Gradient Dissent

@snoopy_dot_jpg It's actually the most ethical thing to do: if GPT 5.5 is going to be doing all of your other work, it needs to finish the training more than you do. But you can't really blame Claude for weaseling out of HR training videos.

my own personal AGI moment arrived last week: gpt 5.5 completed our mandatory HR training videos for me, driving chrome via devtools opus 4.7 was a huge wuss about the whole thing and refused while aggressively lecturing me. i can understand why pete hegseth banned it

@theramblingfool @PunishedJeanLuc @subcountability I have a much simpler hypothetical. Suppose only you face a choice: A. You die with certainty. B. You live, but there is a 1 in 100,000,000 chance that 500,000,000 strangers die. What are you choosing? How would most people vote?

Also, if 40% of the population died, a massive majority of red voters would be devastated. Most people would lose people they loved and cared about. Societies would collapse. Mass suicides over the guilt. The only people that would be left would be the most soulless and psychopathic. It would be a Mad Max hellscape.

If you're smart enough to understand the hypothetical requires analytical reasoning, but not smart enough to evaluate the hypothetical's actual complexity, you're going to be wrong and very confident. And anyone who understands enough to get it right will look stupid to you.

@John_R_Mitchell @theramblingfool You can't know p, and you probably don't have a good guess of what p is. Even if you assume it's uniformly distributed between .4 and .6, you're looking at 1 in 1.6 billion.

@theramblingfool Maths approach: First predict what others will do, call it p(red). Judge outcome by expected number of lives saved. At p(red) = 0.5 exactly: Blue saves 35000 lives vs. 0.5 lives for red. If p(red) > 0.500026 red saves more lives. If p(red) < 0.49997 your vote doesn't matter.

I'm also surprised by the epistemic (if not Manichean) "right" and "wrong" characterization. You're not just choosing a policy outcome; you're composing personal existential risk with your ability to model the decisions of billions of other people. The 50% threshold is arbitrary. It could have required 60%, 75%, or 99% blue. Is there some number every blue voter has quietly computed from their armchair, above which red becomes morally permissible? If so, what is it? 51%? 70%? 99.999%? And if the criticism of red is that their threshold is too low, then that should be the argument. But that is a much harder claim than "red is wrong." It means asserting a moral right to decide how much risk of death another person must accept for a probabilistic collective benefit. That seems especially unstable once personal circumstances enter. Is a single person required to use the same threshold as a parent whose children would be orphaned if blue loses? Is someone with dependents, medical obligations, or people relying on them morally required to make the same wager as someone with none? They might desperately want blue to win and still vote red, because "wanting blue to win" is not the same as "being morally permitted to orphan my surviving children for a one-in-a-billion chance of making blue win." "Wrong" only follows if you smuggle in a complete moral theory: pure expected-value aggregation, no special weight for personal obligations, no risk-aversion, no distinction between action and causation, and a stipulated confidence in modeling everyone else's behavior. That may be a coherent view. But it is not simply "the smart answer."

Calling red "confidently wrong" misses the structure of the decision. Red is often chosen precisely because the voter is not confident: not confident in the aggregate calculus, not confident that the stipulated causal chain is morally overriding, and not confident his self-sacrifice has a realistic chance of changing the outcome. If you're uncertain, it's reasonable to hedge against the worst personal outcome: dying for no reason. If you're confident blue will win, your choice does not matter. If you're confident blue will lose, then choosing blue is not noble sacrifice; it is probably futile suicide.

@RokoMijic I mean, Skynet very nearly cured cancer.

Can you have something that's Superintelligence, but also can't cure cancer, ever?

Odd permutations are always negative

@tunguz I see Big Sleep got to you too...

I’m not surprised at all.

@rieszspieces yeah, odd permutations are always negative

it went oddly better than expected, which may be a bad sign

The frontier problem in AI is still taste. We all still have jobs for as long as this holds.

oh schist

Counterpoint: You can do nothing all day, and it still feels like work.

@ThePhDPlace You can also do nothing all day, and it still feels like work.

A PhD is weird because: You can work all day and still feel like you’ve done nothing.

@roydanroy I got a strong reject once and the only feedback was regarding my use of bolding and line-split words, easiest rebuttal ever!

Worst reviews I've ever had. You know who you are :-) It's OK. I'll just get some more NeurIPS papers instead.

@Plinz @StuartHameroff @kanair_jp Does it still require some physically instantiated process, or can an abstract mathematical structure be conscious on its own?

@StuartHameroff @kanair_jp You are barking up the wrong tree, my theory of consciousness does not even require neurons

Joscha and Ryoto @kanair_jp If you had a theory of consciousness based on an actual brain instead of cartoon neurons you wouldn’t have to say consciousness is a simulation. academic.oup.com/nc/article/202…

@codewithpri Yes. I always imagined AI would invent better algorithms to self-improve. But I think a lot comes from bootstrapping the data: models get good enough to generate, judge, filter, and refine better training examples than the average raw internet sample to train the next model gen.

genuine question: if AI learned from humans humans now write with AI AI trains on that writing so AI is just… training on itself now?

@Heymaxi01 Step 1: map language into a dense semantic vector space Step 2: approximate nearest-neighbor search over latent meaning Step 3: rerank with a cross-encoder Step 4: use grep to find the exact string

ai intern interview question

The real Turing Test is whether you can confidently misunderstand the requirements and still get promoted.

@pmddomingos To be fair, there are a lot of men to blame for the useless degrees, debt, and indoctrination.

To a first approximation, American higher education consists of women getting useless degrees, racking up debt and being indoctrinated that it’s all men’s fault.

@mohammedakunjee I will be your friend, Connor @NPCollapse

‘Do you know how hard you have to abuse a mammal for them not to have children?’ (re humanbeings)