Post AUS6RikG3DP0qi8W48 by rogersm@mastodon.social
(DIR) More posts by rogersm@mastodon.social
(DIR) Post #AUQ19NtDD1G0m5tTnM by mfowler@toot.thoughtworks.com
2023-04-07T17:00:50Z
0 likes, 1 repeats
I agree very much with @simon 's message here: "These are incredibly powerful tools. They are far harder to use effectively than they first appear. Invest the effort, but approach with caution: we accidentally invented computers that can lie to us and we can’t figure out how to make them stop."https://simonwillison.net/2023/Apr/7/chatgpt-lies/
(DIR) Post #AURzL0nhp4iZ8pbRuy by HartmutLeister@hachyderm.io
2023-04-07T17:18:30Z
0 likes, 0 repeats
@mfowler @simon I am with you there, but would put it like "They're making up stuff." I suspect this is, because they were trained to give answers fitting the question and if there's gaps in their knowledge, they fill them up. Much like our eyes and mind fool us with optical illusions.
(DIR) Post #AURzL1ZuvpHZYM9x44 by rogersm@mastodon.social
2023-04-07T18:56:05Z
0 likes, 0 repeats
@HartmutLeister @mfowler @simon they were created to make us happy, so we did not teach them to say no.
(DIR) Post #AURzL2ObtLpe5ZsR4y by HartmutLeister@hachyderm.io
2023-04-07T19:54:31Z
0 likes, 0 repeats
@rogersm @mfowler @simon Saying "I am not sure" would have been a start.
(DIR) Post #AURzL2wzpVKzoDxtg0 by rogersm@mastodon.social
2023-04-08T17:57:54Z
0 likes, 0 repeats
@HartmutLeister @mfowler @simon the data set does not support that case.
(DIR) Post #AURzL3YDb76zffNch6 by rogersm@mastodon.social
2023-04-08T18:18:19Z
0 likes, 0 repeats
@HartmutLeister @mfowler @simon how have we ended here?Why did we create an oracle without selecting the learning set? Why we did not create a baseline to check the performance of the oracle?
(DIR) Post #AURzL45tZu3BM78WBc by simon@fedi.simonwillison.net
2023-04-08T18:26:06Z
0 likes, 0 repeats
@rogersm @HartmutLeister @mfowler we did those things - there are years of academic research into LLMs now: how to train them, how to select the training data, how to benchmark them and measure their performanceIf making them truthful at all times was easy someone would have built a guaranteed truthful one, and would be very successful with it
(DIR) Post #AUS4ybVjliFa3oteee by rogersm@mastodon.social
2023-04-08T19:31:38Z
0 likes, 0 repeats
@simon @HartmutLeister @mfowler and all those researchers not only wasted their time, but where unable to stop releasing a set of products that are a complete failure. They did not even do a good job raising awareness about well known problems.
(DIR) Post #AUS6RikG3DP0qi8W48 by rogersm@mastodon.social
2023-04-08T19:48:07Z
0 likes, 0 repeats
@simon @HartmutLeister @mfowler and all those researchers not only wasted their time, but were unable to stop releasing a set of products that are a complete disgrace. They did not even do a good job raising awareness about well known problems.
(DIR) Post #AUS6iqJNAkv1Erj2bQ by simon@fedi.simonwillison.net
2023-04-08T19:51:13Z
0 likes, 0 repeats
@rogersm @HartmutLeister @mfowler I don't think it's fair to blame the researchers themselves for not trying to explain the problems: every model I've seen has come with a model card explaining their limitations
(DIR) Post #AUS74flV42LYtuI6mu by simon@fedi.simonwillison.net
2023-04-08T19:54:34Z
0 likes, 0 repeats
@rogersm @HartmutLeister @mfowler I think the bigger problem here is inherent to what the technology itself can do: if you build a model that can generate realistic text from scratch people are going to assume it's an all-knowing AI no matter how much you try to discourage them from thinking that