4 ms·

The difference between Theranos and OpenAI is the later has created a product which, while imperfect, does something. While conversational AI is nowhere near a

The difference between Theranos and OpenAI is the later has created a product which, while imperfect, does something. While conversational AI is nowhere near as good as people are made to think, there are two places where it has really shown itself to be good: 1. Writing the first draft of things. Be it code or a email. 2. Turning unstructured data into structured data By and large, those use cases aren’t very sexy, but the have the potential to accelerate the speed at which many things are done. And with less manpower. Talking this through with you, I’m starting to think that two things are going to happen: 1. A significant number of people who are or will be laid off in an effort to reach profitability are going to either found new companies or take senior roles on newly founded ones. 2. A mild recession will occur for some reason (be it layoffs, consumer debt, political unrest, a long tail event, etc). That’ll finally lead the fed to lower interest rates, which in turn starts the process of a true VC driven boom cycle. Reinforces my suspicion that we are very early on in a super-cycle.

talldayo 2y ago
Well, I admire your optimism. Here's a short story about Datadog: Datadog sold what startups wanted. You want an observability platform, they would sell it to you piece-by-piece. This was great to the higher-ups; in an era when data science and dashboards were all the craze, every startup you went to had a monthly Datadog bill in the thousands. They were hooked. Then the economy went to shit, and you know what everyone realized? Datadog was useless. They were taking our money hand-over-fist to give us features that didn't matter and detracted from developer funding. When the higher-ups decided to stop trend-chasing, they made the call to drop Datadog and usually replace it with something else (some lumbering Prometheus stack). The moral of the story isn't that you can see a need and fill a need; the moral is that people will abandon half-measure solutions. OpenAI isn't filling a need, they aren't directly addressing a problem people have been demanding a solution for. They're selling a trend that you can join-in on for a per-call fee from their API. Much like Datadog, one day startups will start removing ChatGPT to replace it with local LLMs, and then probably remove LLMs altogether later down the line. Mark my words, the most successful businesses that come out of the next 10 years will barely even mention AI as a footnote in their product description.
- zer00eyz 2y ago
  Your right about local, but I think your gonna have some LLM around for a long time to come. Video games having rich NPC dialog (so many memes are gonna come from this). An agent that can have some sense of prioritizing email/text/chat to alert you. Someone to proof your emails. DOCUMENTATION has been the bane of every dev's existence, LLM's are useful here. LLM's aren't fire, they arent steel. I suspect that they are going to fit into our lives much the same way "maps on cell phones" did, and just make a lot of stuff smoother.
  - talldayo 2y ago
    We'll see. I'm not quite so optimistic; when you pitch these ideas individually, AI storytelling, AI email generation, AI-generated documentation, they sound scary. I like human storytelling, I like human prose, I like reading carefully-designed documentation. I don't think AI will be generating exemplary content for a while, and given the pace of things I see no reason for that to change; it's just too unpredictable to rely on. I think LLMs will be the robocalling of the 2020s. You'll stand in the self-checkout at the pharmacy, waiting 5 seconds for an LLM to generate a unique, awkwardly-worded way of asking you to clear the bagging area. This is what we're damned to, not necessarily because AI is bad but because executives don't know what to use it for.
  - rainonmoon 2y ago
    Man, I don't want to get into a whole thing here because we can all go around on this for days and it's not what the thread's about. But those just don't seem like selling points where LLMs are persuasively delivering. Rich NPC dialogue is highly dependent on things like thematic context and narrative consistency, both things LLMs are still horrible at on most scales. And LLMs can create a lot of documentation very quickly, but they don't solve the issue of owning and maintaining that documentation; I'm sure we've all worked places where a ton of docs was just as bad as none at all. I'm not trying to naysay, LLMs can do certain specific tasks very quickly and conveniently and I'm a frequent user. But it's like someone invented the dishwasher and now everyone's telling me one day it'll drive my car and make my movies. I don't see it.
zer00eyz 2y ago
Look its not to say that the ML/AI tools aren't powerful or useful... but bigger doesn't seem to be going very far, and there aren't a lot of people working on shrinking them or targeting them yet. There's a ton they can do to "power up" existing workers, but you're gonna need the human in the loop. > That’ll finally lead the fed to lower interest rates, Interest rates are low. Having them lower is DUMB and we saw what happened. The fed likely wants things to stay put incase inflation creeps back and then crank them up again... Not going to happen till well after we see a new president and maybe not even after that. We dont need the rates to change to see the market move with VC money. Look at what they were in 1999 compared to now. The money was flowing then, it will flow again regardless of rates. VC's are just going to have to make some better bets.
8organicbits 2y ago
> Turning unstructured data into structured data Does it do that without hallucinating? I see lots of people talking about this, but not mentioning quality.
- Canada 2y ago
  It does it pretty good, and for a lot of use cases having errors is tolerable
- fleeno 2y ago
  I've found it to be very good and accurate at these sort of tasks. I use it all the time to turn some weirdly formatted data into CSV or something else with structure.
elicksaur 2y ago
> The difference between Theranos and OpenAI is the later has created a product which, while imperfect, does something. I didn’t think it was a fair comparison at first either, but based on this standard it actually might be. Theranos absolutely made something. It had a device that took blood and gave back results. The fraud was that the results were “imperfect”.

