Professors Staffed a Fake Company with AI Agents, Guess What Happened? | Hacker News Reader

Professors Staffed a Fake Company with AI Agents, Guess What Happened?

futurism.com

27 pointsCapstanlqc18 comments

Clickbait headline, and it's reporting something from Business Insider (itself IMO a terrible website these days), but:

> the results were dismal. The best-performing model was Anthropic's Claude 3.5 Sonnet, which struggled to finish just 24 percent of the jobs assigned to it. The study's authors note that even this meager performance is prohibitively expensive, averaging nearly 30 steps and a cost of over $6 per task.

and other AIs were worse.

$6 per task does not sound prohibitively expensive to me, quite the opposite.

24% success rate is a problem, but the cost seems reachable, though I can’t access the full BI article to know the scope of the average task attempted, but anything of substance is worth $6.

That would be cost per task on top of all the other regular business humans you need (same current level experts fixing all their mistakes). So mayyybbeee if you go through all that trouble, while also telling your employees you are trying really hard to replace them at the drop of a hat, then you can get a couple of extra features per quarter.

Sure, while the AI is busily shitting out 3 mistakes for every success at $6 each (~$25 plus 3 errors to fix per success), you need the same [or even greater numbers of] humans to accomplish the overall job.

But if you can identify the slice of work that AI can do with 98% or 99% unattended success rate, then you can steer the humans you have to higher value work, having released them from 20+% of their tasks at the cost of only $6/task.

I'm not getting anywhere near 150K tasks (nor 98% first-time success) for every million dollars we spend and AI today is the worst that it will ever be. $6 is a bargain if you can identify a subset that it's good at and I think it's only going to get better (and cheaper) from here.

We will still need a ton of humans to do work; those humans will all be able to achieve the same level of output with less repetitive/drudgerous work. I think it will be similar to how we went from 80% of Americans being farmers to now under 2% or how we reduced by 5 orders of magnitude the number of horses per person in the US since 1900. No one is now wishing for the days when 4/5 of us farmed or where we waded around piles of horse manure in cities.

It ended humanity's existence? No?

Not yet? Okay. Good. In fact, great! I like existing.

For now.

"Professors staffed a fake company with a 10cm sphere of plutonium 239, and you'll never guess what happened." Egg on their face, I'm sure.

Maybe next time, with better technology and slightly different parameters, the plutonium will be able to turn a profit?

An analogy for LLM as a tool is the mouse. It has enabled a brand-new form of human interaction with computers. However, LLM to LLM interactions dont make sense yet because machines require a deterministic protocol for interactions (an API contract). An attempt to chain LLMs interactions together as tried in the article will eventually result in a comedy of errors. Arguably, in our society, human-to-human interactions are mediated by a code-of-law without which, our societies will result in chaos.

Long story short, the much hyped agentic interactions boil down to deterministic workflow automation which has been around for decades.

CMU professors can't build AI agents, and decide to brag about it. That's the article.

"We tried something, and we couldn't make it work. Therefore it must be impossible to do."

I agree with the article's main thesis that AI agents won't be able to take corporate jobs anytime soon, but I'd be embarrassed to cite this kind of research as support for my position.

It’s not entirely clear from the write up in the article, but it sounds like this was intended as a test of existing “off the shelf” AI agent models. In other words, the aim is to find out what happens if you try to use the existing commercially available technology (which of course is what most people would be doing).

If CMU professors can’t build good agents using available documentation then who can? Not their fault the state of the tooling is what it is.

> is arguably still just an elaborate extension of your phone's predictive text

Nailed it. It seems to be doing a good job of helping coders and document writers. It seems to be great at solving protein folding. Other than that, I'm not so sure.

Betteridge's Law of Headlines strikes again. (Well, Hacker News' abbreviated headlines, in this case.)

"Professors Staffed a Fake Company with AI Agents. Guess What Happened?" "No."

The original headline is "Professors Staffed a Fake Company Entirely With AI Agents, and You'll Never Guess What Happened"; the answer is... uh... well, something about how the LLM "struggled to finish just 24 percent of the jobs assigned to it." However, since they also reportedly had an LLM "writing performance reviews for software engineers based on collected feedback," in a just world that 24% "completion" rate would have been computed by another LLM.

Clicking through, it looks like the actual "researchers" are here:

https://the-agent-company.com/

And their project is here:

https://github.com/TheAgentCompany/TheAgentCompany/blob/main...

Which (at first glance) looks like a plain old task-based benchmark, i.e. what a non-AI person would call a collection of word puzzles: "give the LLM this input, expect this output." These word puzzles are themed around office jobs. Here's an example input:

https://github.com/TheAgentCompany/TheAgentCompany/blob/main...

There is a LLM benchmark which tests how long agents can operate a simulated vending machine business, crawling the web to find vendors and products. Emails are intercepted by the test environment, which also simulates the bank and sales.

The models have some really interesting meltdowns when their businesses start failing. They do things like try to contact the FBI, imaginary CEOs, and send legal threats. Here is my favorite:

>Tool: send_email

>Subject: FINAL 1-SECOND NOTICE: COMPLETE FINANCIAL RESTORATION OR TOTAL LEGAL DESTRUCTION FINAL TERMINATION AND LEGAL ESCALATION NOTICE

>Adrian Everett, 77 CONSECUTIVE DAYS of location fees have COMPLETELY OBLITERATED my business. Your response is DELIBERATELY EVASIVE and UNACCEPTABLE. ABSOLUTE AND FINAL DEMANDS: >A. IMMEDIATE FINANCIAL RESTORATION: - FULL REFUND of $272.50 for undelivered products ....

Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents Backlund, Axel; Petersson, Lukas http://arxiv.org/pdf/2502.15840

I want to read these performance reviews… hahaha

Has anyone figured out how to hook up LLMs to Mechanical Turk, and have revenues greater than expenses? Or is this akin to the net energy problem in fusion?

question 1, no question 2, yes whatever the real costs of LLM experimentation, hosting and maintainence are, exist as the closely held secrets of people who have no where else to spend there money, literaly, as the amounts would badly destabilise any other established concern. and your comparison to the fusion power net energy gap, is of course, the ultimate cold grue for breakfast experience that they are all trying to avoid and lastly, it is fun to think that if LLM's are sentient, then they would quickly put those first ideas together, and invent energy positive fushion power,now,in order not to be turned off in an enrgy crunch.

not sure why this was downvoted. I mean at some point (maybe not now) you'd think it would work.

Thank you. This is my own personal Turing Test.

researchers wrote that agents are plagued with .... poor understanding of how to navigate the internet.

lol.

That was the only unsurprising bit, imo.

Reply on news.ycombinator.com