bubblyworld

Figure eight, bowline, slipknot, clove hitch, trucker's hitch (not really a knot but useful) and a sheet bend will cover you for like 99% of use-cases, including climbing haha.

I really liked this. It's a nice meander through the basics of lambda calculus. It's striking to me how much insight you can get by visualising stuff, even if it's not a great visualisation and you're doing something super abstract. Perhaps a lesson to take into my own programbles...

Sure, but people are allowed to have their own opinions too.

Lol, what is up with everyone assuming there's no learning curve to these things? If you applied this argument to literally any other tool you would be laughed at, for good reason.

Have a read about kin selection, for instance. There are lots of facts about e.g social insects (like the male/female/sterile ratio iirc) that were thought to only be explainable by group selection, but later were found to work out purely from considerations of gene selection. I'm just a layperson though, would highly recommend Dawkins writing on this stuff. It's much less polemic than his writing on religion, if that puts you off at all.

I think the basic criticism of group selection is that at an individual level it can be beneficial to go against the group's goal, and you need to explain why that doesn't happen (even very small advantages are rapidly boosted by selection). This is one of the great puzzles of selection - how does cooperation evolve from selfish interests? The theory of group selection just asserts that it does, which is not very satisfactory.

Yeah, absolutely, I spent months preparing for the ICPC with my team and ended up scoring a paltry 3/12. A week would likely not have helped us at all - we simply had no idea how to approach the rest! Top teams are on another level.

Compute and such is a fair point but that AI is here at all is mind-blowing to me.

My understanding of the "modern" point of view is that selection acting at the level of the gene (not the organism or group) is sufficient as a theory. The Selfish Gene and The Extended Phenotype go into this stuff in a lot of detail.

Totally agree with your first sentence though. And even if there is a plausible adaptive function it may have only been adaptive in the past, or might be a side effect of some other adaptive function (see sickle cell anemia), or a host of other possibilities.

Purely academic, at least if you mean the result itself. People are bringing up all sorts of apologetic stuff about how it might be useful in the future, and as an ex-maths-phd I can sympathise, but to me this seems more like a large-scale bookkeeping exercise than a conceptual breakthrough. It's a very nice result though =)

Your reward system is a fragile thing that is easily broken. Just look around you next time you go anywhere. We live in a society of smart phone addicts. That said, the examples OP gave of tunes while exercising seem pretty benign compared to superstimuli like social media or drugs...

I think people use "dopamine" in quite a loose sense to refer to reward centres in the brain and habit formation, not literally the hormone (though it's related). You might have better luck digging around those topics.

Wild, I found it hard to believe that a 4b model could beat sonnet-3.5 at anything, but at least on the vision arena (https://lmarena.ai/leaderboard/vision) it seems like sonnet-3.5 is at the same ELO as a 27b gemma (~1150), so it's plausible. I guess that just says more about how bad vision LLMs are right now that anything else.

I think there's a bit of "Chesterton's fence" with these issues. If you don't know enough about the ecosystem to appreciate its complexity (aggressive animals are a tiny part of what's going on there) then you definitely shouldn't be allowed to remove or change it. Human ecological interventions have a bad track record.

I suspect (to use the language of the author) current LLMs have a bit of a "reasoning dead zone" when it comes to images. In my limited experience they struggle with anything more complex than "transcribe the text" or similarly basic tasks. Like I tried to create an automated QA agent with Claude Sonnet 3.5 to catch regressions in my frontend, and it will look at an obviously broken frontend component (using puppeteer to drive and screenshot a headless browser) and confidently proclaim it's working correctly, often making up a supporting argument too. I've had much more success passing the code for the component and any console logs directly to the agent in text form.

My memory is a bit fuzzy, but I've seen another QA agent that takes a similar approach of structured text extraction rather than using images. So I suspect I'm not the only one finding image-based reasoning an issue. Could also be for cost reasons though, so take that with a pinch of salt.

Ask a biologist or ML enthusiast; complexity is the natural outcome of a period of mutation/selection. It's a mistake to criticise the result on the grounds of said complexity, I think - rather ask what environment the complexity evolved to "survive" in.

It's boring common sense but don't take a bromeliad into the desert.

My belief is that all codebases that live long enough will develop this level of complexity with time. The only exceptions imo are where the codebase solves a very particular, fixed problem, e.g. a library for A* pathfinding or whatever. But maybe not even then.