talldayo 2y ago

Well, I admire your optimism. Here's a short story about Datadog: Datadog sold what startups wanted. You want an observability platform, they would sell it to you piece-by-piece. This was great to the higher-ups; in an era when data science and dashboards were all the craze, every startup you went to had a monthly Datadog bill in the thousands. They were hooked. Then the economy went to shit, and you know what everyone realized? Datadog was useless. They were taking our money hand-over-fist to give us features that didn't matter and detracted from developer funding. When the higher-ups decided to stop trend-chasing, they made the call to drop Datadog and usually replace it with something else (some lumbering Prometheus stack). The moral of the story isn't that you can see a need and fill a need; the moral is that people will abandon half-measure solutions. OpenAI isn't filling a need, they aren't directly addressing a problem people have been demanding a solution for. They're selling a trend that you can join-in on for a per-call fee from their API. Much like Datadog, one day startups will start removing ChatGPT to replace it with local LLMs, and then probably remove LLMs altogether later down the line. Mark my words, the most successful businesses that come out of the next 10 years will barely even mention AI as a footnote in their product description.

zer00eyz 2y ago
Your right about local, but I think your gonna have some LLM around for a long time to come. Video games having rich NPC dialog (so many memes are gonna come from this). An agent that can have some sense of prioritizing email/text/chat to alert you. Someone to proof your emails. DOCUMENTATION has been the bane of every dev's existence, LLM's are useful here. LLM's aren't fire, they arent steel. I suspect that they are going to fit into our lives much the same way "maps on cell phones" did, and just make a lot of stuff smoother.
- talldayo 2y ago
  We'll see. I'm not quite so optimistic; when you pitch these ideas individually, AI storytelling, AI email generation, AI-generated documentation, they sound scary. I like human storytelling, I like human prose, I like reading carefully-designed documentation. I don't think AI will be generating exemplary content for a while, and given the pace of things I see no reason for that to change; it's just too unpredictable to rely on. I think LLMs will be the robocalling of the 2020s. You'll stand in the self-checkout at the pharmacy, waiting 5 seconds for an LLM to generate a unique, awkwardly-worded way of asking you to clear the bagging area. This is what we're damned to, not necessarily because AI is bad but because executives don't know what to use it for.
- rainonmoon 2y ago
  Man, I don't want to get into a whole thing here because we can all go around on this for days and it's not what the thread's about. But those just don't seem like selling points where LLMs are persuasively delivering. Rich NPC dialogue is highly dependent on things like thematic context and narrative consistency, both things LLMs are still horrible at on most scales. And LLMs can create a lot of documentation very quickly, but they don't solve the issue of owning and maintaining that documentation; I'm sure we've all worked places where a ton of docs was just as bad as none at all. I'm not trying to naysay, LLMs can do certain specific tasks very quickly and conveniently and I'm a frequent user. But it's like someone invented the dishwasher and now everyone's telling me one day it'll drive my car and make my movies. I don't see it.

zer00eyz 2y ago

Look its not to say that the ML/AI tools aren't powerful or useful... but bigger doesn't seem to be going very far, and there aren't a lot of people working on shrinking them or targeting them yet. There's a ton they can do to "power up" existing workers, but you're gonna need the human in the loop. > That’ll finally lead the fed to lower interest rates, Interest rates are low. Having them lower is DUMB and we saw what happened. The fed likely wants things to stay put incase inflation creeps back and then crank them up again... Not going to happen till well after we see a new president and maybe not even after that. We dont need the rates to change to see the market move with VC money. Look at what they were in 1999 compared to now. The money was flowing then, it will flow again regardless of rates. VC's are just going to have to make some better bets.

8organicbits 2y ago

> Turning unstructured data into structured data Does it do that without hallucinating? I see lots of people talking about this, but not mentioning quality.

Canada 2y ago
It does it pretty good, and for a lot of use cases having errors is tolerable
fleeno 2y ago
I've found it to be very good and accurate at these sort of tasks. I use it all the time to turn some weirdly formatted data into CSV or something else with structure.

elicksaur 2y ago

> The difference between Theranos and OpenAI is the later has created a product which, while imperfect, does something. I didn’t think it was a fair comparison at first either, but based on this standard it actually might be. Theranos absolutely made something. It had a device that took blood and gave back results. The fraud was that the results were “imperfect”.