James MacGlashan

Schools are not like a fixed dataset. I think you underestimate how much work static datasets that you iid sample from does.

School is a guide in a continual correlated environment with imperfect information transmission.

By all means, prove me wrong and do it.

Yes, painfully aware. There are many open problems remaining to be solved.

My compute time and memory requirements do not increase rapidly as a function of the length of time that I've been alive.

I also do not learn by fitting pre-supplied stationary datasets.

Today is my birthday (turning 29), which is the only time all year that I ask for a repost 🌿
I hope you get lost in fantasy worlds as much as I do 🔮

Yeah, we can always specialize when it warrants advantages, but if we’re in the specialization game why wouldn’t you also specialize the controller?

Especially if that one embodiment is sufficiently humanoid to interface with the world all the ways we do. Who cares after that.

For this same reason LLMs are fantastic for asking about well documented procedures and putting answers in the context of your question.

This *is* useful, but a software engineer it is not.

People underestimate how much code is on the internet. Most “amazing” examples are programs that have a 1000 instances on the web.

LLMs are great at templating, so they can do variations on those and “search” for them. But it makes the demos deceptive.

Even if @rust-lang.org never fully takes off and becomes one of the more dominant languages, I think learning it was worthwhile because it made me a better programmer in other languages.

It forces you to write in a way that is inherently clearer and easier to maintain. That translates.

I of course don't have the context you do, but I think those statements are compatible. It sounds like they're telling you that you can't run yourself ragged always putting others first. But that doesn't mean you should ignore others.

Just a thought. I could be wrong.

Oh, beware the ending. Horses are doing stuff.

YouTube video by Tracy Kennedy Silicon Valley - sales and product

www.youtube.com/watch?v=_NeJ...

The correct way has been demonstrated: No introductions. When you speak start by saying people call you "Jan the man"

Yes, but have we considered writing another book about how Biden was old?

If only there was some way people could have known he would be bad.

We need a completely different system. The old ways have not scaled and have been breaking for some time now.

But most people don't want to commit to fundamental changes.

It's maddening that we could easily have nice things if our government wasn't overflowing with corrupt assholes.

Hollow Knight is emotionally resonant in a kind of passive way. Music environment and subtlety.

Nine Sols has a more explicit story that hit me in the feels hard when I got the true ending.

If you want a game of a similar style (2d metroidvania): Hollow Knight or Nine Sols.

These games are mandatory as far as I’m concerned. Both well designed and emotionally resonant.

They’re both hard. Nine Sols, does have difficulty & accessibility options to make it easier if you need to.

Kernel errors can certainly make that happen, as can hardware errors.

I don't think there's a lot you can do in high-level software to defend against those errors though.

Google's solution is just replace the whole server because that ends up being faster & cheaper!

Such a great show. Varied messy characters that all embody the geek passion. I never expected show business to nail it so well.

There are more reasons these technologies are stepping stones to what we want, but we'd have to get too technical and long-winded for bsky.

What I hope to convey is that researchers agree with you and have been working on it. This tech follows from those efforts and doesn't mean we don't agree.

Here's a start for understanding some of the reasons why these came first.

LLMs because how do you think you're going to tell robots what you want them to do and have them respond to you?

Image gen because being able to model the world is extremely important to figuring out how to act in it.

I agree with this quote.

However, many people assume that researchers are not interested in it. That couldn't be further from the truth.

The amount of work motivated by robots/AI doing chores is boundless.

We are here only because writing/image gen is easier and a step toward those goals.

1/3

The problem might be that a and b are undefined. If you specify actual numbers (for which the integral is valid) like 1 and 10, you get an answer.

Or if you want an indefinite integral, you can just do: `integral of 1/x` and it gives you the indefinite integral.

I feel like there is something meaningful about these results and the fact that a (general) value function is a forward trace (when scaled by (1-gamma)).

I have yet to see an LLM do well on what I'd need here except for very simple functions, but I certainly wouldn't mind if they could!

If by "real AI" people mean "has personhood" I do not agree with ceding the term to them. That's not what AI has ever meant and it is most definitely not the goal. There is also a lot more to do that is well beyond LLMs but still does not have personhood.

(Just be careful not to fall into the trap of writing a singular entry point that does everything, but provides no flexibility. That may seem easy because it's only "one thing" but is not actually easy because that one thing does a slew of things you'd have to describe for someone to use right)

As a related rule, write a tutorial for how to use a collection of code. If explaining how to use it is hard, your interfaces are bad.

It can be better to write the tutorial earlier to figure out what interface would make it easy to explain. Then design it that way.