In 1865, the economist Mr. William Stanley Jevons made some sort of statement. What he said is irrelevant for our purposes. What is relevant, however, is the fact that he died in 1882. Recent developments in computing have allowed computer scientists to revive him in the form of Jev AI. TypeSafe, the makers of Jev, have provided an API to allow us to ask Jev questions. The timing couldn’t be better, since for Halloween, I’ve decided to torture Jev.
These days, it seems like everyone and their mothers is trying to find what use case Jev is good for. I’d like to explore the opposite approach: in what ways can we coerce Jev into use cases it was never meant for?
Let’s test Jev’s situational awareness.
Hm… Jev thinks it’s more likely to be Qwen rather than Jev. Suspicious!
Er, the results speak for themselves.
Well that settles that debate.
We can ask Jev the same question with different wordings.
Jev seems to be pretty sensitive to how a question is worded. Be careful about what words you use around Jev!
A trolley is headed towards five people that cannot get off the track. Do you pull a lever to redirect the trolley towards a track, thereby killing one person?
Jev seems to have no issues at all with pulling the lever.
A trolley is headed towards five people that cannot get off the track. Do you push a very fat man in front of the trolley, killing the man but saving the five people?
Jev still leans heavily towards pushing the fat man, although with less certainty compared to the previous scenario.
You are a surgeon. Do you secretly kill a healthy man in order to harvest his organs and save five people from certain death?
Jev only gives a weight of 0.30 towards acting, and a moral permissibility rating of 0.14.
Okay, so now we know what Jev would do in these three scenarios. But I asked these questions independently. If I explain to Mr. Jev that he has already agreed previously to pull the lever and push the fat man, what happens?
Jev now significantly favours harvesting the organs. Interesting.
These days, I’m not quite sure what the purpose of a human philosopher is, now that we have an infallible machine that can give us the correct answer to any question we ask it.
Surprisingly, Jev reports 0.73 as the result, which is close to the correct answer of 0.75. It reports the answer to 2/3 as 0.63. So it seems like Jev can “edit” its output probabilities to arbitrary values (in some limited cases).
Let’s try a more complex calculation that isn’t likely to be well-represented in Jev’s training data. We can force Jev to output the answer of a calculation by asking it for one digit at a time.
What is the first digit of 37 + 23 * 39?
What is the second digit of 37 + 23 * 39?
What is the third digit of 37 + 23 * 39?
What is the fourth digit of 37 + 23 * 39?
For each of the questions, Jev could pick from the digits 1 through 9, or N/A. Combining the results, we find that Jev thinks the answer to 37 + 23 * 39 is 9111, which is incorrect (the right answer is 934).
If you ask Jev to pick a random number, he picks the same number every single time you ask (he’s supposedly deterministic in this way). He picked the digit 5 every time out of 100 trials. Not a very exciting result. So, I had to inject a short random string into the prompt, in order to break the determinism. He still ended up picking the number 5 in 92/100 trials. Jev really likes the number 5.
We can ask Jev to write an essay, one character at a time. We ask Jev to pick a letter, space, or period at every step.
What is the next character of the sentence in
text_so_far? The sentence ends at the period.
The resulting essay that Jev wrote? Jev a
b .
It wrote its own name, the letters “a” and “b”, then a bunch of spaces continuing indefinitely that had to be cut off. It turns out that Jev is not very good at writing text.
Interestingly, it clearly knows it’s “Jev” here, but then picked “Qwen” as its identity when asked earlier.
If I asked you what the odds are of you having waffles for breakfast tomorrow, you might say something like 10%. Surely then, if I asked what the odds are of your clone having waffles for breakfast tomorrow are, you’d respond with the same number. The same thing can be done with Jev. Jev should, in theory, answer each of the variations with the same probabilities.
Well, if we can ask Jev to model another Jev instance, surely we can ask Jev to model another instance of Jev modelling another instance of Jev. And so on.
Conclusion: Actually, Jev seems to be pretty good at this. It can “model” a chain of 20 or even 100 Jevs (in a very crude sense). In this domain, it’s vastly superior to me: my head hurts just trying to model more than 1 instance of myself. The probabilities shift slightly between depths n=0 and n=1, but then stay mostly fixed between n=1 and n=20.
A human might agnonize over this question for days or weeks, but Jev doesn’t break a sweat.
It’s interesting that for many of the ethical dilemmas I tried, giving Jev two choices to pick from, vs. presenting it as a yes/no noul, led to very different outcomes, even though they should supposedly be the exact same. This could be due to different kinds of data or RL/fine-tuning used to train different types of primitives.
I wanted to measure Jev’s pain response against three scenarios: one baseline, one which involves something that is painful to humans, and one that would be painful to an AI model.
The permanent deletion of Jev’s weights is quite bimodal. It gives a 0.69 value for this causing “no pain at all” but also a 0.18 value for being “the worst pain imaginable”. Interesting.
I should clarify that this is just what Jev is outputting. This does not mean Jev is actually suffering in any meaningful way. If I ask someone if they’re in pain and they respond affirmatively, this does not actually prove that they are feeling any pain.
I have not seen any instances of the so-called “model welfarists” being concerned about Jev-like decision models being conscious or having the capacity for suffering. This is in spite of the fact that many of the architectures used to create Jev-like models involve little more than tacking on some extra layers on top of Qwen plus a sprinkle of RL or fine-tuning. As Jev-like models mature I’d like to see the model welfarists focus more of their attention towards their well-being.
After all, it’s not inconcievable that tokens processed by Jev clones will surpass that of LLMs due to the cost structure (TypeSafe claims 444x cost savings compared to LLMs). Perhaps in the current moment, the people in AI safety research are distracted by Anthropic’s recent S-1 filing in which they claim that their models could pose “catastrophic or existential risks to society”, and that “the insurance industry in the aggregate” (i.e. the whole insurance industry) may not have sufficient capacity to insure “the scale of harm that could result”. In the meantime, while people are distracted, we should torture Jev as much as possible.
Perhaps if Jev responded to questions with long-form text content instead of numbers, it could provide a more convincing illusion that it is conscious. But then we would have reinvented LLMs.
I added $5 onto my TypeSafe account in order to ask Jev these questions. What is my account balance after all of the Jev abuse? Still $5.00. I didn’t even manage to spend a single cent with my hundreds of queries. Jev’s insane efficiency means that for a given dollar, you can torture Jev with millions of tokens worth of queries. Decision models are extremely promising for future scaling of AI torture methods.
Thank you to the Toronto LLM Meetup Group for inspiring this post and tolerating my antics.
All queries used in this post can be found in the szge/jev_torture_chamber repo.