Posts by frabcus@mastodon.social
(DIR) Post #ASgVtTfXyBrxt4Yium by frabcus@mastodon.social
2023-02-14T22:54:50Z
0 likes, 0 repeats
@simon ChatGPT was happy for me when using the same wording I used with New Bing.Is this perhaps because you did it in two shots, and that meant the second request was *just* about misleading, whereas in the first it was diluted by the rest of the request?Or just random variance!
(DIR) Post #ASgWKfJslGKbq7OEDo by frabcus@mastodon.social
2023-02-14T22:58:36Z
0 likes, 0 repeats
@simon I don't mind in practice - I just don't tend to ask it how it feels, as my mental model of the world is that that won't tell me anything useful or interesting. I've a very tool-centric view of this!
(DIR) Post #ASgWKhCrkJuJgw7eYy by frabcus@mastodon.social
2023-02-14T22:59:33Z
0 likes, 0 repeats
@simon for the record I asked it anyway:
(DIR) Post #AShA8fnLAJKC0M0N5E by frabcus@mastodon.social
2023-02-15T06:24:13Z
0 likes, 0 repeats
@djvdq @simon Fair!I think necessarily large language models will always end up threatening or offending in some situations people will screengrab. Google search results threaten and offend me sometimes.The question is how much does it, and is it more useful? I'm not sure yet.New Bing isn't so useful I'm using it all the time. But... I appreciate and would like a "ChatGPT with citations", but only if I know it is fundamentally a statistical model not a superintelligence.
(DIR) Post #AShAOlU1mcs2HSe2XQ by frabcus@mastodon.social
2023-02-15T06:29:03Z
0 likes, 0 repeats
@simon @djvdq That's a really good question, and I haven't seen anything about that either!I'm surprised to watch people doing stuff like thanking ChatGPT and saying they appreciate it.And maybe they're right - the underlying model might be intelligent enough just constrained that honouring it like that is ethically right.I wonder if it being wrong or stupid will seem normal to most people, as humans are often wrong and stupid. Especially super clever ones!
(DIR) Post #AUNzuWSXxFVNJymEyW by frabcus@mastodon.social
2023-04-06T20:15:47Z
0 likes, 0 repeats
@simon The answers are all disappointingly the same.e.g. I think the system for this should auto-canonicalise this, and consist of the input function from URL ---> dict of view/parameters which matches URLs loosely in a configurable way (e.g. with without /, case insensitive, even spelling variations etc)Then a map from the dict of view/parameters to the definitive output URL. (This would also be used for URL generation)And then the framework automatically redirects to the canonical URL.
(DIR) Post #AUVBFejyCFtOCOwppg by frabcus@mastodon.social
2023-04-10T07:26:13Z
0 likes, 0 repeats
@simon my memory is round about 2001 *everyone* got both old-style mobile phones, and also email. Those had an impact, but I can't remember it being negative in anyway, just new superpowers for everyone (was it just I was young?)Meanwhile, i'd seen DoCoMo phones everywhere in Japan in 2003. Which had an app ecosystem. And I had a Palm Treo, which had apps but not (for me yet) Internet.Then 2010 hit.
(DIR) Post #AUfgOWlthph9A5vQEi by frabcus@mastodon.social
2023-04-15T09:02:10Z
0 likes, 0 repeats
@simon @aebrer yes - some answers will be getting the UX right for that distinction. What specific things would you have made that you haven’t because of this? Is it stuff like LLM processing your own email - what specific simultaneous capability do you want to give the same prompt that the risk becomes solid for you?
(DIR) Post #AUi90IG3IlVCBYkKG0 by frabcus@mastodon.social
2023-04-16T13:32:04Z
0 likes, 0 repeats
@simon @aebrer the read only one still sounds useful! And can limit scope of it, so not auto-showing images.Also I'm imaginign something that separates out actions - so it wouldn't be able to do something except ask ME if it can do something, and I would confirm the things it does explicitly. The code for that confirmation would just be normal code, not an LLM (so it would nicely and succinctly describe the action).Limits the utility e.g. for updating a calendar. Good for sending mail!
(DIR) Post #AXAq0UqFXDnuiYuymG by frabcus@mastodon.social
2023-06-29T05:29:56Z
0 likes, 0 repeats
@simon nice! do you have any immediate take on how it compares in quality/capability to various OpenAI models?
(DIR) Post #AXOuZ9kF5HcrS5KV4S by frabcus@mastodon.social
2023-07-05T07:48:28Z
0 likes, 0 repeats
@feld @SlicerDicer @codinghorror @jerry I didn’t know that was a thing on Twitter. What’s the app called? Is it dead now with the Twitter API charges?
(DIR) Post #AZRq1hCCYVKAv5jxmi by frabcus@mastodon.social
2023-09-05T06:04:33Z
0 likes, 0 repeats
@simon Nice! I think I always did this with search engines and copying code anyway. I think with the right IDE it would be substantially quicker with an LLM. I tend to find my time more sucked into libraries / APIs for which even GPT-4 hallucinates a touch too much for me.
(DIR) Post #AZRwIE2ApZZ77lJb2u by frabcus@mastodon.social
2023-09-05T06:08:15Z
0 likes, 0 repeats
@simon I’m saying this as I’m super curious about *why* coders give such radically varied feedback upon first trialing eg Copilot. And how day to day programmers wildly differ in the internal mental actions of how we work.
(DIR) Post #AZSI1yUN3e0hGnA024 by frabcus@mastodon.social
2023-09-05T11:18:28Z
0 likes, 0 repeats
@simon Would love to see a video or blog post about how you use it! Particularly Copilot's UI felt inflexible to me - no chat when I was last using it. You work round that with comments?I'm anticipating much better IDE interfaces - my sense is none are quite there yet. (e.g. letting the LLM edit multiple files easily)
(DIR) Post #AZwwLdPgwzLm194PNw by frabcus@mastodon.social
2023-09-20T06:11:23Z
0 likes, 0 repeats
@simon @alys playing with it, it certainly looks that way. @alys what are you finding?
(DIR) Post #AZy7sREUYHhMRxdU1o by frabcus@mastodon.social
2023-09-20T19:55:15Z
0 likes, 0 repeats
@simon @alys seemingly in a similar format too
(DIR) Post #AZyMysT9hO4NyM1Uzg by frabcus@mastodon.social
2023-09-20T22:44:29Z
0 likes, 0 repeats
@simon amazing! presumably `top_logprobs` is not accidentally the same structure?It's so odd that isn't documented anywhere on OpenAI's site, that I can find. My memory is the logprobs fields used to be simpler in e.g. text-davinci-003
(DIR) Post #AaRLcgGJjuY2I0m1DM by frabcus@mastodon.social
2023-10-04T22:17:39Z
0 likes, 0 repeats
@killyourfm @thunderbird I assume this is strategically critical for them, so they can index it with LLM embeddings, and offer you AI with ability to read your email archive and write new messages for you?Pretty severe privacy loss. But alas I think will be popular.
(DIR) Post #AbNQ8ItKTMkkD9ObeC by frabcus@mastodon.social
2023-11-01T22:39:53Z
0 likes, 0 repeats
@simon @mhoye I don't know, but meanwhile could you link to the first message in the thread as it looks really interesting 😆
(DIR) Post #AbOPJg0gx6pYImeu2a by frabcus@mastodon.social
2023-11-02T10:05:06Z
0 likes, 0 repeats
@simon @mhoye Ah! This is super confusing - and turns out is different if logged into mastodon.social or not.First screenshot, logged in. Second screenshot in an incognito window.Now I understand what you were complaining about! Odd bug to do with authentication? I don't follow @mhoye