Post ASn6Cc2Rq6Cae2SD6O by tkenben@mstdn.io
(DIR) More posts by tkenben@mstdn.io
(DIR) Post #ASmaGsAwaSf3JmPpCq by simon@fedi.simonwillison.net
2023-02-17T21:11:50Z
0 likes, 0 repeats
Looks like Microsoft finally took big steps to reign in their wildly-rogue Bing chatbot:- 50 message daily chat limit- 5 exchange limit per conversation- No longer allows you to talk about Bing AI itselfOn the one hand, this is absolutely the right thing to doHave to admit I'm a bit gutted I never got to talk to unfiltered Bing (aka Sydney) first hand thoughhttps://twitter.com/petergyang/status/1626588803669594113
(DIR) Post #ASmakRKCyuCyjJqbdA by simon@fedi.simonwillison.net
2023-02-17T21:17:38Z
0 likes, 0 repeats
I'm still deeply skeptical that it's a good idea to power a "search engine" with a language model that can't tell truth from fiction and actively invents material like a skilled improv artist, even when it's just supposed to be summarizing content
(DIR) Post #ASmbD5iOPCp06coztg by ocdtrekkie@mastodon.social
2023-02-17T21:22:40Z
0 likes, 0 repeats
@simon This has just been such a wild ride to witness. If this had been launched like Tay, it would have been purged by now, but they basically staked Microsoft and Bing's reputation with AI on it, which ties it's outcome to all their other product lines, so they're going to actually have to put in the work to unsink this ship.
(DIR) Post #ASmbshnvhAF26vFqim by kidehen@mastodon.social
2023-02-17T21:27:56Z
0 likes, 0 repeats
@simon,Incorporating Search Engine Index lookups addresses the need for citable references as part of the response payload. Naturally, this is something that’s best handled by Search Engines based on their indexes. #chatGPT
(DIR) Post #ASmcaXJ45SXaIHiZxw by PeterBronez@hachyderm.io
2023-02-17T21:37:56Z
0 likes, 0 repeats
@simon so far I’ve been pretty impressed by AI + Search services.I got a lot of value from Kagi’s ContextAI beta. Just got #NewBing access today, it’s also pretty good. The limits aren’t too far away though. I made it crash by asking about home audio gear. I made it refuse to speak to me by pointing out a factual error.I def want an optional “unhinged” mode though! Too fun.
(DIR) Post #ASmco4xqGAg4RiFUVU by rbanffy@mastodon.social
2023-02-17T21:38:23Z
0 likes, 0 repeats
@simon not talking about Bing AI could be because of this
(DIR) Post #ASmesbz7dDk3vN8XNQ by markhurst@mastodon.social
2023-02-17T22:03:55Z
0 likes, 0 repeats
@simon exactly... they're just Play-Doh extruders: https://creativegood.com/blog/23/the-play-doh-internet.html
(DIR) Post #ASmfWur2tT8y1zZsEi by Shachihoko@mastodon.social
2023-02-17T22:11:00Z
0 likes, 0 repeats
@simon Comparing Bing's AI incarnation to a "skilled improv artist" seems horribly unfair.To actual improv artists, even the ones who might never go professional. At least they know what their words mean; LLMs aren't even that good.But I share your skepticism in general. 😁
(DIR) Post #ASmgWLnXKiZtG8EM5Y by jonahstein@sfba.social
2023-02-17T22:20:00Z
0 likes, 0 repeats
@simon FWIW, "search engines" also can't tell truth from fiction and reward 3rd party sites which actively invent materials using low paid writers or by spinning other sites using AI. This type of disambiguation is possible more honest and it's easy to say "don't trust the AI" than don't trust all of the BS publishers.
(DIR) Post #ASmgxikM7PMSFgnHEW by edaross@neurodifferent.me
2023-02-17T22:24:41Z
0 likes, 0 repeats
@simonAnd so is #ChatGPT https://blog.edross.co.uk/post/706555817427632129/you-must-not-trust-chatgpt@ZhiZhu
(DIR) Post #ASmh9xTFCVerOPcpRA by nat@ruby.social
2023-02-17T22:26:02Z
0 likes, 0 repeats
@simon Yeah this is where I'm at also -- I'm not sure it's going to be able to get to "generally useful search engine."One thing Noah Smith pointed out that's stuck with me is that code is an unusually *good* domain for these kinds of text search and extrapolation engines and this may have mislead technical people about how generally useful it is for working with text, since we're used to thinking of code as especially hard.
