ConcurrentSquared

"wir müssen wissen; wir werden wissen!" - David Hilbert

[ConcurrentSquared @ConcurrentSquar]

continuing with o3's motivational posters:

[ConcurrentSquared @ConcurrentSquar]

@bouncy72_ @dm4uz3 dseek!

[ConcurrentSquared @ConcurrentSquar]

i'm going to build a custom harness just to fix this one issue

[ConcurrentSquared @ConcurrentSquar]

... why does claude code still send my /exit and /quit commands to claude?

[ConcurrentSquared @ConcurrentSquar]

relatedly note that this task should be much easier than general language (for one: scheme has a type system, and a very restricted and formal grammar); i have the prediction https://t.co/QWTAR9PBKp, which should be easy if gpt-7 solves interp. via weight-based programming

[ConcurrentSquared @ConcurrentSquar]

as of nov. 2, 2026, 'weight-space programming' (via opus 5.5) doesn't beat adam + even a 3.4m parameter gpt on autocompleting basic scheme (much of it being solutions to sicp exercises)

[ConcurrentSquared @ConcurrentSquar]

https://t.co/Jnvi2Gh6wT
alternatively if i'm wrong and we don't die after gpt-7 solves interp. i will post a personal secret with the sha512 hash dcc01fec544c2eedf88324b57c697d3ab69d9182e7c32b9eb29eb9f62c9d351f0586e3ada0493b845b8171fc0d49c4deef5fd55cc267b090cba29ec241d74623

[Psyho @FakePsyho]

meh, the more you dig into each experiment, the less magical it gets; it feels like a loose collection of cherry-picked demonstrations with a grand narrative built around them

Exp 1-2 are tiny (~20K params) with unusually easy to interpret architecture; Exp 1 is an MLP, where nearly all params are used in the first layer, so you can easily tell what kind of objects the network is trying identify
Exp 5 is about designing a game strategy (we already know that Astra is extremely good at this) and then converting it to a tiny (~20K params) MLP; entirely doable by a human, but of course would require a lot of effort
Exp 3 & 8 are closer to automated mechinterp using python and established techniques, rather than direct weight interpretation (at least according to chat)

to be clear, those are genuinely cool results, but I think people are reading too much into them

I don't think any of this should meaningfully change anyone's intuition about what models are capable of

[ConcurrentSquared @ConcurrentSquar]

nvm i missed the deadline, yay!!! :(

[ConcurrentSquared @ConcurrentSquar]

hmm maybe i should apply to pivotal today

[ConcurrentSquared @ConcurrentSquar]

@ryunuck @cephaloform (note that this was a good assumption to make in the era of gofai, see e.g. https://t.co/PmFGBB0yDN!)

[ConcurrentSquared @ConcurrentSquar]

@ryunuck @cephaloform the 5 millisecond yudwoskyian foom required general ai to have a nice semi-symbolic algorithm whereas it could be easily optimized in vivo by some level of intellect, but neural networks didn't seem to be nice in that way
but this research indicates that maybe they are!

[ConcurrentSquared @ConcurrentSquar]

@theproteindance ok maybe i shouldn't have been given the power to long-tweet
(reference to the original joke in https://t.co/sqVy50LF1y)

[ConcurrentSquared @ConcurrentSquar]

Micromarriages at the IPO party were low, so the Anthropic alignment researcher wrote to the local ACX meetup, asking for help from other rationalists. Samotsvety was assembled, headed by a member of the Bayesian Conspiracy, and weeks of intensive on-site debate took place.
The LW users then returned to the meetup, Guesstimates crammed with priors, where the task of writing the blog post was left to the team leader. Shortly thereafter the beisutsukai returned to the (now entirely empty) party, saying to the researcher, "I have the solution, but it works only in the case of triangular polycules in a vacuum."

[ConcurrentSquared @ConcurrentSquar]

ok but while i do get the safety argument, i think with good classifiers + also the fact no post-training is applied to a base model means that releasing base models to the public would be fine, if only there was no risk of distillment (but sadly there is!)

[ConcurrentSquared @ConcurrentSquar]

i wish for happiness for everybody, free, and also gpt-4.5-base, if not gpt-6.2-base, and claude-fable-5.5-base
i can be trusted with the base text-completion models; no one will go away unsatisfied

[ConcurrentSquared @ConcurrentSquar]

@willdepue (whereas with a robotics model, you need to learn classical physics (or a compression of how humans manipulate objects), without any nice and highly compressed representations of that physics, or of general thought (e.g. math or general problem solving))

[ConcurrentSquared @ConcurrentSquar]

@willdepue i don't think it is that surprising, at least to me.
for robotics, you want a good world model. language is an absurdly good compression of human world-models; the only road-block is just that language is not the literal world, and so you do still need some pictures or video

[ConcurrentSquared @ConcurrentSquar]

@ZyMazza it's not true, in the same way naively comparing a mortgage vs. the annual income of a middle-class family (or student loans vs. a recent graduate) will result in similar oddness