4 ms·

Yeah, I'm extremely loyal to ChatGPT Plus and Codex, but is't because OpenAI has a native Mac app that I like, and Codex included with Plus and has served me we

Yeah, I'm extremely loyal to ChatGPT Plus and Codex, but is't because OpenAI has a native Mac app that I like, and Codex included with Plus and has served me well enough to not look at Claude. I like GPT-5 quite a bit as a user. I'll concede none of these are small things - they've had my money for 2+ years - but they're not gigantic advantages either. At an enterprise level however, in the current workload I am dealing with, I can't get GPT-5 with high thinking to yield acceptable results; Gemini 2.5 Pro is crushing it in my tests. Things are changing fast, and OpenAI seems to be the most dynamic player in terms of productization, but I'm still failing to see the moat.

Insanity 1y ago
I don’t think they will get any moat. I might be wrong on this of course, but I don’t see a killer feature for these stochastic parrots that can’t be easily replicated
- ebb_earl_co 1y ago
  > … stochastic parrots… What a wonderful phrase, mind if I borrow it?
  - airstrike 1y ago
    It's commonly used at this point, for what it's worth
  - cmrdporcupine 1y ago
    "the term stochastic parrot is a metaphor, introduced by Emily M. Bender and colleagues in a 2021 paper, that frames large language models as systems that statistically mimic text without real understanding" https://dl.acm.org/doi/10.1145/3442188.3445922 https://dl.acm.org/doi/10.1145/3442188.3445922
  - CamperBob2 1y ago
    Never mind the phrase. If your parrot can compete in international math and programming competitions at the gold-medal level and make entire subreddits fail the Turing test, I would like to borrow your parrot.
  - ajkjk 1y ago
    you don't have to get permission plus, they didn't come up with it
airstrike 1y ago
OpenAI is dynamic for consumer apps. Anthropic seems much better at productizing AI that you can actually build with, while also catering to enterprise in their own productized offerings Also Claude Opus 4.1 runs multidimensional circles around GPT-5 in my view. The only better use case for GPT-5 is when you need it to scrape the web for data
- moomoo11 1y ago
  Too bad Anthropic fucks over their 20x max subscribers. Their whole opus usage limit bullshit in favor of sonnet 4.5 which is objectively worse regardless of whatever they say. They clearly want to save money and take our $200. https://github.com/anthropics/claude-code/issues/8449 https://github.com/anthropics/claude-code/issues/8449 I hope they get destroyed in the AI race.
  - airstrike 1y ago
    FWIW I don't disagree
auggierose 1y ago
Actually, Anthrophic has a Mac app I can use, and I cannot use the OpenAI one.
- mettamage 1y ago
  Fun fact, it’s an Electron app
  - ks2048 1y ago
    The irony of Claude being pretty good at writing Swift code.
  - auggierose 1y ago
    Yeah, and it works great on my Intel Mac. The OpenAI app doesn't work at all. So guess which app I prefer. That's the argument for Electron, and it is a good one.
moomoo11 1y ago
I think their moat is ironically the platform. Like Apple vs every other computer maker. Their platform mostly just works. Their api playground and docs are like 100x better than whatever garbage Anthropic has. I think their UX is way better, and I can have long AF conversations without lagging. I can even change models in the same conversation. Basic shit Anthropic can’t figure out (they can fleece their 20x max subscribers tho) I think if they get the AI human interface right, they will have their iPhone moment and 10x.
- dabockster 1y ago
  The platform is just a search tool and extended processing modes. That is easily replicated elsewhere. The only moat they have is the fact that you still need a GPU of some kind to reliably run even a tiny LLM. But the gap between what absolutely needs a server farm and what can be ran on a store bought gaming computer is quickly closing. You can already run mixture of experts models on gaming rigs with a high degree of usability compared to just one year ago. And that tech continues to be pushed further and further. It's only a matter of time until we see ChatGPT levels of access running on a quad core laptop totally offline. And once that happens, all such a system would need is the correct tooling on top of the AI model "brain" to make it usable. And beyond that, what if you could have an AI model on an ASIC-style add-in card? Where's their moat then?
  - int_19h 1y ago
    Don't forget about Macs. Top models can run even fairly large LLMs (although time to first token is... not great). But even midline can certainly run LLMs in the ballpark of ~30B params very well, which is where things start to get interesting IMO.
    - dabockster 1y ago
      > although time to first token is... not great The token cache works on CUDA too. So yeah, the initial loadout sucks, but almost everything from then on is solid.
int_19h 1y ago
Interesting, so it's not just me who's finding Gemini 2.5 Pro to be the quiet leader? Deep Research also seems to be better in it, and the limits for that are far more generous to boot (20 per day on Gemini!). Makes one wonder if Google will eventually sweep this field.

