2 ms·

unless you have good reason to believe that they will not improve in quality I do. These LLMs can't tell facts from BS now. With LLMs starting to generate eve

unless you have good reason to believe that they will not improve in quality I do. These LLMs can't tell facts from BS now. With LLMs starting to generate ever more BS, going forward it will only become harder to tell the difference --- for both them and us. What do you think will happen when LLMs start sucking in each others BS from the internet? I don't think it's possible for this sort of global BS feedback loop to produce an improvement in quality.

idopmstuff 3y ago
> These LLMs can't tell facts from BS now. I think this is severely overstated as a problem that would prevent LLMs from being useful. LLMs generate BS if you use them like a search engine, but there are lots of use cases where they don't. Need to summarize a document or rephrase something in a different tone? Hallucination is really a non-issue. I've used LLMs to write scripts - while hallucination does happen, it's a situation where I can easily validate that the scripts are doing what I would expect, so hallucination isn't a serious issue.
- MattPalmer1086 3y ago
  Exactly. There are many, many cases where AI significantly speeds up doing something. I've already used it productively for writing code, understanding how to set up new tech, summarising documents and academic papers, and writing clearer text. And I'm not a power user at all. So this will result in fewer existing jobs, and the productivity of the remaining jobs must increase. And some new jobs will be created along the way.
- jqpabc123 3y ago
  while hallucination does happen Once new LLMs start to consume hallucinations from older LLMs, the hallucination problem will only start to grow and expand. With LLMs producing more content and producing it faster (this is their purpose), it will become prohibitively expensive if not impossible to check and verify everything that is fed into these models. LLMs are multipliers --- garbage in equals orders of magnitude more garbage out for the next generation of LLMs to consume.

idopmstuff 3y ago

> These LLMs can't tell facts from BS now. I think this is severely overstated as a problem that would prevent LLMs from being useful. LLMs generate BS if you use them like a search engine, but there are lots of use cases where they don't. Need to summarize a document or rephrase something in a different tone? Hallucination is really a non-issue. I've used LLMs to write scripts - while hallucination does happen, it's a situation where I can easily validate that the scripts are doing what I would expect, so hallucination isn't a serious issue.

MattPalmer1086 3y ago
Exactly. There are many, many cases where AI significantly speeds up doing something. I've already used it productively for writing code, understanding how to set up new tech, summarising documents and academic papers, and writing clearer text. And I'm not a power user at all. So this will result in fewer existing jobs, and the productivity of the remaining jobs must increase. And some new jobs will be created along the way.
jqpabc123 3y ago
while hallucination does happen Once new LLMs start to consume hallucinations from older LLMs, the hallucination problem will only start to grow and expand. With LLMs producing more content and producing it faster (this is their purpose), it will become prohibitively expensive if not impossible to check and verify everything that is fed into these models. LLMs are multipliers --- garbage in equals orders of magnitude more garbage out for the next generation of LLMs to consume.