7 ms·

I'm convinced GPT is running separate helper functions on input and output tokens to fix the 'tokenization' issues. As in, find items of math, send it to this h

I'm convinced GPT is running separate helper functions on input and output tokens to fix the 'tokenization' issues. As in, find items of math, send it to this hand made parser and function, then insert result into output tokens. There's no other way to fix the token issue. For reference, Let's build the GPT Tokenizer https://www.youtube.com/watch?v=zduSFxRajkE https://www.youtube.com/watch?v=zduSFxRajkE

nine_k 3y ago
I personally find approaches like this the correct way forward. An input analyzer that finds out what kinds of tokens the query contains. A bunch of specialized models which handle each type well: image analysis, OCR, math and formal logic, data lookup,sentiment analysis, etc. Then some synthesis steps that produce a coherent answer in the right format.
- michaelt 3y ago
  Then you might enjoy looking up the "Mixture of Experts" model design.
  - numeri 3y ago
    That has nothing to do with the idea of ensembling multiple specialized/single-purpose models. Mixture of Experts is an method of splitting the feed-forwards in a model such that only a (hopefully) relevant subset of parameters is run for each token. The model learns how to split them on its own, and usually splits based not on topic or domain, but on grammatical function or category of symbol (e.g., punctuation, counting words, conjunctions, proper nouns, etc.).
    - michaelt 3y ago
      An ensemble of specialists is different to a mixture of experts? I thought half the point of MoE was to make the training tractable by allowing the different experts to be trained independently?
- CuriouslyC 3y ago
  Yeah. Have a multimodal parser model that can decompose prompts into pieces, generate embeddings for each of them and route those embeddings to the correct model based on the location of the embedding in latent space. Then have a "combiner/resolver" model that is trained to take answer embeddings from multiple models and render it in one of a variety of human readable formats. Eventually there is going to be a model catalog that describes model inputs/outputs in a machine parseable format, all models will use a unified interface (embedding in -> embedding out, with adapters for different latent spaces), and we will have "agent" models designed to be rapidly fine tuned in an online manner that act as glue between all these different models.
- hackerlight 3y ago
  Doesn't the human brain work like this? Yeah it's all connected together and plastic and so on, but functions tend to be localized, e.g vision is in occipital area. These base areas are responsible for the basic latent representations (edge detectors) which get fed forward to the AGI module (prefrontal cortex) that coordinates the whole thing based on the high quality representations it sees from these base modules. This strikes me as the most compute efficient approach.
Workaccount2 3y ago
I'd almost say anyone not doing that is being foolish. The goal of the service is to answer complex queries correctly, not to have a pure LLM that can do it all. I think some engineers feel that if they are leaning on an old school classically programed tool to assist the LLM, it's somehow cheating or impure.
- uoaei 3y ago
  Of course. But we must acknowledge that many have blinders on, assuming that scale is all you need to beat statistical errors.
  - sigmoid10 3y ago
    Well, these people are not wrong per se. Scale is what drove what we have today and as hardware improves, the models will too. It's just that in the very short term it turns out to be faster to just code around some of these issues on the backend of an API rather than increase the compute you spend on the model itself.
    - uoaei 3y ago
      Monkey sees moon. Monkey climbs tree. "See? Monkey is closer to moon than before. To reach moon, monkey just needs taller tree." How long before monkey finds tall enough tree to reach moon?
      - sigmoid10 3y ago
        We're rapidly approaching the compute capacity of the human brain in individual server racks. This "moon" is neither unreachable nor is there any doubt that we will cross the threshold soon.
        uoaei 3y ago
        I find it incredibly hard to believe we stumbled upon an efficient architecture that requires nothing but more compute not 10 years after the AI winter thawed. That's incredibly optimistic to the point of blind hope. What is your background and what makes you think we've somehow already figured everything out?
vidarh 3y ago
GPT has for some time output "analyzing" in a lot of contexts. If you see that, you can go into settings and tick "always show code when using data analyst" and you'll see that it does indeed construct Python and run code for problems where it is suitable.
Jabrov 3y ago
What if we used character tokens?
Der_Einzige 3y ago
I wrote a whole paper about ways to "fix" tokenization in a plug-and-play fashion for poetry generation: Filter the vocabulary before decoding. https://paperswithcode.com/paper/most-language-models-can-be-poets-too-an-ai-1 https://paperswithcode.com/paper/most-language-models-can-be...
data-ottawa 3y ago
ChatGPT definitely has a growing bag of tricks like that. When I use analysis mode to generate and evaluate code it recently started writing the code, then introspecting it and rewriting the code with an obvious hidden step asking "is this code correct". It made a huge improvement in usability. Fairly recently it would require manual intervention to fix.
bevekspldnw 3y ago
You can often see it write and execute python code to answer a question which is awesome.

