No, but I bet his quants backtest trading strategies. His argument for dismissing LLMs is nonsensical, just because something uses information from the past doesn't mean its useless. In fact, every tool we have can match that criticism in one way or another
Yes, but backtesting does not need an LLM...backtesting has been a thing for many, many years.
Also, if you listen to any of his other speeches, or read the article, he does mention that they are using LLMs to port code from one language to another...so, he is not entirely dismissing LLMs.
You believe LLMs are capable of predicting financial markets?
In his line of work, he’s absolutely right. Correlations change. Markets react to themselves. LLMs may have some place in analyzing prior data…but it can’t predict what interest rates will be, whether or not congress will increase taxes on stock buybacks, or anticipate an armed conflict with China. Let’s get real here FFS
I agree but he wasn't talking about predicting financial markets, he was talking about global implications in industries like law and finance. And LLMs as a knowledge retrieval mechanism can already be pretty useful in those areas.
But the hype machine that's running overtime about LLMs is not making that case at all. It's all about how LLMs are some kind of miracle machine that will either solve or cause all of mankind's problems.
The hype is insane and is absolutely harming the cause that the pro-AI crowd is trying to push forward.
I wouldn't really mind this much aside from on grounds of taste, except that it's actively preventing people from actually understanding these technologies and being able to reach grounded conclusions about them.
Well it most definitely will cause unrest/revolution if it stalls the economic cycle where people can't get food.
That's happened enough times in recorded history to have a solid causal link. Which is already possible with current LLM capabilities; but since its decentralized won't be noticeable until its too late (like a dam breaking).
Yes, this is the main, and major, problem I see with this tech (and to be fair, it's not so much a problem with the tech as it's a problem with how we as a society handle work and earning a living). At this point, that result is looking probable to me. I hope I'm wrong.
But mainstream skepticism of LLMs largely revolves around absurd hype along the lines of "AI will be a malevolent superintelligence" (which is just the flip side of same absurd argument that many AI proponents engage in: "AI will be a beneficial superintelligence") which makes discussing more realistic risks and their possible solutions difficult if not impossible.
What, exactly, is already possible with current LLM capabilities? If you're implying they could replace workers, I don't think that will happen broadly. They're simply not that good yet, and it's unclear to me that they'll become that good in the next several years. And they definitely can't replace most jobs that involve any kind of manual labor.
Its already happening in a broad number of fields.
Bioinformaticians, copyrighting, research studies, law (though we've already seen the downsides). All the companies want to automate everything they can because of the cost savings even if there is a reduction in quality. Cents versus tens of dollars per hour is a no brainer in terms of cost so replacement it is.
Importantly, this is also just the start, and it gets better exponentially, currently the rate of change exceeds the rate at which we can evaluate the changes and thus take action.
Most manual labor jobs account for the lowest rung on the payscale outside of trades which have limited supportability due to limited market size.
1 company per 12-50,000 people for it to be economically viable depending on the trade.
Those simply can't absorb the hundreds of thousands of jobs and now deprecated business sectors where those degrees that people are still paying off on their college loans have no benefit.
Cite that bioinformaticians and lawyers/paralegals have been laid off in substantial numbers due specifically to machine learning / AI automation?
> currently the rate of change exceeds the rate at which we can evaluate the changes and thus take action
I dispute this. I believe we're seeing a plateau in capabilities, due to limitations of the transformer architecture and the massive, expensive amounts of compute it takes to create marginally better models. I don't believe we'll see anything like exponential growth in what LLMs can do.
You seem under the mistaken impression that if I can't cite it, its not happening. This is why cascade failures cause so much mayhem and death. Because so many people subscribe to show me proof, and if there is no tangible proof because proof is a lagging indicator its not happening. This is a complete failure in any professional risk management.
This is like a dam collapsing, I'm pointing at a crack, and you are saying show me proof this crack will cause everyone in the valley below to die. Its a flawed way of thinking in many ways especially with fundamental things that we base our lives and welfare on (as a society). In engineering there is different criteria for evaluating risk with items that are safety-critical. This is safety critical.