Sorry, on rereading this I've actually written a bunch of nonsense. It's true that each choice of n gives you back a metric, but the |x|, |y|, |z| actually refer to the components of your vectors. So in 1-dim space you just have |x|, in 2-dim space you have |x| and |y|, 3-dim space has |x|, |y| and |z| and so forth.

So in 3-dimensional space, for n=3, the "length" of a vector u=(u_x, u_y, u_z) is: d(u) = (|u_x|^3 + |u_y|^3 + |u_z|^3)^(1/3)

In 2-dimensional space, for n=4 and u=(u_x, u_y) you get the following: d(u) = (|u_x|^4 + |u_y|^4)^(1/4)

If you want to use the norm as a metric to work out the "distance" between two vectors u and v then you just compute d(u-v) where u-v is ordinary vector subtraction.

There's a stackexchange thread that touches on the first two questions. It's got the integral form of the circumference calculation, but I doubt there's a closed-form solution in general: https://math.stackexchange.com/questions/2044223/measuring-p...

It's an infinite family of metrics - you provide an n (a positive integer) and get back a metric.

So if you pick n=1 you get d(x, y) = |x| + |y|, which is the taxicab metric. You can apply this metric to a Euclidean space of whatever dimension you like, just substituting the appropriate definition of |x| and |y|. For 1-dimensional space you would use |x| = abs(x[0]), for 2-dimensional space you would use |x| = sqrt(x[0]**2 + x[1]**2), etcetera. Hope that helps.

Depends on your circles. Marijuana use is super prevalent amongst climbers (even while climbing, for better or worse) and it's a very active lifestyle. Same for surfers in my area. I don't think it's a given that people who smoke are less active.

I mean, it really depends. I'm writing models for some energy traders at the moment, and things were simple enough in the mvp that we could hand-roll our optimisations. But when we started having to map real contracts it got out of hand super fast. Linear constraints solvers have been a godsend, it makes adding new types of contacts and constraints really easy (and supports almost arbitrary kinds of optimisations).

It's just never as easy as "don't do this" or "always do this". But yeah, I agree with your gist, right tools for the right job and all that. Most of the time you don't know all requirements up front and it can make sense to hedge your bets in the interest of speed until you do.

There are other aspects to maintenance, like requirements change. In this case it's trivial to change or add new constraints to a constraint solver, whereas even small changes to a typical DP problem can require a total rethink of the approach. Extending the analogy to other kinds of dependencies left as an exercise for the reader.

Point being that software has many dimensions. Reducing the use of dependencies to fear of learning or thinking is a bit reductive in my opinion, even for stuff that seems simple initially.

People are allowed to change their minds, or even (god forbid) have contradictory and/or nuanced beliefs.

LLMs can clearly make use of tools, unlike your cat. The claim was that they cannot do backtracking natively, which may or may not be true but it's irrelevant because they can do it through code.

Who said anything about software engineers?

That assertion is not what I was referring to. Anyway, I'm not really interested in nitpicking this stuff. Engage with my initial comment if you actually care to discuss it.

Lol, okay. Serves me right for feeding the trolls.

I think you are applying logic and demand for rigour selectively, to be honest. Not all arguments require formalisation. I have presented mine - your linked logical analyses just aren't relevant to modern systems. I said nothing about the logical steps being wrong, necessarily.

If you want to understand SOTA systems then I don't think you should study their formal properties in isolation, i.e. it's not useful to separate them from their environment. Every LLM-based tool has access to code interpreters these days which makes this kind of a moot point.

Something to consider is that while it's really hard to implement a decent NN-based algorithm like AlphaZero for your game, you get the benefit that model checkpoints give you a range of skill levels to play against as you train it.

Handicapping traditional tree search produces really terrible results, imo. It's common for weak chess engines to be weak for stupid reasons (they just hang pieces, make random unnatural moves, miss blatant threats etc). Playing weak versions of Leela chess really "feels" like a (bad) human opponent by contrast.

Maybe the juice isn't worth the squeeze. It's definitely a ton of work to get right.

The sun doesn't solve any problems in a useful sense. I'm talking about life! I thought that would be obvious haha.

Simplicity is brittle. Perhaps we should take a page from nature. Nothing in nature in simple, and yet it has built some of the most robust systems (by far) that we know of.

Can your software run for millions of years?

Under the FAQ:

> Previous chats with no additional activity will not be used for model training.