4 ms·

Ask HN: What needs to happen for ChatGPT to iron out inaccuracies?

3 points by dizzydes 4y ago

smoldesu 4y ago
They'd have to stop training ChatGPT on inaccurate or contradictory information.
- PaulHoule 4y ago
  But the commonsense domain is all about inaccurate information. In your theory of mind, for instance, you have to model wrong beliefs that people have: https://en.wikipedia.org/wiki/Phlogiston_theory https://en.wikipedia.org/wiki/Phlogiston_theory A general purpose chatbot should be able to talk about Star Trek, Doctor Who , Don Quixote and other fiction and that is a myriad of worlds that are more or less internally consistent but yet need to be understood in how they derive from the "base" knowledge base (James Tiberius Kirk inherits nearly all of the attributes of a Pᴇʀsᴏɴ as a fictional human character) and influence the real world (what does it mean when a person says their boss is like "Captain Ahab?")
  - smoldesu 4y ago
    Right. It would seem that without a heuristic model, mixing nonfiction with our fiction is a recipe for disaster.
qualudeheart 4y ago
Homans paid $15/hour to correct it. Deepmind self play mechanisms. Fine tuning on blog posts by Gary Marcus.
theGeatZhopa 4y ago
They need bard as fact checker.. joke. The problems I see is beside wrong answers, e.g. some Excel logic, that it loses its state in the middle. As LLM work in a 3d graph space, it seems to often take some wrong shortcut in that graph space causing it to lose some important parts of the information. They need to clean the rubbish connection between the graphs out, whats may be impossible. Or to train it even more to make certain connections in the graph more "used". Refine..
PaulHoule 4y ago
A wider attention window. ChatGPT can see 4096 sub-word tokens which is not a lot in the grand scheme of things. You could get "explainable" explanations of topic if it went and read 10 articles and was able to attend to the full text of all the articles to quite literally link the part of the article to the output with both generated text and a citation. That would take an attention window in the 400,000-4,000,000 tokens range. There are very long-range transformers but as of yet they don't work so well. They will be a continuing research topic because a longer attention window allows applying LLMs to more problems. For instance, document classification or retrieval of documents longer than 4000 tokens.
sp332 4y ago
Well it’s a language model. It’s fundamentally not about learning facts unless the facts are statistics about the corpus text. Hook it up to a database and have it translate questions into queries instead.
sp332 4y ago
Include a confidence score next to the output. If it’s just babbling to fill space, that should get a lower score than if it actually recognizes something from the prompt.

3 points by dizzydes 4y ago

smoldesu 4y ago

They'd have to stop training ChatGPT on inaccurate or contradictory information.

PaulHoule 4y ago
But the commonsense domain is all about inaccurate information. In your theory of mind, for instance, you have to model wrong beliefs that people have: https://en.wikipedia.org/wiki/Phlogiston_theory https://en.wikipedia.org/wiki/Phlogiston_theory A general purpose chatbot should be able to talk about Star Trek, Doctor Who , Don Quixote and other fiction and that is a myriad of worlds that are more or less internally consistent but yet need to be understood in how they derive from the "base" knowledge base (James Tiberius Kirk inherits nearly all of the attributes of a Pᴇʀsᴏɴ as a fictional human character) and influence the real world (what does it mean when a person says their boss is like "Captain Ahab?")
- smoldesu 4y ago
  Right. It would seem that without a heuristic model, mixing nonfiction with our fiction is a recipe for disaster.

qualudeheart 4y ago

Homans paid $15/hour to correct it. Deepmind self play mechanisms. Fine tuning on blog posts by Gary Marcus.

theGeatZhopa 4y ago

They need bard as fact checker.. joke. The problems I see is beside wrong answers, e.g. some Excel logic, that it loses its state in the middle. As LLM work in a 3d graph space, it seems to often take some wrong shortcut in that graph space causing it to lose some important parts of the information. They need to clean the rubbish connection between the graphs out, whats may be impossible. Or to train it even more to make certain connections in the graph more "used". Refine..

PaulHoule 4y ago

A wider attention window. ChatGPT can see 4096 sub-word tokens which is not a lot in the grand scheme of things. You could get "explainable" explanations of topic if it went and read 10 articles and was able to attend to the full text of all the articles to quite literally link the part of the article to the output with both generated text and a citation. That would take an attention window in the 400,000-4,000,000 tokens range. There are very long-range transformers but as of yet they don't work so well. They will be a continuing research topic because a longer attention window allows applying LLMs to more problems. For instance, document classification or retrieval of documents longer than 4000 tokens.

sp332 4y ago

Well it’s a language model. It’s fundamentally not about learning facts unless the facts are statistics about the corpus text. Hook it up to a database and have it translate questions into queries instead.

sp332 4y ago

Include a confidence score next to the output. If it’s just babbling to fill space, that should get a lower score than if it actually recognizes something from the prompt.