3 ms·

> Some companies believe they can scale the data wall by using synthetic data — that is, data that is itself generated by A.I. systems — to train their models.

> Some companies believe they can scale the data wall by using synthetic data — that is, data that is itself generated by A.I. systems — to train their models. But many researchers doubt that today’s A.I. systems are capable of generating enough high-quality synthetic data to replace the human-created data they’re losing. I'm thinking synthetic datasets are how it's going to go. Of course you can't get information from nothing, but they might not need nearly as much seed data to generate lots of examples that are specifically designed to train reasoning skills. Maybe it will take a while to get over the hump, but they'll be motivated to make it work.

Out_of_Characte 2y ago
Garbage in Garbage out is real and shouldn't be underestimated. Its about as useful as trying to understand quantum physics from the perspective of a flat-earther
- WalterSear 2y ago
  Synthetic and generated data is also real and shouldn't be underestimated. LLM training can be very counterintuitive at times.
  - Out_of_Characte 2y ago
    Synthetic and generated data might not be garbage. Consider random chess positions. Those are generated,yet there's an engine out there that has objectively a higher chance of winning from that position. There was also a paper on persistent computer programs after running random data iteratively over millions of cycles. This doesn't even have a 'goal'. which by itself is very interesting because of how similar it feels to how organisms develop by the act of persisting after a few billion years. That is to say I largely agree with the premise that you can train and improve alot by absense of data. But if your original dataset is garbage, then training over and over might get limited by a hidden nash equilibrium that the algoritm can't escape due to lack of new information. Obligatory Anton petrov: https://www.youtube.com/watch?v=L_IWVZPmc-E https://www.youtube.com/watch?v=L_IWVZPmc-E Paper: https://arxiv.org/pdf/2406.19108 https://arxiv.org/pdf/2406.19108
- Workaccount2 2y ago
  Garbage in garbage out feels like a road bump more than a blockade. It doesn't seem far fetched at all for a future model to use known clean datasets to train on, and then become a curator of larger datasets.
- DiscourseFan 2y ago
  >Its about as useful as trying to understand quantum physics from the perspective of a flat-earther Wouldn't it make more sense to say trying to understand general relativity from the perspective of a flat-earther? At the quantum level, the world is flat--I mean, ontologically. "Dimension" is not used to describe space and time in the same way as in Einsteinian physics which retains a very Kantian model of the cosmos. But then you could certainly re-interpret Kant to make it cohere with quantum physics, I'm thinking of people like Deleuze and Latour here, which is why I alluded to a "flat" ontology. But in truth, it sort of negates your point, since when it comes to gathering things from the totality of the productive process (which is what an LLM does), indexing them and creating a surrealist wet-dream of output, there is not going to be a "right" answer, in just the same way that a quantum physicist and a cosmologist are not going to be able to agree on the correct definition of time. Truth is at the level of feeling, not sense. No scientific instrument can detect the feeling that one thing caused another, even though we are absolutely sure of its continuity, and it is such continuity that stands as the basis of all scientific investigation. In the end, empirical laws, just like moral laws, simply "feel right," and that's it. The LLM tells us things that feel right. It is a good-feeling machine, yet it is precisely when we notice something wrong, when it seems totally off, in those sublime gaps, that it becomes the most enigmatic and, in fact, the most useful. The real beauty of the AI is in those gaps, in its imagination, because it is actually our own, projected outward and reflected back like a society-wide mirror. There is something wonderful in the dreams of a machine.
- tensor 2y ago
  Synthetic data doesn't require using another ML models output as training data. One incredibly successful way of generating useful synthetic data is using rules. For example, AlphaZero uses synthetic data (playing against other versions of itself) but it does gain information because the data is generated against the rules of the game. Similarly, word2vec and other vector representations use synthetic training data, but that data is derived from observations of text (no, it's not using raw text, each training instance is generated from the text, with varying levels of simplicity of course). Tesseract is trained using synthetic data generated from font files and simulated "rendering errors." Synthetic data should not be underestimated.
itsumoiru 2y ago
As you mentioned, the data processing inequality[1] applies here, but I imagine synthetic data could help training squeeze out more from the existing data. [1] https://en.m.wikipedia.org/wiki/Data_processing_inequality https://en.m.wikipedia.org/wiki/Data_processing_inequality
- worstspotgain 2y ago
  It's neat how a longer "digestive tract" loses entropy, but can make up for it by making more sense of things. It's akin to adding a NN layer, to a more computationally-intensive lossy compression algorithm, or to asking a LLM to explain the problem domain and the relevant variables (populating attention) before getting to the point. It's probably true for people too. Instead of asking an expert for an opinion right away, ask them to discuss the options out loud first.
- skybrian 2y ago
  There are probably a lot of applications where the LLM could rely more on data that's supplied to it just-in-time in the context window, and less on specialist knowledge from its training set. Also, "natural" data taken from the Internet is probably quite inefficient as training material. It's going to have a lot of duplication. You only need each fact once to be able to synthesize more examples of it.
skeledrew 2y ago
No need even for synthetic data. Advance away from doing statistics on troves of human-generated data, to doing statistics on raw environmental data, like humans do. This will also likely mean moving away from the GPT architecture completely.
- FeepingCreature 2y ago
  Seems silly to say: "Instead of training this LLM on internet data, we will train the LLM on the output of an OCR model pointed at a screen scrolling internet data. This is clearly more ethical."
  - skeledrew 2y ago
    I made no allusion to ethics here; merely logic. Humans don't learn by scraping the web, but by observing the environment. The web is an incomplete distillation of humanity's knowledge of facts and various interpretations of them (which also includes "creative" works as one cannot create anything without basing said creations on environmental observations).
- Kostchei 2y ago
  This is the way. Robotics, self driving cars, mobile devices themselves, either phone or glasses.... They can generate the data needed for the next steps. (pun intentional) Text was easy, the process has been defined. Multi-modality improves the models. The next set of data is the world. Everything we currently observe using our senses and data outside our senses like radio waves. Build a better world model.