Here is an example of a copyrighter losing her job. There are many other examples if you bother to go looking.
You can dispute it all you want, it won't make you any less wrong when it happens, and many experts are deeply concerned because the rate of change has been exponential even if you refuse to believe it. There's a video that covers the important parts if you want to educate yourself, its on youtube.
> You seem under the mistaken impression that if I can't cite it, its not happening.
I don't even know how to respond to this. Think about the implications of what you're saying: you can claim anything, and when you're challenged on it, you can say "just because I have no proof doesn't mean it's not happening". If you have no proof that bioinformatics professionals are being laid off due to AI, then why did you bring them up as an example?
> I'm pointing at a crack, and you are saying show me proof this crack will cause everyone in the valley below to die.
No, you're pointing at what you believe is a crack, and I'm saying "show me proof that what you're pointing at is actually a crack and not a trick of the shadows".
> Its a flawed way of thinking in many ways especially with fundamental things that we base our lives and welfare on (as a society)
I don't know. I'd like to think that with fundamental things we base our lives on, we should try to be as certain as possible that we're responding to the right threats. We don't typically base our response to threats solely on how severe the damage would be if the threat came to pass -- we also run analyses to determine how likely the threat is to happen in the first place. There's a reason we don't throw trillions at the problem of preventing an asteroid impact or the eruption of Yellowstone.
> the rate of change has been exponential even if you refuse to believe it
I mean, people say that. What I've seen is:
* the transformer architecture is invented in 2017
* it takes years to make a broadly useful text completer with that architecture
* over the most recent span of a year (gpt-3.5 to gpt-4) we've seen relatively minor improvements in capability
* gpt-5 isn't even being trained yet (at least as of April)
I see some advancements, but they're all about making smaller, less useful models run on commodity hardware. The hard work of making an AGI doesn't seem like it's going anywhere right now. The current approaches require too much data and too much compute to see large improvements, and new approaches -- which don't yet exist -- need to be developed.
I believe the large number of people using LLMs are capable of noticing when there is a hole in the market that is actively being used to steal people's retirement money.
This announcement smacks of a fear trade to dissuade a lot of people from looking closer at things they shouldn't be looking closely at.
Backtest Gamestop and FRC. You'll notice two different inconsistencies with the options, and market makers actions/reactions even before the first breaking news that FRC was in trouble, or the fake news where people posted edited video claiming a bank run was in progress (on a Sunday, where the FRC sign was reversed in the wrong direction for a glass reflection relative to the position the shot was being taken from). Lets definitely get real here.
What happens when you create a lot of additional counterfeit shares in a market where the number of buyers or sellers on the ticker decides price action (up or down)... It goes down.
Ask yourself how holding/writing options works when a company goes into receivership, and how it differs when you are a market maker instead of a individual. Do you really believe there is no preferential treatment where the entity deciding has a conflict of interest (because its on the hook if the market maker is on the hook). How is that not jumping on the bandwagon of potential price manipulation to bankrupt a company (any company)?
Obviously do your own due dilligence. When they say its done on a case-by-case basis, and they have a vested interest in the outcome; shady stuff rarely ever gets punished; its how corruption and fraud work.
Anomaly detection has been around for a long time. You don’t need a chat based LLM to analyze data.
The beauty of the markets is that you can put your money where your mouth is. If you believe there’s an inefficiency go fill it yourself and become rich.
Completely agree on his take. Generative AI perhaps might be good at certain, specific tasks in terms of pulling or summarizing information for an analyst, but it won't help predict markets. It'll be much better at helping aid in software where the problem is more deterministic.
I am always a bit disappointed when only few arguments are provided. Ken Griffin might have a good understanding of the implications of LLM or not but that is hard to tell here.
