we're humans, because we're limited in our ingenuity,
9:34
Ejaaz: because it's so hard of a field to figure out all these different problems,
9:37
Ejaaz: because our mental cognition and our context windows compared to AI LLMs
9:41
Ejaaz: are tiny, small, we need sleep, we're moody, we're irrational,
9:45
Ejaaz: we're emotional beings.
9:47
Ejaaz: Now we have a computer that can just work in the perfectly optimized way that we need to.
9:52
Ejaaz: And the thing that was holding back AI before, prior to this announcement,
9:57
Ejaaz: was that it was just trained on existing human data. It was dumbed down and
10:01
Ejaaz: held back by the limits of human capacity by even the smartest people.
10:05
Ejaaz: And now we see AI surpass that ingenuity. And that is a genuinely uncomfortable
10:10
Ejaaz: feeling, understandably.
10:11
Ejaaz: But you have to understand, this story isn't actually about the mathematicians.
10:15
Ejaaz: The way that they're feeling today is the way that experts across any single
10:19
Ejaaz: field are going to feel in the very near future.
10:22
Ejaaz: So I want to get into what that's going to look like and the major implications
10:25
Ejaaz: of that on today's show brought to you by Qualcomm.
10:31
Ejaaz: So the first thing we have to understand about this, aside from maths,
10:34
Ejaaz: is if reasoning is solving mathematics, which is one of the hardest fields, most complex
10:40
Ejaaz: to understand, learn, higher barrier of adoption to even enter the field,
10:44
Ejaaz: then it seems logical that this can apply to
10:48
Ejaaz: any other field that requires intense reasoning.
10:51
Ejaaz: The first thing that pops into mind that is closest to maths,
10:54
Ejaaz: it's kind of like the sister of maths, is physics.
10:56
Ejaaz: And then you kind of think of chemistry, all the numbers, equations that are
11:00
Ejaaz: required to create chemical compounds that could end up being a pharmaceutical
11:03
Ejaaz: product that cures a major ailment or illness. That is a big part of chemistry.
11:08
Ejaaz: And then it kind of rolls into systems biology and different aspects,
11:10
Ejaaz: but it's still very mathematical.
11:12
Ejaaz: And so those folks will probably push back and they'll say, but you do realize
11:16
Ejaaz: you need human intuition, you need judgment. And any feel that requires judgment
11:20
Ejaaz: is something that AI LLMs can't solve with just reasoning.
11:24
Ejaaz: And then you realize, well, actually, that's not strictly true,
11:28
Ejaaz: because you can technically take that logic and tell an LLM,
11:32
Ejaaz: hey, like, can you also try and reason and figure out human intuition and apply
11:36
Ejaaz: that? And it could end up doing that.
11:38
Ejaaz: But it's also about what OpenAI is not saying. And this is arguably the biggest
11:42
Ejaaz: point of the entire episode. This is post by Justin Drake. He is very renowned in cryptography.
11:49
Ejaaz: And he says, today I call upon the blockchain industry to calmly begin planning for bunker mode.
11:54
Ejaaz: What he goes on to explain is bunker mode is, we need to shift off of blockchain
11:59
Ejaaz: addresses because I think cryptography has been broken, which is weird because
12:04
Ejaaz: there's been no proof that an AI has done this.
12:07
Ejaaz: But he referred to OpenAI's announcement of solving all these mathematical problems.
12:10
Ejaaz: And he said the biggest problem about this is what they haven't said.
12:14
Ejaaz: He said, what is the coincidence of OpenAI's internal model being so smart to
12:19
Ejaaz: solve these niche problems that have existed for hundreds of years and somehow
12:22
Ejaaz: not been able to break the fundamentals of cryptography that has secured an industry
12:28
Ejaaz: worth trillions and trillions of dollars?
12:31
Ejaaz: And so the point he makes is, if an internal lab such as OpenAI has been able
12:35
Ejaaz: to do that, they wouldn't tell the public.
12:38
Ejaaz: And then I started to realize, oh wait, this model that OpenAI just used to
12:42
Ejaaz: do all these cool stuff isn't even accessible to you or I.
12:45
Ejaaz: It's accessible to OpenAI and they've had it for a while now, for months, for weeks.
12:50
Ejaaz: And look how much damage that they've done in that very little time.
12:54
Ejaaz: And then you start to think, well, if I were OpenAI and I had this model that
12:58
Ejaaz: could solve all these hard problems, I wonder what would happen if I pointed
13:02
Ejaaz: it at a problem that would give me a huge commercial advantage against my competitor
13:07
Ejaaz: Anthropic or Meta or Google.
13:09
Ejaaz: That's something that I would keep internally. And so we're starting to realize
13:13
Ejaaz: this gap being created where the model labs are forming these models that are
13:18
Ejaaz: very capable of self-improving, doing it in a very cost-effective manner.
13:22
Ejaaz: So the cost of human intelligence is rapidly declining, but it's not at the
13:27
Ejaaz: benefit of humans, it's at the benefit of these models, which are owned by these private companies.
13:32
Ejaaz: Have the incentive to not publicly release this for their own commercial or specific advantage.
13:39
Ejaaz: But I guess to challenge myself, on the flip side, if these models were to become publicly accessible,
13:45
Ejaaz: it creates another interesting challenge, which is anyone and everyone with
13:50
Ejaaz: access to this model who can think of a smart enough prompt can end up being as smart
13:55
Ejaaz: as Einstein, or at least achieving some of the greatest things that these former
13:59
Ejaaz: researchers have built over history in humanity.
14:03
Ejaaz: The future of human innovation is not really human at all.
