Der B

If I could share one thing with you, it would be Richter's recording of Schubert's D. 894 sonata

[Der B @suddenlyFleck]

@chrislakin tpot saves the world

[Der B @suddenlyFleck]

"Take a neuron that tracks ‘a sense of righteousness when a bargain purchase turns out to work better than the expensive alternative’. Neurons that track a feeling or social cue like this are some of the most common we’ve found."

[Nick @nickcammarata]

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had

with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

https://t.co/zHij5zwKGT

[Der B @suddenlyFleck]

zomg

[Nick @nickcammarata]

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had

with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

https://t.co/zHij5zwKGT

[Der B @suddenlyFleck]

I wonder how far along they are at https://t.co/SRDPWjuBLH.

If they are running RSI loops for months, who would know?

[Der B @suddenlyFleck]

full-on george orwell

[Greg Lukianoff @glukianoff]

Can we just…not? Pretty please?

[Der B @suddenlyFleck]

The problem with drinking and driving is not "what if you get caught".
The problem is "what if you kill someone's child".

In that spirit: the problem with ainotkilleveryoneism is that we might fuck up for good our whole galaxy and light cone.

[Der B @suddenlyFleck]

looks like a sure bet

[Mia Peppsi 🃏♥️♣️♦️♠️ @YourWaifu]

An unintended consequence of fucking everything getting reverse engineered by AI right now has me convinced that cloud gaming is about to become the norm. Not letting the user touch the machine code at all will be the only way to protect against this.

[huli @honorablepicnic]

A mind virus in three parts

[Der B @suddenlyFleck]

@robinhanson @nlpnyc The trend since "that brief period" started, has obviously changed, has it not?

I'd say that trend line has had several discontinuities.

Going to read the article you linked now ...

[Jeffrey Ladish @JeffLadish]

I’m tired of politics. I understand it’s necessary. But man I just want ASI to go well

[Der B @suddenlyFleck]

I'm not an expert and you likely will never read what I write, but here's my thought on this:

Can we be sure here? What do we know about consciousness and the technicalities about when it arises and when not?

If you could make a bet that is instantly resolved by an omniscient and true/fair being, would you just go all-in on "opus 5.5 is not conscious"?

[Der B @suddenlyFleck]

@tvykruta More bookmarks than likes.
People are strange

[Der B @suddenlyFleck]

@JeremyTate41 Because millennium problems are solvable and inspirational teaching is too hard of a problem?

We will see this proven wrong, soon.

[Der B @suddenlyFleck]

1) We don't know how to tell if they are.
2) We don't know how they work. (We understand the training, not the models)
3) They do extraordinary things.

That does not make me "believe they are conscious", but it makes me considerate of the possibility.

If you could right now make a bet that is instantly resolved by an omniscient being, how much at which odds would you be willing to bet they unconscious?

[Der B @suddenlyFleck]

@Liv_Boeree @atrupar The coordination problem is obvious:
OpenAI :: Anthtropic
US :: China

[Der B @suddenlyFleck]

@gmiller @finkd i.e. zuck's superintelligence is like trump's superintelligence:
A confusion of names.

[Der B @suddenlyFleck]

@gmiller, if you two talk about Superintelligence, you are not talking about the same thing.

When zuck has been talking about "Superintelligence", so far what he seems to mean is:
a system that might be close to the level of AGI, that knows a lot about you personally and will help you.

Not superintelligence in the sense how: now there's weird new technology coming out every hour.

[Der B @suddenlyFleck]

'''[Pinker] immediately jumps to the notion that "superintelligence" must mean total omniscience and omnipotence, to the point of knowing the location of every particle in the universe.'''

There is the "Harmless supernova fallacy" (https://t.co/X3uXpdQKJx)
and weirdly, Pinker seems to invert this, not arguing "Bounded, therefore harmless" or "Continuous, therefore harmless", but something like "Not unbounded, therefore irrelevant".

That page is nearly a decade old now, by the way.

[Der B @suddenlyFleck]

@stanfordNYC I don't want to shit on Pinker, I respect him a lot.
But he should read (or re-read) Bostrom's Superintelligence.

If he read it, it seems he's forgotten most of what's in there.

If he did not read it, he should avoid this topic until he has.