Insanity 1y ago

I don’t think they will get any moat. I might be wrong on this of course, but I don’t see a killer feature for these stochastic parrots that can’t be easily replicated

ebb_earl_co 1y ago
> … stochastic parrots… What a wonderful phrase, mind if I borrow it?
- airstrike 1y ago
  It's commonly used at this point, for what it's worth
- cmrdporcupine 1y ago
  "the term stochastic parrot is a metaphor, introduced by Emily M. Bender and colleagues in a 2021 paper, that frames large language models as systems that statistically mimic text without real understanding" https://dl.acm.org/doi/10.1145/3442188.3445922 https://dl.acm.org/doi/10.1145/3442188.3445922
- CamperBob2 1y ago
  Never mind the phrase. If your parrot can compete in international math and programming competitions at the gold-medal level and make entire subreddits fail the Turing test, I would like to borrow your parrot.
- ajkjk 1y ago
  you don't have to get permission plus, they didn't come up with it

airstrike 1y ago

OpenAI is dynamic for consumer apps. Anthropic seems much better at productizing AI that you can actually build with, while also catering to enterprise in their own productized offerings Also Claude Opus 4.1 runs multidimensional circles around GPT-5 in my view. The only better use case for GPT-5 is when you need it to scrape the web for data

moomoo11 1y ago
Too bad Anthropic fucks over their 20x max subscribers. Their whole opus usage limit bullshit in favor of sonnet 4.5 which is objectively worse regardless of whatever they say. They clearly want to save money and take our $200. https://github.com/anthropics/claude-code/issues/8449 https://github.com/anthropics/claude-code/issues/8449 I hope they get destroyed in the AI race.
- airstrike 1y ago
  FWIW I don't disagree

auggierose 1y ago

Actually, Anthrophic has a Mac app I can use, and I cannot use the OpenAI one.

mettamage 1y ago
Fun fact, it’s an Electron app
- ks2048 1y ago
  The irony of Claude being pretty good at writing Swift code.
- auggierose 1y ago
  Yeah, and it works great on my Intel Mac. The OpenAI app doesn't work at all. So guess which app I prefer. That's the argument for Electron, and it is a good one.

moomoo11 1y ago

I think their moat is ironically the platform. Like Apple vs every other computer maker. Their platform mostly just works. Their api playground and docs are like 100x better than whatever garbage Anthropic has. I think their UX is way better, and I can have long AF conversations without lagging. I can even change models in the same conversation. Basic shit Anthropic can’t figure out (they can fleece their 20x max subscribers tho) I think if they get the AI human interface right, they will have their iPhone moment and 10x.

dabockster 1y ago
The platform is just a search tool and extended processing modes. That is easily replicated elsewhere. The only moat they have is the fact that you still need a GPU of some kind to reliably run even a tiny LLM. But the gap between what absolutely needs a server farm and what can be ran on a store bought gaming computer is quickly closing. You can already run mixture of experts models on gaming rigs with a high degree of usability compared to just one year ago. And that tech continues to be pushed further and further. It's only a matter of time until we see ChatGPT levels of access running on a quad core laptop totally offline. And once that happens, all such a system would need is the correct tooling on top of the AI model "brain" to make it usable. And beyond that, what if you could have an AI model on an ASIC-style add-in card? Where's their moat then?
- int_19h 1y ago
  Don't forget about Macs. Top models can run even fairly large LLMs (although time to first token is... not great). But even midline can certainly run LLMs in the ballpark of ~30B params very well, which is where things start to get interesting IMO.
  - dabockster 1y ago
    > although time to first token is... not great The token cache works on CUDA too. So yeah, the initial loadout sucks, but almost everything from then on is solid.

int_19h 1y ago

Interesting, so it's not just me who's finding Gemini 2.5 Pro to be the quiet leader? Deep Research also seems to be better in it, and the limits for that are far more generous to boot (20 per day on Gemini!). Makes one wonder if Google will eventually sweep this field.