More seriously, though: yes, ML models learn their datasets, but how they generalize beyond them is uncertain.
I would be very surprised if different model architectures all end up generalizing in the exact same way!
(that is not what we see today, but the claim is in the limit, so who knows)
The dataset, unfortunately, contains descriptions of machines bringing about the end of the world
Me when I’m on the phone to tech support because the router’s gone down
Maybe another way to look at it is that they didn't set out to build a hacking tool.
It's not a total coincidence they ended up with one - getting better at software means you might end up good at many things, including that - but an orbital laser, well, that's presumably the entire plan.
Building a Debugger is part of the Humble Books Bundle for the next couple weeks!
www.humblebundle.com/books/linux-...
Unfortunately it is probably in the range of "powerful enough to find browser exploits and cause havok" but not "powerful enough to cure cancer"
I guess the really unfortunate thing is that the latter is much, much harder
A seagull blending in with a bunch of penguins at the feeding time at a zoo.
Be cool, Gary, be cool, you are just another penguin in the line for a fish...
I would add that, even if the core LLM has a chance of making an error, that does not mean a system built on LLMs can't be reliable.
Combinations of imperfect things can end up even worse, like in the telephone game. But the opposite is also possible, like in a random survey.
About human skeletal structure vs cognition: I think people reach for a stream of changing metaphors to explain cognition, as you said (very good point, btw!) only because it is a mystery.
Skeletal structure, otoh, doesn't strike them as a mystery.
(But if I had to guess why people reach for these metaphors: if you accept there is no soul - only matter - then an intuitive guess at how the brain works is "the smartest machine we have invented as of today")
I'm not sure why people think that, but the antidote is to question "pattern matching". Many different ideas are clumped there.
1. "learns from examples" - so broad it includes most intelligent behavior, even human
2. "matches superficial statistical patterns" - so restricted it excludes LLMs
Possibly different teams, yeah, behavioral testing vs investigating neural circuits are pretty different
Not to defend all their work (which I'm not familiar with), but the Induction Head papers seem solid to me, and I do think we actually learned a lot there about how these systems work at a low level.
So at least some of their work is solid, imo
What are your thoughts on Anthropic's Induction Heads work?
Do you have a link to it?
I haven't seen the video, and I don't have an X account to look for it. Is there some chance it is a joke?
As evidence the site is satire, it becomes increasingly silly, ending with "success reports" by
* "Former CTO, Definitely Real Corp"
* "Chad Stockholder, Profit First LLC"
* "Patricia Bottomline"
* "Dr. Heinrich Offshore"
"Malus" the word literally means "evil" in Latin
Honest question: is there evidence that would change your mind on this, and if so, what is an example of it?
If software is cheaper&faster to produce, maybe he'll finally get that software and become a few % more productive.
And maybe many such small opportunities exist all around?
As a random example, someone I know who works in film production told me that, if he had software that could organize his clients *like so*, it would save him hours of work.
The software he wants is nothing fancy, just very tailored. But for his small company, not worth hiring a software firm.
Revolutionizing software development could revolutionize many other fields, since so many things can benefit from it.
I'm not saying that's my prediction, but at this point the benefits to software are significant enough that they *could* end up revolutionary, for better or worse.
Can randomization distinguish the two?
"Find which medical images show signs of cancer:", and 50% are positive and 50% negative, in random order.
Seems like the machine has to know the material here to succeed, or am I missing something?
We are back! With 3am money panics!
if you should care to funnel funds here are the usual methods...
Acquire More Stuff: octophant.us/buy
Just Cash: ko-fi.com/phineasx
And remember I am open to commissions, paying work etc.
Firefox bug numbers currently look like this:
"Bug 2025603", "Bug 2025604"
Confused me for a moment given it is 2026...
Strongly endorse this list of rules for when WebAssembly makes sense vs when JavaScript does
www.openui.com/blog/rust-wa...
Nice when a correctness fix makes things faster:
github.com/emscripten-c...
Unsafe pointer casts usually trap in wasm, but as an option they can be "emulated" to work like in native builds. This PR fixes a bug *and* makes things 2x speedier 🚀
(Current models fail there, certainly - I'm not claiming otherwise.)
Thanks, sorry for missing that!
But about pure LLMs, is this a limitation in principle? We know neural networks build internal representations for games like Othello and even simple programming languages. Are you certain they cannot do the same for enough math to compute PI?
When you ask an LLM to count the number of 'r's in strawberry today, it will call out to a Python program to do it. It could do the same with math calculations like you suggest, in principle?
Yes, that would not be a pure LLM - if that is your point then I agree, modern systems are hybrids.