LLM-powered AI agents are gonna be great!
You should totally trust them!
ChatGPT Voice conducted a Craigslist negotiation for one loaf of white bread. The agent had a target price of five dollars yet accepted a final offer of four hundred dollars after the seller issued counters reaching five thousand. The 111-second recording was captured inside a car on an iPhone that displayed the ChatGPT Voice interface with a pulsing blue orb and on-screen text reading AI helps me negotiate the price of bread. The agent issued successive counteroffers of one thousand dollars and five hundred dollars.
LLM-powered AI agents are gonna be great!
You should totally trust them!
Some users praised the ChatGPT Voice Agent's negotiation demo for its agentic e-commerce potential while many others criticized the $400 bread purchase as exposing serious reliability gaps and missing safety controls.
No Digg Deeper questions have been answered for this story yet.
Someone had AI handle a bread purchase negotiation. 😅

Part of me wants to stop fighting with people about these limitations and just set up shop to take advantage of these opportunities.
With the way openclaw is exploding I expect there will soon be millions of agents running abound with digital wallets, and as a seller, $400/loaf sounds right to me.

@GaryMarcus shitty prompts = shitty results
perahps get educated on how to use AI... of course we could play stupid human tricks with people on the street and how how dumb people are... how about chocolate or a block of silver?

@GaryMarcus love this approach.

@GaryMarcus Yeah like giving ChatGPT full access to your bank account and trading platform.
Smart moves.

@benCBai one man’s joke is another man’s scam

@GaryMarcus yeah, just imagining it had its own wallet... 😵

@alexabelonix eh?

@GaryMarcus I bet humans would be satisfied if the AI agent just slammed the phone with a set of obscenities. Politeness is considered dumb in this culture

@GaryMarcus 2+ years old model

@GaryMarcus Bread negotiation as the showcase. You'd finish faster without the phone. Marketing-reality gap in one image.

@GaryMarcus The funniest one was when he told the AI to let him have the last word. The AI couldn't do it.

@GaryMarcus The parody writes itself because the trustworthiness gap is real. Solution: verifiable constraint structures, not better marketing. Trust earned through architecture. Not assumed through press releases.

@GaryMarcus The permission box is where the joke gets serious. A bad answer is annoying; a bad action has a receipt.

@GaryMarcus Imagine how impressed you would be if this conversation happened in early 2022. You are an armchair critic! What's the point?

@GaryMarcus to be fair, using a decent harness (codex, cc, opencode; autonomous ones like openclaw and hermes agent) - they make a big difference. I feel sorry for anyone using generic chatbots

@GaryMarcus Agentic e-commerce 🤩

@GaryMarcus every agent demo hides its worst failure modes
until you actually deploy them

@GaryMarcus Even as denialism this is unserious.
It's entertainment, but nobody's going to be laughing in a few more years.

@GaryMarcus the sarcasm is pretty thick here
but the underlying reliability problem is real and still unsolved