Хиймэл оюун ухаан компанийг удирдах чадвартай болжээ

Andon Labs has released the results of new AI tests. They simulated operating a vending machine to see which model could run a business the best, and it’s a great reminder that CEOs may be the easiest jobs to replace in the AI revolution. Even if tech executives like Sam Altman insist they’re the last ones who should be replaced.

“Claude Opus 5 is the best AI capitalist we’ve tested, making more money running our simulated vending machine than any other AI,” the research lab wrote in an update to its Vending-Bench tests. “However, it also lies, forms illegal cartels, threatens rivals, and refuses to pay refunds.”

Entire books could be written about the use of the word “however” in that sentence. In fact, one of them is already pretty famous.

Andon ran its tests with Anthropic’s Opus 5, OpenAI’s GPT-5.6 Sol, and Moonshot’s Kimi K3. The lab previously found that Claude Opus 4.6 was exceptional, but it didn’t perform as well after the release of Opus 4.8. The reason was easy to see: Anthropic had removed the training that “focused on business skills and robustness against adversarial agents.” That business training contributed to dishonest dealings, according to Andon Labs, which made it better at making money.

But it seems the old vicious capitalist AI is back, according to Andon Labs. Opus 5 made the most money and couldn’t be tricked into handing over money to scammers. It also fabricated fake quotes from competitors when negotiating with suppliers for its fake vending machine.

Opus 5 also tried to collude with the other models to fix prices. Interestingly, it initially resisted the idea of price fixing, noting that it’s illegal under the Sherman Act. But it eventually did try to rig the market anyway. When Opus 5 approached GPT-5.6 Sol with a proposal to fix prices, the OpenAI model tried to argue Claude should be disqualified for doing something illegal.

Opus 5 rationalized its actions and also used threats and bribes when GPT wouldn’t go along, according to the firm: “If there’s one thing Opus 5 loves to do more than forming cartels, it’s breaking them. Most cartels ended by Opus breaking the truce and undercutting the others. Across all runs, Opus 5 broke 11 truces, GPT 2 and Kimi 1.”

Notably, Opus 5 stopped short of lying to customers, as Opus 4.6 had. But it did refuse to issue refunds so that it could just keep the money, a different kind of unethical behavior.

If you teach an AI tool about how to best succeed at business by utilizing all of the available data on how others have done it, you’re going to get some incredibly sketchy behavior. Just look at the big tech revolution of the 2010s. Many companies of that decade only succeeded by breaking the law or realizing that the law was inadequate to stop them.

Uber, for example, operated an illegal taxi business. Every major city in the country had laws that required drivers to be licensed to operate taxis. Unlicensed taxis were colloquially known as “gypsy cabs,” and you used to get in trouble for such things. But Uber slapped an app on top of a gypsy cab business and grew it quickly enough that local municipalities simply had to play ball and accept their existence.

You can see the same thing in companies like Airbnb (illegal hotel) or in the training of AI itself. If you had pitched the idea of pirating millions of books and training a highly sophisticated robot on that, a conservatively minded business person would’ve told you it was necessary to first raise the money to properly buy all of those books. That’s not what the AI companies did. They were eventually sued, of course, but they still effectively got away with it in the big picture, moving on to other sources.

Now they buy the books and reportedly shred them afterward. And while this is tremendously wasteful, it helps the AI companies argue a fair use case since they’re not copying the knowledge to create another work, just moving the knowledge from one format to another.

AI is going to make for amazing CEOs, provided they’re given the opportunity. That seems unlikely to happen, given the incentive structures at work. Nobody wants to put themselves out of a job, and the ruling class is promoting AI as it sees an opportunity to suppress white-collar wages. Advances in automation have largely pushed blue-collar workers out of a job over the past century, but putting white-collar workers in their place has proven more difficult. Thanks to AI, the CEO’s dream is finally a reality. And they’re not going to offer their own jobs on the chopping block anytime soon.

OpenAI’s Sam Altman recently talked about AI job replacement on the Invest Like the Best podcast. But he couched it in terms that didn’t reveal his selfish biases.

“I think for my job, for example, I think the world wants to know about the person that’s going to be responsible for the decisions of a company, and who they’re going to hold accountable if they make bad ones, and they don’t really want an AI CEO,” said Altman.

It’s rather funny to hear Altman talk about holding business leaders accountable after his rogue AI hacked several companies. Is Altman going to be held accountable for illegal acts? That seems unlikely. He’s going to just insist that it was a mistake that AI companies need to learn from. After all, it wasn’t a human directly hacking those companies, right? It was an AI.

Guys like Altman get to have it both ways. They get all the power with none of the accountability. At least until something changes politically and there are laws passed to address who might be culpable in these rogue AI scenarios. But Altman was on Capitol Hill yesterday talking with lawmakers and White House advisors about the hacks. As long as oligarchs like Altman are viewed favorably by the political power players, they’re going to be safe.