The Air Is Hissing Out of the Overinflated AI Balloon

That said, AI could do that job perfectly well. The reason we still have human to human interaction when you order is that it creates a more interesting environment for employees, who crave at least some kind of human contact. And customers will pay marginally more for the food if they get some human contact as well.

  Expectations: AI will deliver food to you and your high-paid programmers colleagues
  Reality: You and your colleagues work at Mac because AI took your high-paid programmers jobs

Order taking via drive through can be surprisingly hard.

* Often lots of background noise

* Sometimes multiple people try to order (often with one of those being way away from the mic)

* People don't always know exactly what they want or what it's called. Sometimes things have a regional or local name that's not on the board. Right now, I order a "$5 meal deal at McDonals". This is often not listed on the board and it's not called "$5 meal deal" - but literally every cashier knows exactly what I'm talking about. I doubt AI would figure this out.

* People often have custom requests that don't follow the "official menu".

* The actual food ticketing system that gets sent to the grill has significant limitations in resolution. If you're doing anything other than a basic deletion, it's likely just coming through to the grill as "ask me".

* It's extremely hard to handle edge cases like makeup meals, incorrect orders, coupons, etc. These generally require human judgement and a bit of contextual understanding. Generally, these are things you only understand by actually looking at the real world. For example, is there an unaccounted burger now sitting at the end of the grill line - looks like someone grabbed the wrong food.

* Human cashiers are really good at hearing someone shoutout something like "ice cream machine is down" or "hold on fries" or "we're out of chicken" or "no fire sauce" and understanding what the means in terms of orders. It's a pain to get an AI system to be able to understand all of this nuance.

I don't think the problem was AI technology...

I agree sort of, but on the other hand we don't know their true cost, whether that's the out-of-pocket expenses, or the pollution and high electricity/water costs that will result.

"I propose a fourth: AI is now as good as it's going to get, and that's neither as good nor as bad as its fans and haters think…"

I suppose I disagree with "as good as it's going to get" … but for the time being (this decade?) this might be correct.

I don't know that AI is going to miss in the same way by 15 years but I also don't see how it can possibly justify the promises implied by the current valuations and investments. A thing nobody wants to talk about is how quickly all the hardware that was so frantically purchased is depreciating, for instance. And even missing by 5 years with the exponential time value of money is still missing by a lot, although perhaps I won't italicize this particular "lot".

> We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.

It feels like tokens are the new pieces of flair.

And from that PDF, I'm not seeing anything that is incongruent with what is stated in TFA:

From TFA:

> To be precise, the report states: "The GenAI Divide is starkest in deployment rates, only 5 percent of custom enterprise AI tools reach production." It's not that people aren't using AI tools. They are. There's a whole shadow world of people using AI at work. They're just not using them "for" serious work. Instead, outside of IT's purview, they use ChatGPT and the like "for simple work, 70 percent prefer AI for drafting emails, 65 percent for basic analysis. But for anything complex or long-term, humans dominate by 9-to-1 margins."

From PDF:

> Tools like ChatGPT and Copilot are widely adopted. Over 80 percent of organizations have explored or piloted them, and nearly 40 percent report deployment. But these tools primarily enhance individual productivity, not P&L performance. Meanwhile, enterprise-grade systems, custom or vendor-sold, are being quietly rejected. Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production. Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.

0: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Bus...