4 ms·

> We'll have a core directable glued-together stack that can take instruction like "make me a shopify site to sell cute mini traffic cones that are oscar-the-gr

> We'll have a core directable glued-together stack that can take instruction like "make me a shopify site to sell cute mini traffic cones that are oscar-the-grouch themed, but make sure they don't violate any copyright themes, make me 100 ads to sell them" by the end of 2024 lol I’d be VERY surprised if they can do that by the end of 2030. 1. You need relationships with manufacturers to do that 2. You need to validate the physical results prior to a full run 3. These LLMs have no idea about copyright and even the best ones tell you stuff that isn’t remotely true. These models predict the next construct (phrase, word, etc.) and they’re decent at it. They’ll eventually do images, with text, hopefully both actually relate and make sense. I’d give that 2 years really to refine. They’re really good predictors, but when you start mixing things and gluing 4-5 things together, the error rate grows exponentially and the understanding equally had a massive drop.

vessenes 2y ago
I think you're very wrong about this timeline. 1. I've never sourced hardware, so I can't comment directly on manufacturers. I have done plenty of business in Asia, and, I have purchased from alibaba. One side of those alibaba transactions is definitely using AI translation already. I don't think manufacturers in Asia will care at all "who" is sourcing product from them. For speccing a car - sure, you're right that a relationship is needed (and probably will involve doing some drinking near the factory). For speccing injection mold / printable goods? I'm skeptical that this is cut and dried in favor of humans with IRL relationships. 2. Trivially done by an AI; "Mr. Bossman do you want to validate this sample, or have me hire someone to do it?" 3. This is demonstrably untrue; OpenAI disagrees with you about this, for instance. ChatGPT pre-instructions include multiple paragraphs about this topic already. (here's a link to info about it if you're curious: https://x.com/krishnanrohit/status/1755122786014724125/photo/1 https://x.com/krishnanrohit/status/1755122786014724125/photo...) Generative AI that can spell correctly is here now, although not widely commercially deployed, but is in no way needed to AI CEO things. My experience as the 'gluer' is that the most recent visual/multimodel models are good enough to, say, fix a CSS bug in HTML they wrote by looking at a screenshot of the output. So, I'd say the opposite is true in my experience -- with multimodal feedback, they are able to self correct. Anyway, to me this doesn't add up to 2030. Maybe for the snazzy mac app version. But, I think we are VERY close right now, and it is a major engineering effort at OpenAI and Microsoft to get there quickly.
- sarchertech 2y ago
  $1000 to the charity of your choice that this isn’t available in any generalizable way before 2030.
  - vessenes 2y ago
    Deal! I’m at a graduation right now but I’m responding to keep this in my threads. let’s set up a long bet. If your email isn’t in profile mail me
  - vessenes 2y ago
    https://longbets.org/951/ https://longbets.org/951/
- raincole 2y ago
  > Trivially done by an AI; "Mr. Bossman do you want to validate this sample, or have me hire someone to do it?" I honestly don't understant this sentence. If you hire someone to do it then it's not "trivially done by an AI", no?

vessenes 2y ago

I think you're very wrong about this timeline. 1. I've never sourced hardware, so I can't comment directly on manufacturers. I have done plenty of business in Asia, and, I have purchased from alibaba. One side of those alibaba transactions is definitely using AI translation already. I don't think manufacturers in Asia will care at all "who" is sourcing product from them. For speccing a car - sure, you're right that a relationship is needed (and probably will involve doing some drinking near the factory). For speccing injection mold / printable goods? I'm skeptical that this is cut and dried in favor of humans with IRL relationships. 2. Trivially done by an AI; "Mr. Bossman do you want to validate this sample, or have me hire someone to do it?" 3. This is demonstrably untrue; OpenAI disagrees with you about this, for instance. ChatGPT pre-instructions include multiple paragraphs about this topic already. (here's a link to info about it if you're curious: https://x.com/krishnanrohit/status/1755122786014724125/photo/1 https://x.com/krishnanrohit/status/1755122786014724125/photo...) Generative AI that can spell correctly is here now, although not widely commercially deployed, but is in no way needed to AI CEO things. My experience as the 'gluer' is that the most recent visual/multimodel models are good enough to, say, fix a CSS bug in HTML they wrote by looking at a screenshot of the output. So, I'd say the opposite is true in my experience -- with multimodal feedback, they are able to self correct. Anyway, to me this doesn't add up to 2030. Maybe for the snazzy mac app version. But, I think we are VERY close right now, and it is a major engineering effort at OpenAI and Microsoft to get there quickly.

sarchertech 2y ago
$1000 to the charity of your choice that this isn’t available in any generalizable way before 2030.
- vessenes 2y ago
  Deal! I’m at a graduation right now but I’m responding to keep this in my threads. let’s set up a long bet. If your email isn’t in profile mail me
- vessenes 2y ago
  https://longbets.org/951/ https://longbets.org/951/
raincole 2y ago
> Trivially done by an AI; "Mr. Bossman do you want to validate this sample, or have me hire someone to do it?" I honestly don't understant this sentence. If you hire someone to do it then it's not "trivially done by an AI", no?