Large Quantitative Models with Will Zhang of Openhouse.ai

Speaker 0: one thing was interesting is, we are very deep in the, cutting and research, is that, current large language model, there's a lot of hallucination problem. and, that's actually what we are working very hard to solve that ourselves. that's why we call it large quantitative model. and if interest there, i can elaborate more about how do we actually see about embedding the physical phenomenon of the laws of physics into the language model. that's what we're working on right now for the home building industry.

Speaker 1: yeah. so i would love let's drill down on that in a bit. so you're calling your it's not a large language model. you're calling it again a what?

Speaker 0: a large quantitative model. quantitative model. and

Speaker 1: to reduce, the hallucinations, you're putting physics into this model?

Speaker 0: yeah. so it's the cutting edge challenge is actually how do we make it has reasoning capability. how do you bound the constraint into the system to say it has to follow this rule? i'll use a simple example. in the newtonian physics, right, so this, second law of motion, we know that f equals m a, which, like, the new new newtonian, discovery actually set the foundation of modern physical world leading to all this great discovery. and the second law of motion says, force equals mass multiple by acceleration, meaning something moving, you have to apply as the force to get it going, or you have to apply as force to make it stop. sounds trivial and obvious today. but for the last ten thousand years of history, mankind, this is only found only four hundred years ago. so they're not that far away if you look at the time horizon. the second thing is actually not trivial is that for newton to understand this acceleration, he had to invent calculus to do that. so if you think about llm, large language model, it's almost like the new invention of calculus. so that's actually tremendously powerful so that we can discover those new phenomenon. but one challenge in large language model is that it does not understand the causal relationship. meaning, you can easily just mathematically change your formula to say, well, you know what? if i can, increase the force or reshoot acceleration purely by just looking at math, i can change the mass because because mass equals f divided by a. right? so, oh, if i get more mass without changing acceleration sorry. if i give more forces without changing acceleration, then the thing will get heavier. that's what human is nonsense, but from mathematical standpoint, a large language model do not understand it. it's actually as far as the causal relationship to the acceleration. so that's actually what we're working very hard to actually think about deploying large language model to a industry that actually moving dirt, building houses in a physical environment that you are governed by the laws of physics. and that's why we realized that if we basically need to teach your algorithm that math is something that you do not change. it's the force that cause acceleration to change. and that's actually those are phenomena we basically need to feed into the language model before and after, and actually have this physical, loss of basic bounding the, the large language language model to make it applicable for this industry. and, that's actually got it very exciting about how do we also be part of advancing the technology development as well.

Speaker 1: it sounds very interesting. you know, when you're training large language models, oftentimes, you're you're fielding, feeding it a scenario and then a solution, a correct solution, a wrong solution, whatever, on these tables to train it. how do you train it, to understand that this formula equals the amount of physical dirt you have to move. like, it's a thing that's a leap for lack of a

Speaker 0: yeah. so that's actually of course, like, we do not have unlike openai or entropic, we don't have billions of dollars to spend to build our foundation model. we basically take advantage of existing large language model, also taking some open source model. and what is open source model by absorbing the data from internet in their latent feature, essentially, they build a understanding or abstraction of the world. so that allow it actually to create new solution in inference time. so, of course, when we feed data into it, we feed data by providing some like, for example, like, we say, hey. math is something you cannot change. we feed those into the model. but then when they're coming out, we then deploy tons of classical machine learning approach, layer on top of that to basically do optimization theory. so there's a field that is actually not typically you can put it as part of ai. some people might not. there's a field that's actually called operational research. it's how do you minimize or maximize the objective function. like, basically, for example, minimize the cost or maximize the, profit of the business or maximize the utility of a resource with constraint. and those constraint are the physical law the laws of physics. and, also, when there are uncertainty, how do we play out to know if this situation, what would happen? if the other situation, what would happen? when we're making decision under uncertainty, we use a lot of monte carlo simulation technique which involve a lot of volume compute, but very different compute compared to, and so we do a lot of things before and by basically wrapping large language model. but on top of that, which is not just wrapping it, we put layers of or more classical machine learning model that actually injecting the physical, physical phenomena into the technology to help them to make better decision.

Speaker 1: and here's another question. just again, i'm looking on the outside coming in my mind. when you say math doesn't change, you know, i don't know if that's actually true. and when i say that, i'll i'll explain that in a second.

Speaker 0: if you're looking

Speaker 1: at a scenario, say, i gotta move this volume of dirt. k? yeah. if it's clay, the mass might be different. if it's wet dirt, the mass might be different. if it's dry dirt, it might be different. if it's so how do you explain that to, say, a computer ai model? because the math actually doesn't change, but the environment changes enough that the formula math actually changes because of it.

Speaker 0: oh, abs i was more referring to that analogy, more like, well, for example, you can say, hey. like, you can make a demand protection, but the demand doesn't go negative. for example, like, in but the mathematic would say, oh, allow you to put negative view. paste now, put boundary into math model, but that actually all of a sudden. it's more like math now. it's not a continuous function. so that caused a lot of math just immediate breakdown. so your question is how do we feed that in is basically that's where it's actually in the post processing of the language model after you actually extract it out. so in input, we basically say, these are the phenomena that you have to be grounded in the input. in the training dataset, we actually also do a ton of feature engineering to basically bias the data intentionally to certain distribution. that's actually in the early days, the rgfl technique for large language model, how do they train? it's almost like, i'm gonna use another analogy to explain it. the internet might have a lot of cat information, a lot of cute cat. but if you want to ask a lot of language model, what's the most common picture you will see in the world? they might spit out as cat. so so how do you actually correct that? it's actually you re you you rebias those information back to the phenomena you know is governed by the laws of physics. so that's in the input phase. but on the output, then when the outcome is actually go reach beyond those laws of physics, we basically then kick in other machine learning technique to say, well, that's the outcome that we're gonna reject. there's a lot of governance. there's a lot of guidance. there's a lot of checks and balance around those to basically eliminate the obvious problem. and then the second layer is now we deploy simulation to understand the sensitivity exposure of such decision. how much risk are we actually talking about? so, it's getting a bit more deeply into the mathematic, but it's before and after, you basically put those rules and balance and policy around it so they make the right decision for people.

Speaker 1: i'm always fascinated about, you know, how this is done and how different companies are mhmm. are training ai and this large quantitative model versus, say, small language models and large language models and how everyone's having a different approach and now large action models that are out there. it's really nice to see, you know, how all these agents and things are coming into play.