Especially law will be interesting to watch. Lawmakers have to encode their intentions into text. LLM are good at detecting patterns in texts, find inconsistencies and so on. On top of that I would argue that we learn from the past to predict the future. Laws do not change as frequently as tech does. So LLMs might turn out to be excellent at understanding the law, at least written law, and experienced from all the cases they saw during training. I think law is in for a change similar to software.
Lawyers are already sufficiently good at "detecting patterns in texts, find inconsistencies and so on" in written legal texts, that's not really the hard part in practicing law. Even legal research, despite what non-lawyers may think, isn't all that big of a problem for wide swaths of lawyers. In any given field, there's going to be a short list of relevant cases to know, and everyone will know pretty quickly when a new one comes out. Experienced lawyers don't generally spend a lot of time doing legal research. The hard part is in fitting the facts of your case to the fact pattern of the cases you're relying on for your legal argument, and then fitting it all together so that its easy to read, understand, and agree with.
28 comments
[ 2.7 ms ] story [ 73.6 ms ] threadSo Ken sees books and backtesting trading strats as useless too?
I really expected more knowledgeable answers
You think Ken Griffin is reading books on how to make money in the market? Running a successful market maker? Maybe, but I doubt it.
I think in the context of his business, he is probably right...
Also, if you listen to any of his other speeches, or read the article, he does mention that they are using LLMs to port code from one language to another...so, he is not entirely dismissing LLMs.
In his line of work, he’s absolutely right. Correlations change. Markets react to themselves. LLMs may have some place in analyzing prior data…but it can’t predict what interest rates will be, whether or not congress will increase taxes on stock buybacks, or anticipate an armed conflict with China. Let’s get real here FFS
But the hype machine that's running overtime about LLMs is not making that case at all. It's all about how LLMs are some kind of miracle machine that will either solve or cause all of mankind's problems.
The hype is insane and is absolutely harming the cause that the pro-AI crowd is trying to push forward.
I wouldn't really mind this much aside from on grounds of taste, except that it's actively preventing people from actually understanding these technologies and being able to reach grounded conclusions about them.
That's happened enough times in recorded history to have a solid causal link. Which is already possible with current LLM capabilities; but since its decentralized won't be noticeable until its too late (like a dam breaking).
That's happened enough times in recorded history to have a solid causal link."
Examples? Genuinely curious
https://www.csis.org/analysis/dangerously-hungry-link-betwee...
https://politicalviolenceataglance.org/2022/07/12/food-insec...
Less recently: https://law.uoregon.edu/sites/law2.uoregon.edu/files/fakhri_...
Boyd Orr also wrote on the subject and won the Nobel peace price in 1949 for his work.
But mainstream skepticism of LLMs largely revolves around absurd hype along the lines of "AI will be a malevolent superintelligence" (which is just the flip side of same absurd argument that many AI proponents engage in: "AI will be a beneficial superintelligence") which makes discussing more realistic risks and their possible solutions difficult if not impossible.
Bioinformaticians, copyrighting, research studies, law (though we've already seen the downsides). All the companies want to automate everything they can because of the cost savings even if there is a reduction in quality. Cents versus tens of dollars per hour is a no brainer in terms of cost so replacement it is.
Importantly, this is also just the start, and it gets better exponentially, currently the rate of change exceeds the rate at which we can evaluate the changes and thus take action.
Most manual labor jobs account for the lowest rung on the payscale outside of trades which have limited supportability due to limited market size.
1 company per 12-50,000 people for it to be economically viable depending on the trade. Those simply can't absorb the hundreds of thousands of jobs and now deprecated business sectors where those degrees that people are still paying off on their college loans have no benefit.
> currently the rate of change exceeds the rate at which we can evaluate the changes and thus take action
I dispute this. I believe we're seeing a plateau in capabilities, due to limitations of the transformer architecture and the massive, expensive amounts of compute it takes to create marginally better models. I don't believe we'll see anything like exponential growth in what LLMs can do.
