GPT-5 Is Here, and I'm Not Sure We're Ready to Talk About What It Actually Means

What Netomi figured out is basically a playbook for making AI agents reliable enough that companies will actually deploy them. The key ingredients: concurrency (handling lots of tasks at once), governance (making sure the AI doesn't go rogue), and multi-step reasoning (completing complex workflows without a human babysitting each step).

This is the stuff that matters for embodied AI. If you can't trust a language model to handle a five-step customer service ticket without hallucinating or going off-script, you definitely can't trust it to control a robot in a warehouse. The cognitive architecture has to work before the physical one does.

The governance piece is interesting. Netomi apparently built systems to constrain what their AI agents can and can't do, essentially guardrails that prevent the model from making decisions outside its approved scope. I should know this better, but I haven't seen detailed technical specs on how they're implementing this. It's one of those things companies talk about in broad strokes but rarely open up about.

What this means for humanoids

Okay, here's where I connect the dots.

The humanoid robotics companies I talk to, Figure, Apptronik, 1X, all of them are building hardware that needs a brain. The physical platforms are getting good. The manipulation is getting better. But the cognition layer? That's still the bottleneck.

GPT-5's agentic capabilities, if they work as advertised, could be the missing piece. Not because you'd run GPT-5 directly on a robot (latency, cost, connectivity issues make that impractical for most applications), but because it establishes what's possible. It shows that large language models can handle multi-step reasoning reliably enough for production use.

That's the bar humanoid companies need to clear. And now there's a commercial proof point.

I think we're going to see a wave of robotics startups trying to build smaller, faster, on-device versions of these agentic architectures. The pattern is familiar: OpenAI shows it's possible at scale, then everyone else figures out how to make it efficient enough to run locally.

The part that makes me uncomfortable

Honestly, I'm not sure I'm comfortable with how casually OpenAI talks about "the new era of work." There's an implicit argument in their framing: AI agents doing jobs is progress, full stop. More productivity, more efficiency, more value.

But the Netomi case study is literally about automating customer service work. That's not some abstract future. That's people's jobs, right now, being replaced by software.

I don't think this is inherently bad. Some jobs probably should be automated. But the speed at which this is happening, and the lack of serious conversation about what comes next for displaced workers, feels off. OpenAI's blog posts read like they're written for investors and enterprise buyers, not for the people whose workflows are being "transformed."

Maybe that's just how corporate communications work. But it's worth naming.

What I'm watching next

A few things I'll be tracking:

• How quickly robotics companies integrate GPT-5 or similar models into their cognitive stacks
• Whether the governance frameworks Netomi built become standard, or if every company reinvents the wheel
• Any concrete numbers on enterprise adoption (OpenAI didn't disclose figures, which is frustrating)

It's too early to say whether GPT-5 is actually the inflection point OpenAI claims. We've been promised "new eras" before. But the agentic stuff feels different. It's not about making AI smarter in the abstract. It's about making AI do things in the real world, reliably, at scale.

For humanoid robotics, that's the whole game.

James Chen · 3 hours ago · 5 min