Out_of_Characte 2y ago

Garbage in Garbage out is real and shouldn't be underestimated. Its about as useful as trying to understand quantum physics from the perspective of a flat-earther

WalterSear 2y ago
Synthetic and generated data is also real and shouldn't be underestimated. LLM training can be very counterintuitive at times.
- Out_of_Characte 2y ago
  Synthetic and generated data might not be garbage. Consider random chess positions. Those are generated,yet there's an engine out there that has objectively a higher chance of winning from that position. There was also a paper on persistent computer programs after running random data iteratively over millions of cycles. This doesn't even have a 'goal'. which by itself is very interesting because of how similar it feels to how organisms develop by the act of persisting after a few billion years. That is to say I largely agree with the premise that you can train and improve alot by absense of data. But if your original dataset is garbage, then training over and over might get limited by a hidden nash equilibrium that the algoritm can't escape due to lack of new information. Obligatory Anton petrov: https://www.youtube.com/watch?v=L_IWVZPmc-E https://www.youtube.com/watch?v=L_IWVZPmc-E Paper: https://arxiv.org/pdf/2406.19108 https://arxiv.org/pdf/2406.19108
Workaccount2 2y ago
Garbage in garbage out feels like a road bump more than a blockade. It doesn't seem far fetched at all for a future model to use known clean datasets to train on, and then become a curator of larger datasets.
DiscourseFan 2y ago
>Its about as useful as trying to understand quantum physics from the perspective of a flat-earther Wouldn't it make more sense to say trying to understand general relativity from the perspective of a flat-earther? At the quantum level, the world is flat--I mean, ontologically. "Dimension" is not used to describe space and time in the same way as in Einsteinian physics which retains a very Kantian model of the cosmos. But then you could certainly re-interpret Kant to make it cohere with quantum physics, I'm thinking of people like Deleuze and Latour here, which is why I alluded to a "flat" ontology. But in truth, it sort of negates your point, since when it comes to gathering things from the totality of the productive process (which is what an LLM does), indexing them and creating a surrealist wet-dream of output, there is not going to be a "right" answer, in just the same way that a quantum physicist and a cosmologist are not going to be able to agree on the correct definition of time. Truth is at the level of feeling, not sense. No scientific instrument can detect the feeling that one thing caused another, even though we are absolutely sure of its continuity, and it is such continuity that stands as the basis of all scientific investigation. In the end, empirical laws, just like moral laws, simply "feel right," and that's it. The LLM tells us things that feel right. It is a good-feeling machine, yet it is precisely when we notice something wrong, when it seems totally off, in those sublime gaps, that it becomes the most enigmatic and, in fact, the most useful. The real beauty of the AI is in those gaps, in its imagination, because it is actually our own, projected outward and reflected back like a society-wide mirror. There is something wonderful in the dreams of a machine.
tensor 2y ago
Synthetic data doesn't require using another ML models output as training data. One incredibly successful way of generating useful synthetic data is using rules. For example, AlphaZero uses synthetic data (playing against other versions of itself) but it does gain information because the data is generated against the rules of the game. Similarly, word2vec and other vector representations use synthetic training data, but that data is derived from observations of text (no, it's not using raw text, each training instance is generated from the text, with varying levels of simplicity of course). Tesseract is trained using synthetic data generated from font files and simulated "rendering errors." Synthetic data should not be underestimated.

itsumoiru 2y ago

As you mentioned, the data processing inequality[1] applies here, but I imagine synthetic data could help training squeeze out more from the existing data. [1] https://en.m.wikipedia.org/wiki/Data_processing_inequality https://en.m.wikipedia.org/wiki/Data_processing_inequality

worstspotgain 2y ago
It's neat how a longer "digestive tract" loses entropy, but can make up for it by making more sense of things. It's akin to adding a NN layer, to a more computationally-intensive lossy compression algorithm, or to asking a LLM to explain the problem domain and the relevant variables (populating attention) before getting to the point. It's probably true for people too. Instead of asking an expert for an opinion right away, ask them to discuss the options out loud first.
skybrian 2y ago
There are probably a lot of applications where the LLM could rely more on data that's supplied to it just-in-time in the context window, and less on specialist knowledge from its training set. Also, "natural" data taken from the Internet is probably quite inefficient as training material. It's going to have a lot of duplication. You only need each fact once to be able to synthesize more examples of it.

skeledrew 2y ago

No need even for synthetic data. Advance away from doing statistics on troves of human-generated data, to doing statistics on raw environmental data, like humans do. This will also likely mean moving away from the GPT architecture completely.

FeepingCreature 2y ago
Seems silly to say: "Instead of training this LLM on internet data, we will train the LLM on the output of an OCR model pointed at a screen scrolling internet data. This is clearly more ethical."
- skeledrew 2y ago
  I made no allusion to ethics here; merely logic. Humans don't learn by scraping the web, but by observing the environment. The web is an incomplete distillation of humanity's knowledge of facts and various interpretations of them (which also includes "creative" works as one cannot create anything without basing said creations on environmental observations).
Kostchei 2y ago
This is the way. Robotics, self driving cars, mobile devices themselves, either phone or glasses.... They can generate the data needed for the next steps. (pun intentional) Text was easy, the process has been defined. Multi-modality improves the models. The next set of data is the world. Everything we currently observe using our senses and data outside our senses like radio waves. Build a better world model.