← Feed

Posts

• View post

  @hotdogsladies time should be behind that stoplight imo

• View post

  @caseyliss sorry to butt in to this conversation, but you need to not think about LLMs in terms of "out of date world knowledge" All the modern LLMs are hooked up to search engines and are going to search to answer *any* question like this. https://chatgpt.com/share/685cd122-f4c8-800c-acab-540aa685d07b

• View post

  @daringfireball @gruber you were unsure if "hold for me" was a never before seen feature, but indeed Android has had it for like 3 years https://overcast.fm/+AAB7NBV4mEM/33:02 https://support.google.com/pixelphone/answer/10071878?hl=en

• View post

  @gruber yes of course, Pixel 😆

• View post

  @siracusa at least use one that's hooked up to a search engine and cites sources!

• View post

  @siracusa both of these provide website snippets that I find are quite satisfactory (Gemini’s UI attached) But I think the important part is that I have much much more confidence in an LLM’s ability to summarize content that’s loaded into the context window.

• View post

  @siracusa I’m not sure how likely that really is. That was your whole point back in episode 589. If the answer is in the context window, you shouldn’t be surprised if it gets it right. You should expect it to be right basically every time! Pulling stuff out of the context window is quite easy.

• View post

  @siracusa that “sources” view shows snippets, not ai summaries.

• View post

  @siracusa I mean, yea, I checked before telling you they’re snippets. But one doesn’t need to do that each time. You wouldn't re-read the terms and conditions every time you open an app

• View post

  @siracusa on desktop you still get the standard snippet for that. Would you have really admonished @marcoarment for posting this??

• View post

  @siracusa that’s a “featured snippet”. It’s not generated and never has been as far as I know.

• View post

  @siracusa but why does it end there? You’re deciding whether to believe the Google snippet is really on the NIH website. Why do you think it's accurate just because it’s published on the NIH website, you haven’t looked at the spreadsheet they loaded the results. Maybe they made a math mistake. And that’s only one study. You really want a literature review or meta analysis. No such analysis exists, but a search results screenshot can fit a bunch of snippets from multiple sources all next...

• View post

  @zhaozilong you're responding to a screenshot of the snippets view though. I think a bunch of snippets of authoritative websites should be strong evidence, possibly stronger than a screenshot of a quote on a single authoritative website.

• View post

  @marcoarment @gruber @siracusa Dye's team needs someone who owns this shirt https://cottonbureau.com/p/29CRME/shirt/stuff-with-computers-command-key-white-logo

• View post

  @atpfm I don’t think the “pre populate phone data” idea was so crazy. They wouldn’t restore the data decrypted, they’d just put the encrypted data from iCloud servers onto the iPhone, and you’d be decrypting it on first unlock. Apple could already do that, this doesn’t present extra danger? If there’s an issue it’s that it would be a day old backup. Maybe incremental restore would fix that issue? I’m not saying this would be easy, but it’s totally possible!

• View post

  @atpfm @siracusa regarding accidental input: the solution of this problem used to be having a folio case that would lock the phone when it closed. I’m so upset that Apple got rid of this feature back around the 12. But it just used the compass (and magnets in the folio). It was free! Every once in a while, I send Tim an email about it. Can’t hurt right?

• View post

  @siracusa I knew there was a second ‘real’ key that your own password is actually just decrypting, but somehow I thought that the real key could be extracted. But in retrospect: no of course not, that would be insecure!