ben_w

3 ms·

1.
by ben_w 5d ago
> You can actually discover those in open weight artifacts, reproduce them, study them and issue a security bulletin. No, you actually cannot. Not in general and without already knowing what the whole trigger pattern is. It's absolu
2.
by ben_w 5d ago
Good. If you think closed source software/binaries only is bad, wait until you see how awful the state of the art is with a clear-as-mud bucket of matrix weights. We know it's possible to train an LLM to secretly respond to certai
3.
by ben_w 5d ago
Thanks! Good to see a huge range of the same thing in so many styles. I've done similar at a smaller (free) scale to explore art styles and avoid the clichés that the image models prefer, with some success. My feelings, when presented
4.
by ben_w 5d ago
I'm an optimist by the standards of people who have ever written "p(doom)", I'd put p(doom before 2030) at 2-3% or so. On that scale, mostly political risk, e.g. this kind of thing going hot: https://www.yahoo
5.
by ben_w 5d ago
I disagree. They wrote "we're just different enough to feel that we're somehow special". I think in that sentence "special" is bound to "feel", it's not separable into an "is" statement
6.
by ben_w 5d ago
> Most animals are different from most others, we're just different enough to feel that we're somehow special. That's probably a similarity we share with other species, and I'd expect to a similar degree. We own pets,
7.
by ben_w 5d ago
Lot of that, for sure. But also, today, a lot of lawsuits about Facebook knowingly getting people addicted, and governments changing the laws to resist the general category. But not, as you say, for Windows (and others*) spyware. * I'm
8.
by ben_w 5d ago
I believe the argument you're responding to is "a set of dictionary definitions does not suffice to define intelligence"? I will admit the first time I read the thing you're replying to, I had a similar thought as you; F
9.
by ben_w 5d ago
> I don't think I agree with your characterization of the second definition. Time scales matter. It's not much use to be able to solve human-scale problems if it takes millennia. And it only takes months to train an LLM to the
10.
by ben_w 5d ago
> I'm not sure how Sundial and Gnomon are relevant to this conversation, since they're among the most benign, attention-seeking designs that Teller contributed to within his lifespan. Are you denying that the weapon could have
11.
by ben_w 6d ago
Not. A thousand cable channels all licensed by a government, is much the same as five broadcast channels licensed by the same government. A million YouTubers grinding The Algorithm while secretly sponsored by various world governments, isn&
12.
by ben_w 6d ago
> Why don't you think that AI doom is marketing? Because in the entire history of marketing, the only single example I can think of where someone tried to sell "this is so dangerous it has a good chance of killing everyone&quot
13.
by ben_w 6d ago
Information is an upper bound on knowledge. > 10 billion year training period, something many orders of magnitude more expensive than an LLM. I'm saying both definitions are valid definitions, they both point to important and differ
14.
by ben_w 6d ago
I can literally point to how much information is encoded in our DNA, because it's four bases (so 2 bits per base pair) and ~3.1 billion base pairs. 6.2 gigabits total, or slightly less than 1 gigabyte. A 1 gigabyte LLM isn't going
15.
by ben_w 6d ago
We invented controlled fire perhaps a million years ago; at a generation gap of 25 years, that's 40,000 opportunities for evolution to pass on a mutation that does anything. Written language is around 210 generations old, the capacity
16.
by ben_w 6d ago
Because our evolutionary environment doesn't contain cars, poetry, calculus, Star Craft, hamburgers, touch screen computers, or doors, and yet we are able to learn these things with (relative to a computer) very few examples. Most of t
17.
by ben_w 6d ago
How many examples you need to get good. Don't misunderstand: I'm happy saying AI models "think" or "have learned a thing", and for in-context learning I'd call them smart even by this definition… …but also
18.
by ben_w 6d ago
Man whose net worth depends on there not being problems, insists there are no problems. Unlike all the times I've seen people say "AI doom is marketing", I can easily believe this headline is just marketing.
19.
by ben_w 6d ago
That's the trick: in the early years, most people didn't know Zuckerberg's history.
20.
by ben_w 6d ago
I've heard many variants; Paris in an earlier war, too. I try to keep an open mind about propaganda fooling me today; the people who were fooled in the past often were not fools themselves.
21.
by ben_w 6d ago
What's the old quote from WW2? I couldn't help but notice how each successive headline reporting our glorious victories seemed to draw closer to Tokyo. Something like that. Well. I can't help but notice how each succes
22.
by ben_w 6d ago
They're each in different grey areas that were previously the commons. Facebook didn't invent "talking to friends" or "showing adverts", but made itself "the place" hard enough most of the advertisers
23.
by ben_w 6d ago
I'm old enough to remember when people had expectations of privacy on Facebook, when the discovery that they didn't made for schocked headlines.
24.
by ben_w 6d ago
Almost everything I interact with was made by an organisation (or a disorganisation), not by an individual. Say I download an app. Who made it? The programmer? Their PM? Apple's store requirements? The US government, for whom there was
25.
by ben_w 6d ago
> I look at the person wielding the tool. I look at the whole chain of influence behind them, and it's quite a lot more upsetting than the smiling face I interact with. The whole "AI is a Lovecraftian tentacle monster wearing a
26.
by ben_w 6d ago
> Then they don't need to train on github, no? Why not release a new model trained from Knuth's Art of Programming, Cormen's Introduction to Algorithms and the C specification. Because they're really really stupid and
27.
by ben_w 7d ago
I understand (and to extent I agree), but he does seem to respond to some headlines at least.
28.
by ben_w 7d ago
Some did, some didn't, government acted. Trump acted, even if he now dissociates from vaccinations in general and by extension takes no credit for Operation Warp Speed.
29.
by ben_w 7d ago
Could go either way, TBH. So long as there are in fact those things, and so long as it didn't sneak something else in there at the same time, it being just on the knife-edge between sense and word-salad is better than "fix".
30.
by ben_w 7d ago
It's not impossible, of course. But AI is also being hyped beyond its competences* by e.g. Hegseth, to the degree that Anthropic saying "we don't want to be used like this" resulted in an illegal attempt to punish Anthro

More ›