14:08
Ejaaz: It's human and AI assisted. There is no future where we do not use this tool
14:13
Ejaaz: because it has now become cheap enough and more capable than us to be able to
14:17
Ejaaz: do this. So we need to work in confluence with this.
14:20
Ejaaz: But there is one major issue, which is humans, particularly the experts,
14:24
Ejaaz: have flipped from being the experts to becoming a complete bottleneck. Let me explain.
14:31
Ejaaz: With the math problems that have just been solved, there are 90 of them that
14:35
Ejaaz: have what's known as a lean resolution.
14:37
Ejaaz: So this is a formula that's been tested in a computer and it's come out with
14:42
Ejaaz: a big green check, 100 out of 100.
14:45
Ejaaz: But according to research practices today.
14:48
Ejaaz: You need a human to actually verify that it's true. We can't trust the computer.
14:53
Ejaaz: We need to trust the human. Problem is the humans, the smartest mathematicians,
14:58
Ejaaz: don't understand the formulas, don't understand the solutions that's in front of them.
15:02
Ejaaz: So it's going to take a lot of time.
15:04
Ejaaz: That time can cost us major innovation and leaps in the future.
15:08
Ejaaz: So I'm sitting here and I'm thinking about this and I'm thinking,
15:10
Ejaaz: hmm, what could help humans do this faster? AI.
15:14
Ejaaz: We should just use AI to check the AI solution.
15:18
Ejaaz: And then it becomes this weird thing where it's like, okay, well,
15:20
Ejaaz: why do we need the human in the loop anyway? Why do we need humans at all?
15:24
Ejaaz: Now, that is a very doomer and bearish take and I just want to be very upfront and clear.
15:28
Ejaaz: That's not actually where I think things will settle. I think to my earlier
15:32
Ejaaz: point, human intuition is a very important thing that is required to make the right judgment calls.
15:37
Ejaaz: I think there are plenty of fields out there that require this and it's not
15:41
Ejaaz: something that I think an AI model can intuitively figure out anytime soon.
15:45
Ejaaz: However, it's something that we should be wary of.
15:47
Ejaaz: And most importantly, that researchers of any other field should be wary of
15:50
Ejaaz: because it's going to slowly roll over.
15:52
Ejaaz: This is like a tidal tsunami that's going to sweep in very unexpectedly.
15:57
Ejaaz: And if you're wondering where I'm betting the next field is going to be,
16:01
Ejaaz: I think it's going to be biology, particularly focused on medical services or
16:07
Ejaaz: medical cures for the pharmaceutical industry.
16:10
Ejaaz: And there's good reasoning behind this. For the last six to 12 of months,
16:14
Ejaaz: major AI labs have been buying up as much scientific and biological data as they can.
16:20
Ejaaz: This data includes things like lab experiments and how to run these lab experiments
16:26
Ejaaz: and details on DNA and sequences that they can run through AI models.
16:31
Ejaaz: Anthropic has released a bunch of stuff on this. In fact, Anthropic two weeks
16:34
Ejaaz: ago announced that they have a automated wetware lab that they're building that
16:38
Ejaaz: you plug Claude into and it can operate robotic arms to try and run experiments autonomously.
16:44
Ejaaz: Now, the thing that a lot of critics very validly say is.
16:49
Ejaaz: AI models don't work in the physical world. You still need humans and intuition to go do the thing.
16:54
Ejaaz: But what happens when robotics accelerate and become good enough over the next
16:58
Ejaaz: three to five years, and these models are already vastly more capable than human
17:02
Ejaaz: intelligence, and you can just kind of plug them in and, you know,
17:05
Ejaaz: they create a cure for something the next day.
17:07
Ejaaz: That is a very weird world to kind of think of.
17:10
Ejaaz: And I fully think that humans are going to be heavily integrated into this.
17:14
Ejaaz: I guess I just don't know how. It's honestly hard for me to wrap my head around,
17:18
Ejaaz: if I'm being very honest with you.
17:19
Ejaaz: But I think it's an important one to talk about, obviously, as I'm sitting on
17:23
Ejaaz: this hotel floor and my fiance is trying to sleep in the bed next to me.
17:27
Ejaaz: So I hope you found this interesting. I felt very compelled to talk to you guys
17:32
Ejaaz: about this, because this isn't a story about maths at all.
17:35
Ejaaz: It's a story about AI finally being able to surpass human cognition,
17:40
Ejaaz: but doing it much quicker and cheaper than we could have anticipated.
17:45
Ejaaz: And so I'm scared or slightly nervous about how quickly this rolls into any
17:49
Ejaaz: and every other industry.
17:51
Ejaaz: And the ramifications that have on degree holders, workers, job roles that have
17:57
Ejaaz: existed for centuries that have been highly revered suddenly gets turned off overnight.
18:02
Ejaaz: Let me know what you guys think. Do you agree with my takes?
18:05
Ejaaz: Are there anything that I potentially missed?
18:08
Ejaaz: But thank you so much for listening. If you've listened this far,
18:11
Ejaaz: if you enjoyed this episode, you should definitely go check out some of our
18:14
Ejaaz: other episodes. We are back to posting day in and day out five times a week.
18:18
Ejaaz: If you have a friend that you think might find this episode or other episodes
18:21
Ejaaz: interesting, please share it with them. It helps us out massively.
18:24
Ejaaz: I'm anywhere and everywhere, by the way, on X in the description below,
18:27
Ejaaz: or you can just leave a comment here. I read all of them and I try and reply to all of them as well.
18:31
Ejaaz: But thank you so much for listening. This was today's episode on In The Loop
18:35
Ejaaz: brought to you by Qualcomm. See you on the next one.