nine_k 3y ago

I personally find approaches like this the correct way forward. An input analyzer that finds out what kinds of tokens the query contains. A bunch of specialized models which handle each type well: image analysis, OCR, math and formal logic, data lookup,sentiment analysis, etc. Then some synthesis steps that produce a coherent answer in the right format.

michaelt 3y ago
Then you might enjoy looking up the "Mixture of Experts" model design.
- numeri 3y ago
  That has nothing to do with the idea of ensembling multiple specialized/single-purpose models. Mixture of Experts is an method of splitting the feed-forwards in a model such that only a (hopefully) relevant subset of parameters is run for each token. The model learns how to split them on its own, and usually splits based not on topic or domain, but on grammatical function or category of symbol (e.g., punctuation, counting words, conjunctions, proper nouns, etc.).
  - michaelt 3y ago
    An ensemble of specialists is different to a mixture of experts? I thought half the point of MoE was to make the training tractable by allowing the different experts to be trained independently?
CuriouslyC 3y ago
Yeah. Have a multimodal parser model that can decompose prompts into pieces, generate embeddings for each of them and route those embeddings to the correct model based on the location of the embedding in latent space. Then have a "combiner/resolver" model that is trained to take answer embeddings from multiple models and render it in one of a variety of human readable formats. Eventually there is going to be a model catalog that describes model inputs/outputs in a machine parseable format, all models will use a unified interface (embedding in -> embedding out, with adapters for different latent spaces), and we will have "agent" models designed to be rapidly fine tuned in an online manner that act as glue between all these different models.
hackerlight 3y ago
Doesn't the human brain work like this? Yeah it's all connected together and plastic and so on, but functions tend to be localized, e.g vision is in occipital area. These base areas are responsible for the basic latent representations (edge detectors) which get fed forward to the AGI module (prefrontal cortex) that coordinates the whole thing based on the high quality representations it sees from these base modules. This strikes me as the most compute efficient approach.

Workaccount2 3y ago

I'd almost say anyone not doing that is being foolish. The goal of the service is to answer complex queries correctly, not to have a pure LLM that can do it all. I think some engineers feel that if they are leaning on an old school classically programed tool to assist the LLM, it's somehow cheating or impure.

uoaei 3y ago
Of course. But we must acknowledge that many have blinders on, assuming that scale is all you need to beat statistical errors.
- sigmoid10 3y ago
  Well, these people are not wrong per se. Scale is what drove what we have today and as hardware improves, the models will too. It's just that in the very short term it turns out to be faster to just code around some of these issues on the backend of an API rather than increase the compute you spend on the model itself.
  - uoaei 3y ago
    Monkey sees moon. Monkey climbs tree. "See? Monkey is closer to moon than before. To reach moon, monkey just needs taller tree." How long before monkey finds tall enough tree to reach moon?
    - sigmoid10 3y ago
      We're rapidly approaching the compute capacity of the human brain in individual server racks. This "moon" is neither unreachable nor is there any doubt that we will cross the threshold soon.
      - uoaei 3y ago
        I find it incredibly hard to believe we stumbled upon an efficient architecture that requires nothing but more compute not 10 years after the AI winter thawed. That's incredibly optimistic to the point of blind hope. What is your background and what makes you think we've somehow already figured everything out?

vidarh 3y ago

GPT has for some time output "analyzing" in a lot of contexts. If you see that, you can go into settings and tick "always show code when using data analyst" and you'll see that it does indeed construct Python and run code for problems where it is suitable.

Jabrov 3y ago

What if we used character tokens?

Der_Einzige 3y ago

I wrote a whole paper about ways to "fix" tokenization in a plug-and-play fashion for poetry generation: Filter the vocabulary before decoding. https://paperswithcode.com/paper/most-language-models-can-be-poets-too-an-ai-1 https://paperswithcode.com/paper/most-language-models-can-be...

data-ottawa 3y ago

ChatGPT definitely has a growing bag of tricks like that. When I use analysis mode to generate and evaluate code it recently started writing the code, then introspecting it and rewriting the code with an obvious hidden step asking "is this code correct". It made a huge improvement in usability. Fairly recently it would require manual intervention to fix.

bevekspldnw 3y ago

You can often see it write and execute python code to answer a question which is awesome.