My doubts is: AI will just be a feature to implement into existing product or can exist as based product for hundreds new startup?
Since most models are open source anyone can train it and make everything available as API (eg open AI)
Since most models are open source anyone can train it and make everything available as API (eg open AI)
Why would we need an AI for X tool? Why not just pay for an AI tool? Perhaps the AI tool needs some integrations to work well, but when the ai can build arbitrary programs to achieve its goal… durable value of tools becomes dubious.
My suspicion is that many startups are trying to surf a wave of rising AI capability. Many will be swamped by the AI simply getting better and eliminating their differentiators. See GPT-4/Plug-ins… whatever is next.
Because (1) GPT-4 is giant and resource intensive, and (2) has several areas where it is not particularly strong without additional, external tools. Focussed AI in particular domains can be stronger and lighter in the domain.
Also, because even where GPT-4 or its successor is the ultimate backend model, actual AI tools are more than the backend model, and the specialization of that piece for the target domain can make a big difference.
External tools are great - but why would we expect the outcome not to be GPT-5 uses all external tools? unless you are building a mote that you must use your LLM to work with the tool, in which case - why not just charge GPT-5 customers through the plugin marketplace?
Why do most business not hire private armies at each location, even if it would solve all of their security needs? Utility/$ matters.
> External tools are great - but why would we expect the outcome not to be GPT-5 uses all external tools?
The fact that task specific AI may be called by GPT-n is... not a contradicition to the argument that GPT-n doesn’t obviate the need for task-specific AI, any more than the fact that the board interfaces with developers through managers means that managers existing obviates the need for developers.
LLMs like GPT-n may well end up as the most important AI-powered interactive user interface — mapping things to and from language is, after all their specialty, and language is an interface humans tend to be optimized for. But browsers are pretty much that today, and they are hardly the interesting part of the ecosystem.
And the thing with external tools—often AI themselves—is that they commoditize LLMs. There’s no moat, because the interfaces are natural languahr that any LLM can handle. And if I can have a good enough local LLM as my prinary interface, which can call out directly to external (to the LLM, but including local to me) services (including bigger, more expensice remote LLMs that add an additional failable service dependency) then I’m not screwed out of both my ptimary interface and the tools it uses when something fails remotely, and if its also cheaper to use, that’s another win.
If a cheaper llm exists, but it isn't as good - then the customer has to spend time evaluating it. Typically, these evaluations don't yield a tangible result.
I agree with you; the more time passes, the harder it is to differentiate from a generalist AI coupled with specialized plugins.
That said, it does activate a lot of opportunities to be part of this future plugin ecosystem.