Using shell as an intermediary is the same kind of indirection as tool search and tool use from code, so I think you are largely agreeing with their substantive sentiment while disagreeing with their word choice.

Not exactly. Proliferation of tools built into agents for computer user is anti-thematic given that computer use is a key focus for model development.

Why build a tonne of tool-use infra when you could simplify instead?

> less tools, better skills to apply those tools

All models have peaked (the velocity of progress is basically zero compared to previous years) -there are not going to be "better skills" (any time soon).

All these bubbled up corps(es) have to try to sell what they can, agent this, tool that, buzzword soup to keep the investors clueless one more year.

That’s one narrative.

It’s typical for the foundation to settle before building on top of it.

Additionally do agree there is immense commercial pressure.

I’m quite curious how it all shakes out across the majors. If the foundation is relatively similar then the differentiator (and what they can charge) will determine their returns on this investment.

As a user, I love the competition and the evolution.

As an investor, am curious how it shakes out.

The problem with this, unless I'm misunderstanding what you're saying, is that the model's responses go into the context. So if it has to reinvent the wheel every session by writing bash scripts (or similar) you're clogging up the context and lowering the quality/focus of the session while also making it more expensive. When you could instead be offloading to a tool whose code never comes into the context, the model only has to handle the tool's output rather than its codebase.

I do agree that better tools, rather than more tools, is the way to go. But any situation where the model has to write its own tools is unlikely to be better.

Yeah I kind of agree. I think there's demand for an connector ecosystem because it's something we can understand and market, but I think it's the wrong paradigm

While maybe the model could do everything from first principles every time, once you have a known good tool that performs a single action perfectly, why not use that tool for that action? Maybe as part of training, the model could write, test, and learn to trust its own set of tools, rather than rely on humans to set them up afterwards.

In this case LLM would have to write a bunch of stuff from scratch though and might call APIs wrongly.