Post ASfpBYKv3kplXvHccC by ggdupont@mastodon.social
(DIR) More posts by ggdupont@mastodon.social
(DIR) Post #ASfftK5kNovwZDQ2xE by TedUnderwood@sigmoid.social
2023-02-14T13:14:08Z
0 likes, 0 repeats
The problem with using #LLM for search may turn out to be mostly a problem with the internet, which contains lots of false, unverified, or conflicting information. Knitting that into a single judicious account, for all possible questions, is a superhuman task. #Bard might be well advised just to decline some missions.
(DIR) Post #ASfgl6fqUCoUadwx60 by scott_bot@hcommons.social
2023-02-14T13:23:49Z
0 likes, 0 repeats
@TedUnderwood "search" and LLMs as currently construed are both limited to the internet.It's clear that a ChatGPT-fueled hype train tipped off a fear-of-being-scooped feedback loop between Microsoft and Google, neither of whom had a product anywhere near ready for rollout. Before the betas I still afforded some room to the possibility LLM search (at least in its current form) could avoid being a disinformation noise machine, but now? We're a long way out. Trying to talk the machine into accuracy and out of being susceptible to prompt engineering is a fool's errand.
(DIR) Post #ASfh1RKbRyjDI6nn7o by TedUnderwood@sigmoid.social
2023-02-14T13:26:49Z
0 likes, 0 repeats
@scott_bot “Accuracy” is not something the internet currently has and could lose. This looks to me like a rhetorical problem; they need to find a way to position these models as search assistants or explicators rather than oracles.
(DIR) Post #ASfhYQwvtHUXfBSmHo by scott_bot@hcommons.social
2023-02-14T13:32:46Z
0 likes, 0 repeats
@TedUnderwood Indeed. LLMs are a solution seeking a "search" problem.
(DIR) Post #ASfi1UN6FLXsfsYbc8 by scott_bot@hcommons.social
2023-02-14T13:30:51Z
0 likes, 0 repeats
@TedUnderwood Put another way: as long as the interface between product owner and LLM rhymes with the interface between you and I, and as long as LLMs are trained on *gesticulates tiredly*, I'll trust LLMs for search about as much as I trust a random redditor (even/especially one with great karma).
(DIR) Post #ASfi1Uu4GluuK7yw08 by TedUnderwood@sigmoid.social
2023-02-14T13:38:01Z
0 likes, 0 repeats
@scott_bot Yes. Agreed. I think that’s a marketable product though. There are lots of queries that don’t require “trust,” or require no more trust than that. There’s also the Galactica path of training on better sources, and although Galactica itself was oversold I think that can work.
(DIR) Post #ASfiTkNiJd2oR8JpoG by afamiglietti79@mastodon.social
2023-02-14T13:39:58Z
0 likes, 0 repeats
@scott_bot @TedUnderwood at least some of the false info we're seeing in these early examples is straight up hallucinated by the llm. That's a genuine new problem, above and beyond issues of source reliability. It's also potentially difficult to resolve, given that accurate factual statements from LLMs are basically an accidental finding. The proper role for the current generation llm is far away from public facing search in any capacity.
(DIR) Post #ASfiTmMMxb9oZXhmzY by TedUnderwood@sigmoid.social
2023-02-14T13:43:07Z
0 likes, 0 repeats
@afamiglietti79 @scott_bot Yes, hallucination is a new problem. How difficult to resolve we shall see. I apply a huge discount for the incentive structure around commentary rn. The (non-cherry-picked) examples of Bing I’ve seen have been good stuff: a service I would pay for.
(DIR) Post #ASfiduNpz1TiLxDpT6 by TedUnderwood@sigmoid.social
2023-02-14T13:44:58Z
0 likes, 0 repeats
@afamiglietti79 @scott_bot but then I would also pay for a model that argues fiercely with me about whether it’s 2022 or 2023, so maybe I’m not the best test case
(DIR) Post #ASfj84Vysb76SjPjma by afamiglietti79@mastodon.social
2023-02-14T13:50:24Z
0 likes, 0 repeats
@TedUnderwood @scott_bot I really need to finish the generative text chapter in the chollet textbook. My instinct after working through the first few chapters is hallucinations could be very difficult to stamp out, but my confidence in that instinct is low. In any event, even rare hallucination could be problematic, especially if they crop up in unexpected ways. And there is still Scott's point: users feeling like they are speaking to a hyper intelligent...
(DIR) Post #ASfjGEWVYIBvotYIr2 by scott_bot@hcommons.social
2023-02-14T13:51:52Z
0 likes, 0 repeats
@TedUnderwood Even then, you and I both know how science works: disagreements, supercessions, a sprinkling of vaccines-cause-autism publications, etc. I'm not sure "trust" should be a problem that scales, but even if it is / should be, certainly not with this type of technology.Google Translate showed rule systems weren't as good as LLM-style kitchen sink approaches for good-enough translation. And that's basically been our approach to everything for the last two decades (chess, etc. "end of theory" stuff), but we've been seeing cracks in the " all you need is scale" approach for years now.
(DIR) Post #ASfjSvvtyx7rvYUCCO by afamiglietti79@mastodon.social
2023-02-14T13:54:11Z
0 likes, 0 repeats
@TedUnderwood @scott_bot just pay for it as a game! I will be right there with you! We can roll characters together, the AI can't possibly be more abusive DM then some of the guys I played with in college
(DIR) Post #ASfkmeLx9Y5p9NdX8q by TedUnderwood@sigmoid.social
2023-02-14T14:08:50Z
0 likes, 0 repeats
@scott_bot I think you’re conflating lots of different things there. What Bing is doing now is not “All you need is scale”; it’s mixing *retrieval* with an autoregressive language model. I get that people want to see “this tide” go back out, but it’s not a single tide; it’s quite adaptable. “Oracle” is not a job it should attempt—and maybe not one we should attempt either. That seems okay to me.
(DIR) Post #ASflkqEY6texN8sJjU by scott_bot@hcommons.social
2023-02-14T14:19:49Z
0 likes, 0 repeats
@TedUnderwood That's fair, I was uncarefully pithy. The point I'm trying to make is when expensive solutions search for lucrative problems, they creep into domains in which they're not appropriate.
(DIR) Post #ASfpBYKv3kplXvHccC by ggdupont@mastodon.social
2023-02-14T14:58:15Z
0 likes, 0 repeats
@TedUnderwood 100%!The ability to assess trustability of a source (and thus the reliability of the information it shares) has been a problem for intelligence analysts for ever...By design, the #LLM approach seems to be really not made to address it.