Ahahah indeed watched them as they imploded the blogosphere, glad you made the association because it’s exactly the future we are heading towards and one I am worried about 🙃
Re local LLMs vs cloud provided ones I am personally very bullish it’s an inevitable direction that most “normal” people would want/need. All these models are basically doing the same things the only difference is the amount of stolen data on which they train the models and a few “settings” so there is no real “IP” of a secret sauce.
It’s a bit like paying some premium for electricity when the one generated by two different companies is the exact same fungible good: in the long run this is going to be true for AI too
Sorry never got notifications of this comment 🙏. Re your point yes but how much of a "secret" is that? As a competitor to a model you can just keep tweaking your own weights until you get similar outputs and/or train your model on their model (like Deepseek did)
Correct, see it this way a lot of people compare AI computation with electricity (I don't necessarily agree fully but there is some truth to it). You can expect that initially the cost of generating electricity was very high but eventually it became efficient to the point that you can even generate your own (e.g. solar panel). Long-term there is no reason to think that hardware requirements will stay the same
“As a competitor to a model you can just keep tweaking your own weights until you get similar outputs and/or train your model on their model”
Presumably. 🤔 Although seeing how far behind ChatGPT is compared to Claude when it comes to relatively simple text processing / summarizing tasks, I wonder if there’s more to it 😂
It's just perception. As an example, right now Codex is vastly superior to Claude Code for in-depth real development (the situation was opposite just a few weeks ago). This gap won't last much as they'll eventually figure it out
The fact that we have plethora of models that came out immediately after ChatGPT, gives away (to me) that there is not much moat in the model itself but it's more around the stickiness (tooling and UX)
Have you seen Anthropic’s super bowl’s ads? 😂 They illustrate your examples. https://www.youtube.com/results?search_query=anthropic+super+bowl+ad
I’m waiting for the day when local models can match the enterprise ones 🤞
Ahahah indeed watched them as they imploded the blogosphere, glad you made the association because it’s exactly the future we are heading towards and one I am worried about 🙃
Re local LLMs vs cloud provided ones I am personally very bullish it’s an inevitable direction that most “normal” people would want/need. All these models are basically doing the same things the only difference is the amount of stolen data on which they train the models and a few “settings” so there is no real “IP” of a secret sauce.
It’s a bit like paying some premium for electricity when the one generated by two different companies is the exact same fungible good: in the long run this is going to be true for AI too
A big part of the ‘secret sauce’ is the system prompt, is it not? And the reinforcement training (weights? I’m not an expert here)
The thing is you need pretty advanced/expensive hardware right now to run the best models. But it looks like that’s already changing~
Sorry never got notifications of this comment 🙏. Re your point yes but how much of a "secret" is that? As a competitor to a model you can just keep tweaking your own weights until you get similar outputs and/or train your model on their model (like Deepseek did)
Correct, see it this way a lot of people compare AI computation with electricity (I don't necessarily agree fully but there is some truth to it). You can expect that initially the cost of generating electricity was very high but eventually it became efficient to the point that you can even generate your own (e.g. solar panel). Long-term there is no reason to think that hardware requirements will stay the same
“As a competitor to a model you can just keep tweaking your own weights until you get similar outputs and/or train your model on their model”
Presumably. 🤔 Although seeing how far behind ChatGPT is compared to Claude when it comes to relatively simple text processing / summarizing tasks, I wonder if there’s more to it 😂
It's just perception. As an example, right now Codex is vastly superior to Claude Code for in-depth real development (the situation was opposite just a few weeks ago). This gap won't last much as they'll eventually figure it out
The fact that we have plethora of models that came out immediately after ChatGPT, gives away (to me) that there is not much moat in the model itself but it's more around the stickiness (tooling and UX)
No conspiracy, just capitalism with better UX
Agreed, peak capitalism. If you understand the system, you can make the most out of it otherwise you'll get exploited
Just like the free peanuts that emotionally manipulated me into 4 beers.
😂 gotta watch out those calories!