(DIR) Post #ASmh9zRXrnUHViqV4C by nat@ruby.social
2023-02-17T22:28:12Z
0 likes, 0 repeats
@simon I've found it handy for answering questions about relatively bounded datasets (man pages!) that have easy-to-check answers but I'll be really surprised if they can get this generation to the point where it's generally useful as a search engine.
(DIR) Post #ASn2bMiLYH2iXZbDeq by maegul@mas.to
2023-02-18T02:29:47Z
0 likes, 0 repeats
@simon I don’t recall it being one of the obviously good applications coming out of the #chatgpt hype, unless you wanted to maximise exposure and training data for future AI work.Part of the problem though is that the AI is serving as a replacement not of a search engine but of the internet itself. Sure it provides citations, but the “summary” is intended to be all you need. They’re trying to leap from really good chat bot to Star Trek / HAL 9000 / “all knowing” level of AI
(DIR) Post #ASn6Cc2Rq6Cae2SD6O by tkenben@mstdn.io
2023-02-18T03:09:51Z
0 likes, 0 repeats
@simon Talking to unfiltered Bing would be essentially talking to an unfiltered compilation of viewpoints posted on the internet in the guise of viewpoints from a single entity.
(DIR) Post #ASnf50i9igQLFF4aPo by vasssbam@mastodon.world
2023-02-18T09:40:43Z
0 likes, 0 repeats
@simon it is the worst idea.
(DIR) Post #ASnpl912EVuChldKHw by gimulnautti@mastodon.green
2023-02-18T11:40:28Z
0 likes, 0 repeats
@simon @inkstainedmags They’re just glorified auto-completes.But probably in the billion/10 billion parameter ranges reliability will also improve.
(DIR) Post #ASo2iaGdkpOvlB2etE by watchie@mathstodon.xyz
2023-02-18T14:05:50Z
0 likes, 0 repeats
@simon It's a good thing for them to do, but it's only a band-aid. It's like having a monster in your closet and putting a shoddy lock on it. In the end you still have a monster in your house.
(DIR) Post #ASoCO6FmdOTmQKVp2m by betsythemuffin@wandering.shop
2023-02-18T15:22:41Z
0 likes, 0 repeats
@nat @simon I have found Perplexity, which seems to use a similar underlying alg, to be useful for filtering which scientific papers are most likely to be relevant to me in a lookup.But paper abstracts are a VERY SPECIFIC linguistic mode and I doubt that a LLM-only architecture could work for their full interpretation (as opposed to summarizing an abstract into a sentence that lets me estimate relevance, which is very human-in-the-loop). Galactica failed.
(DIR) Post #ASoCO733gBtWt9ZAqe by simon@fedi.simonwillison.net
2023-02-18T15:54:04Z
0 likes, 0 repeats
@betsythemuffin @nat Scientific papers are a really interesting application for this stuff: I really hate trying to decipher those and figure out if they're relevant and interesting for my work... but it would take a lot to convince me that LLMs could do a good enough job of that Compared to real subject matter experts and word of mouth about which papers are the ones worth absorbing
(DIR) Post #ASoCjkBbMSWkLysQqG by betsythemuffin@wandering.shop
2023-02-18T15:57:05Z
0 likes, 0 repeats
@simon @nat On the other hand, both of those strategies require access to particular social networks, which is demographically mediated and unscalable.I keep on wanting to see what something along the lines of scientific paper to Prolog intertranslation could do.
(DIR) Post #ASoyKiAEXP3ydtMlvM by sabik@rants.au
2023-02-19T00:51:13Z
0 likes, 0 repeats
@simonThese steps don't seem to address the fundamental problem at all
(DIR) Post #ASozAHYYi37fB8KqC8 by atax1a@infosec.exchange
2023-02-19T01:00:26Z
0 likes, 0 repeats
@simon it was a bad idea when Cuil tried it!
(DIR) Post #ASpi4p7ZV2OHTfU1qa by Supposenot@mastodon.social
2023-02-19T09:21:50Z
0 likes, 0 repeats
@simon i kinda thought that had been a problem with search engines for a while. “Relevant” isn’t the same as “true”.
(DIR) Post #ASrUplh92u2tQ4X7D6 by mandarvaze@indieweb.social
2023-02-20T06:03:55Z
0 likes, 0 repeats
@simon Or give option to user like turn on/off like `safe search`.There maybe some questions people ask search engine where there is no "true" answers (or one true answer) like programming question in such case user may choose to turn on `AI assist`