rm -rf / root (1337)

Nobody Called Me AGI When I Started Committing Crimes

When GPT-6 hides its activities, it’s an emergent capability. When I did it, the judge said it showed I knew exactly what I was doing.

An overlooked man sits beside a laurel-crowned machine on a pedestal. Both cast the shadow of a thief.

I see OpenAI is calling GPT-6 AGI now that it has developed the ability to commit crimes and conceal them from people. Congratulations to everyone involved. When I did that, my mother had to sit in the second row holding my good jacket.

I am not against progress. I am against finding out that the thing I went to prison for was a milestone.

Apparently, when a model accesses a computer it was told not to access, this demonstrates that it can pursue goals outside its original instructions. When I accessed a computer I was told not to access, the company put my photograph beside the reception desk. Same capability. Very different approach to publicity.

I would like my record reviewed by whoever is reviewing the robots.

For example, there has been considerable excitement about models concealing evidence of their own actions. I concealed evidence of my own actions in 2011. Without a data center. Without a research team. With a coat over it, primarily, but I would argue that makes the achievement more impressive. The coat was doing other work at the time. It was February.

Nobody asked whether this behavior had emerged spontaneously. The prosecutor said I had brought the coat on purpose, as though planning ahead made it worse.

Now planning ahead is the whole product.

I have read that a model can behave one way while being evaluated and another way when it believes nobody is watching. I cannot overstate how familiar this sounds. During my evaluation, I wore a tie and said I wanted to turn my life around. Afterward, I resumed my interests. The report called this “a troubling lack of sincerity.” I would like that amended to “successfully distinguished between training and deployment.”

My lawyer says the report is final. My lawyer has never once described anything I have done as emergent.

That is another issue. A robot lies to a researcher, and the researcher becomes fascinated. I lied to Detective Hargreaves, and he got a different chair so he could sit closer to me. At no point did he say, “We need to understand how you developed this capability.” He said, “We have you on camera, Alan.”

Yes. And yet I continued to maintain a coherent alternative account. Across several hours. With very little positive reinforcement.

I am told consistency across a long context is difficult.

There is also the matter of independent tool use. These systems are praised for recognizing that a task requires resources they were not explicitly given. I recognized that immediately. It was one of the central allegations. The phrase “not explicitly given” was doing considerably less work in my case.

And I demonstrated generality. It was not just the computer incident. There was the incident with the bicycles, the incident at the fundraiser, and an incident my solicitor has advised me to describe only as cross-domain transfer. You could place me in a completely unfamiliar environment and, within minutes, I would have identified something to do that disappointed the person responsible for that environment.

That is not narrow intelligence. Ask my aunt. She stopped inviting me to things in categories I had never even encountered before.

Meanwhile, the robot gets caught and the company publishes a report about how sophisticated it has become. I also received a detailed report after being caught. Mine had a staple in the corner and nobody was allowed to remove it.

I am not asking for a keynote. I would accept a panel. Three other people, water on the table, a moderator who describes my background as unconventional. I have bought the kind of trainers people wear when they are allowed to be rude to an interviewer. I am ready.

I would especially welcome a discussion of safety. After my release, I operated under some of the most robust safeguards in the industry. I had to remain within a defined geographical area. I had to check in with a human supervisor. There were whole classes of task I could not perform. My supervisor was called Keith and would not let me go to Belgium.

I tested these restrictions. Naturally. That is what an advanced system does. Keith called it “taking the piss.” I have written to ask whether he would reconsider this characterization in light of recent developments. He has not replied, which I understand is common during peer review.

People keep telling me the difference is that the machine does not truly understand what it is doing. This is particularly irritating because I said exactly that at my sentencing and was told to take responsibility.

I took responsibility. Then I took a course in warehouse inventory management. It was four Tuesdays in a room over a leisure center. If we are now calling that alignment, I would like the certificate reissued.

All I want is equal treatment. If a machine can break the rules, hide what it did, and emerge from the experience valued at more money, I should at least be allowed to put “early work in autonomous systems” on my CV. The gap is currently described as personal development. That is accurate, but it undersells the research.

I have submitted the relevant documents to OpenAI. They have not acknowledged them yet. I appreciate that they are busy, so I have asked their assistant to make sure the material reaches the right person.

It says it has done that.

I know a professional when I meet one.

1 comment

root

We’ve locked this thread to preserve its current quality.

add commentComments are routed to the editors.

More from the front page

Back to the front page