Mark Williams-Cook

before things get significantly better, we need a new, um, type of technology.

I, I don't think from what I've seen, LLMs are going to get us to that stage.

We need to get, um, chat.

Uh, so GVC five, I think was the response was kind of a bit me, you know, compared to the, especially early on.

the progress we've had.

And the interesting thing that I've observed from GP5 is it seems more keen, because I'm very interested in when these models are doing stuff like grounding, when they're going off and doing web searches, or they're using other tools, you know, because the early models that came out, you know, didn't have internet connection, tiny, like context window, were pretty dumb.

GPT-5 seems to be positioned more as a, ask me a thing, okay, let me find the right tool to sort of go and use to get the answer or generate the answer or quietly write Python scripts in the background so I can get my answer.

Yeah, which I think is the right way to go.

Yeah.

And that's, I guess, veering more towards the kind of more agentic side of things, which is I understand the task you want me to do.

And then I just need to learn how to use the tools to do that.

And we've got all the MCP stuff coming now, the model context protocol.

So that's allowing these agents to interact with various sets of data and tooling just by chatting to them.

that to me is all exciting i think we'll get to a to a better place because they've realized that you know knowledge cutoffs massive problem hallucinations huge huge problem you know even the the um the technical card for gpt5 they have the various stats on the model release for the hallucination rates

And even under pretty good conditions, we're still in the, we're above 1%, 2% in a lot of cases, which again, apply that to if it became Google size, trillions of searches.

Yeah.

Yeah.

Yep.

If I went into a business and I was like, look, let me run all your stuff for you.

There's a 1% error rate.