This is like a dam collapsing, I'm pointing at a crack, and you are saying show me proof this crack will cause everyone in the valley below to die. Its a flawed way of thinking in many ways especially with fundamental things that we base our lives and welfare on (as a society). In engineering there is different criteria for evaluating risk with items that are safety-critical. This is safety critical.
Here is an example of a copyrighter losing her job. There are many other examples if you bother to go looking.
https://www.indiatoday.in/technology/news/story/tech-copywri...
You can dispute it all you want, it won't make you any less wrong when it happens, and many experts are deeply concerned because the rate of change has been exponential even if you refuse to believe it. There's a video that covers the important parts if you want to educate yourself, its on youtube.
https://www.youtube.com/watch?v=xoVJKj8lcNQ&t=5s&ab_channel=...
I don't even know how to respond to this. Think about the implications of what you're saying: you can claim anything, and when you're challenged on it, you can say "just because I have no proof doesn't mean it's not happening". If you have no proof that bioinformatics professionals are being laid off due to AI, then why did you bring them up as an example?
> I'm pointing at a crack, and you are saying show me proof this crack will cause everyone in the valley below to die.
No, you're pointing at what you believe is a crack, and I'm saying "show me proof that what you're pointing at is actually a crack and not a trick of the shadows".
> Its a flawed way of thinking in many ways especially with fundamental things that we base our lives and welfare on (as a society)
I don't know. I'd like to think that with fundamental things we base our lives on, we should try to be as certain as possible that we're responding to the right threats. We don't typically base our response to threats solely on how severe the damage would be if the threat came to pass -- we also run analyses to determine how likely the threat is to happen in the first place. There's a reason we don't throw trillions at the problem of preventing an asteroid impact or the eruption of Yellowstone.
> the rate of change has been exponential even if you refuse to believe it
I mean, people say that. What I've seen is:
* the transformer architecture is invented in 2017
* it takes years to make a broadly useful text completer with that architecture
* over the most recent span of a year (gpt-3.5 to gpt-4) we've seen relatively minor improvements in capability
* gpt-5 isn't even being trained yet (at least as of April)
I see some advancements, but they're all about making smaller, less useful models run on commodity hardware. The hard work of making an AGI doesn't seem like it's going anywhere right now. The current approaches require too much data and too much compute to see large improvements, and new approaches -- which don't yet exist -- need to be developed.
This announcement smacks of a fear trade to dissuade a lot of people from looking closer at things they shouldn't be looking closely at.
Backtest Gamestop and FRC. You'll notice two different inconsistencies with the options, and market makers actions/reactions even before the first breaking news that FRC was in trouble, or the fake news where people posted edited video claiming a bank run was in progress (on a Sunday, where the FRC sign was reversed in the wrong direction for a glass reflection relative to the position the shot was being taken from). Lets definitely get real here.
What happens when you create a lot of additional counterfeit shares in a market where the number of buyers or sellers on the ticker decides price action (up or down)... It goes down.
Ask yourself how holding/writing options works when a company goes into receivership, and how it differs when you are a market maker instead of a individual. Do you really believe there is no preferential treatment where the entity deciding has a conflict of interest (because its on the hook if the market maker is on the hook). How is that not jumping on the bandwagon of potential price manipulation to bankrupt a company (any company)?
Obviously do your own due dilligence. When they say its done on a case-by-case basis, and they have a vested interest in the outcome; shady stuff rarely ever gets punished; its how corruption and fraud work.
The beauty of the markets is that you can put your money where your mouth is. If you believe there’s an inefficiency go fill it yourself and become rich.
Especially law will be interesting to watch. Lawmakers have to encode their intentions into text. LLM are good at detecting patterns in texts, find inconsistencies and so on. On top of that I would argue that we learn from the past to predict the future. Laws do not change as frequently as tech does. So LLMs might turn out to be excellent at understanding the law, at least written law, and experienced from all the cases they saw during training. I think law is in for a change similar to software.