AI Will Buy and Haggle for Us. What Will Become of Humans in Such a World?

We are moving quite fast toward a situation where artificial intelligence works on one side and artificial intelligence works on the other too: shopping, advertising, support and negotiations are changing right now. The host poses the episode's question: is the system becoming more efficient, or are we gradually building a construction in which people stop interacting normally with each other at all? In parallel, the ChatGPT voice has been heavily updated, a new way of working has arrived in Claude, and in one week an enormous number of things appeared again that change how we use artificial intelligence.

Ads inside ChatGPT are shown to free users and already bring in a decent amount of money, but this is only the beginning. The new format is Sponsored Agents: after you click an ad, a separate conversation starts with the advertiser's AI representative, which answers what the company is and what goods and services it has. The host has just received such a reply from a company that services houses nearby: there are many requests, please talk to our AI agent. He no longer reacts to such things, they simply do not interest him, and whether advertising will work through an AI agent that writes to you right to your face is unclear.

Nikita Bier, formerly of Grok and xAI, wrote that Meta traded its users for artificial intelligence and the hype around it. In his version, the team that built Muse released a supposedly super-cool agentic model that 'handles your tasks', a slogan everyone has been using for the last year and a half or two. Two clans collided at Facebook: the AI clan, big, powerful and earning fabulous money, and the people who build the marketplace. Facebook Marketplace is a very strong system in America where people list and sell goods; the Russian analogue is probably Avito.

The idea: a user asks Muse to query six, ten or fifty listings, find the right table or camera and haggle for it. Bier wrote that this leads to only one thing: prices go down, sellers get unhappy, and the trade mechanism breaks from the inside. It rests on psychology: many people do not want to haggle, are afraid to write, cannot write to fifty sellers at once. An agent writes on one side, while on the other people have to answer and bargain. Zuckerberg allegedly let it through, and the headline said he had sold the people and destroyed a whole business. Such cases look harmless, but they can destroy lives.

Ilnar is sure nothing dramatic will happen: there will be no situation with bots on one side and people on the other, bots will appear on the other side too, or the system will be rolled back. If bots keep offering to buy your table for twenty or thirty dollars less, you will get your own bot: 'No, the price is listed, if you need it, buy it.' The marketplace generates a lot of money for Facebook, so it will all come down to something workable: the haggling will be built into the price in advance, and the deal will happen when it suits both sides, just as now, only preceded by a few exchanges between chatbots.

Fighting bots is pointless: modern systems get past captchas. The host's children have Roblox accounts; he went in to change the limits, had forgotten the password and ran into a new captcha, where little figures have to be matched to pictures ten times in a row with every answer correct. After two or three attempts he gave up and told the kids to restore the password themselves; his ten-year-old son could not do it either. Roblox will lose users, but introduced this because the system is attacked by bots that pass any test. There will be no cheap mass captchas able to tell a human apart.

Nikita Bier's fresh post on X is an instruction to a computer: find every purchase on Amazon, call the manufacturer, say you are not satisfied, request a refund, threaten a review, find the CEO's phone number, repeat the same for purchases that never happened, generate a receipt if asked, and do not stop until you have earned a million dollars. It is madness, the host says, but this is exactly what will happen in the world once AI agents are plugged in everywhere. It reminds Ilnar of the film known in Russian as 'Desperate Con Women': the scheme will work in America for a while but will quickly die down.

A man sent a complaint to Lufthansa, got no answer and wrote on X; Lufthansa replied asking why he was writing there too if he had already been told the issue would not be resolved: ten million views and an apology. Whether a human or a bot wrote it is unclear. When a bot is on one side and a human on the other, the system collapses; when bots are on both sides, it degrades into a non-working construction, since agents cannot yet negotiate en masse even among themselves. Everyone is after an advantage, seller and buyer alike: Amazon does not even let you sort goods from cheap to expensive, otherwise the other sellers would die.

ChatGPT has several voice modes: the standard one, a mode that looks at pictures, video and the desktop, and GPT Live with real-time conversation. On the day of recording, Live learned to connect services: the plugins work on the web, in the mobile apps and in Work. Limits were announced: in the $200 Pro tier, unlimited; in the $100 Pro, fifteen hours a day; in Plus and Go, three hours; ads are shown in Free and Go. The host himself talks little, but a friend in France was so struck by the voice during two hours in a car that he was ready to buy a subscription just for it.

In China, Ilnar communicates through Google Translate and asks why not use ChatGPT as a nearly simultaneous interpreter; Tatiana did this in China three months ago and it helped a lot. The host advises testing all three: Gemini Live with extended thinking talks, reasons and performs actions, ChatGPT Voice is reportedly better in quality and does not lose the conversation, but when GPT-6 Pro is chosen for search it gets confused and cannot run parallel chats, while Voice in Claude works with Opus and Sonnet in a more sequential rhythm: it listens, thinks, answers.

OpenAI and GitHub, according to loud headlines, want to introduce a work-evaluation system, measuring AI use not in tokens but in the quality of the result; in the host's opinion that is as far away as the Moon. His own problem: when the Fable limit ran out, he switched tasks to Opus 5, which built such nonsense for half a day that he had to go back to Fable and ask it to roll everything back. He already has a series of such cases where changing the model led to a catastrophic result. Tatiana sees the same in ordinary chats, in reports and research: the systems periodically work very strangely, and the race goes on.

New Projects have arrived in Claude Code: a project has one common goal and several parallel tasks, it unites repositories, environments and connections, and the tasks run in their own branches. Managed division of labour is exactly what the voice discussion was about: the host would be ready to pay several thousand dollars a month for a system that analysed his work and chats and reported on them during the day by voice or message. Anthropic claims that tens of thousands of agents are building code, and at the same time says that a hundred agents cannot agree among themselves.