Post ATo445j5iFmlVeyuEy by scott_bot@hcommons.social
 (DIR) More posts by scott_bot@hcommons.social
 (DIR) Post #ATo43u15FX31BzJ6cS by scott_bot@hcommons.social
       2023-03-20T12:10:08Z
       
       0 likes, 1 repeats
       
       Prompted by @TedUnderwood's call to respond to what AI might look like in 3-5 years, @emckean & Vivek Seshadri's ideas of a data co-op, & @ct_bergstrom's discussion of LLM's propensity for bullshit, I've been reflecting on one possible future for large language models.Let's assume from the current state of text-to-image models that LLMs will keep improving. By improving, I mean at meeting benchmarks and passing for human, not at being better for society. The bullshit will become more difficult to identify.It's a problem for a million oft-discussed reasons. One of them is false information presented confidently through an AI assistant. The training set (the internet) is awash with enough false information that one of these models will eventually give a parent advice that will kill their kid.As they say: garbage in, garbage out. Companies may respond by doubling down on what Meta did with Galactica, training models on scientific articles and other trusted sources.1/n
       
 (DIR) Post #ATo445j5iFmlVeyuEy by scott_bot@hcommons.social
       2023-03-20T12:12:22Z
       
       0 likes, 0 repeats
       
       Galactica failed because LLM tech wasn't quite there yet, still making stuff up unwittingly, but I expect that will be fixed. The next Galactica will fail because because even if it doesn't outright make stuff up, the sum of scholarly & reference sources has its own issues.A model trained on scientific articles will lie succinctly, and one trained on humanities will bullshit like a high school debate champion; both will do so with the whiff of authority. Encyclopedias and other trusted sources will come with their own flavors of problems.As one example of many (pushing aside Big Stuff like geographic skew): science disagrees, science changes. Scientific consensus now is different from thirty years ago. A model trained on All Of It would lead us astray (see COVID and wrong ideas about airborne transmission).2/n
       
 (DIR) Post #ATo4dFDaPju15F0WnI by TedUnderwood@sigmoid.social
       2023-03-20T12:21:09Z
       
       0 likes, 0 repeats
       
       @scott_bot I quite like this idea. I would tend to present it as a positive contribution to human knowledge rather than a remediation for LLMs. (Since it’s not like we had well-organized consensus about scientific credibility *before* LLMs came around.) But that’s immaterial; a good idea is a good idea, and one should never let a crisis go to waste, even if you suspect it’s a crisis of rising standards.
       
 (DIR) Post #ATo6v1Hl9c6YzcYyem by scott_bot@hcommons.social
       2023-03-20T12:46:45Z
       
       0 likes, 0 repeats
       
       @TedUnderwood I should have pre-registered the response I expected from you, since apparently my internal ted_underwood model meets benchmarks admirably. 🤖