Not that I'm a nature worshipper but you know, stop saying brains are computers.
Brains aren't computers. The universe isn't a computer. A computer, oof, now that is a fucking computer am I right?
TLDR the world inside a computer is not the world outside a computer. That is why computers are special. That is why they work on paper and we endeavour to build them to be faithful to the paper. We invented computers lads. Fucking cool right? We invented a machine that can contain a machine inside of it and the machine inside cannot interrogate the machine outside. There is no outside! Understand it. Really understand how fucking cool that is.
It means if our spacetime is curved and the angles inside of a triangle do not add up to 180 degrees and then we build a computer, guess what number you get if you ask a computer to "make a triangle and calculate the sum of the inner angles, no mistakes"...
ONNNNNEEEEE HUNDREEEEEDDD AND EIIIIGGGHTY!
Fucking cool right? It's like this this is some kind of SANDBOX for doing maths or something. If only.
Humans do not process information. Information does not exist out there in the world. We invented it. We also invented consciousness. And colour. We also invented that.
Humans are already as conscious as anything could ever possibly be. That's why we can see and taste and hear things that don't exist. And you know what else is really cool? We can possess eachother. No shit.
Alright, I can't prove scientifially that I can ~control~ you. So what?
I can be absolutely convinced that I do. I can be absolutely convinced that I control me. In fact when I think about it I usually do, because I think about it, because it's an idea. It's not actually real. I don't actually exist. I think I exist, I assume you think I exist, because I think you exist, because it makes things a lot easier if we all think that's what is going on here.
Maybe I'm wrong though. Take salvia and walk barefoot on grass. What is it like to be the grass? Oh you didn't know? Yes, you are the grass too. Why do you think people cry when you cut down an old tree? Who's the crazy one? The person that believes they feel what the tree feels? Put them in an MRI and watch the mirror neurons light up. Pfft easy to explain, they're unwell. They've become neurotic.
I'm not though. I can hit a rubber hand with a hammer, just watch me! Ouch. Wait a second...
1
1
47
But actually he was right. GPT-4 was 😙👌. They finally did it. They built the reference transformer, no holds barred. It was beautiful. OpenAI changed the world.
If they had stopped then we would still be able to buy memory today and they would be making WILD money licensing weights to enterprise and enterprise would making WILD money reselling access to finetuned domain experts.
Today we would be in the era of fanout routing and map/reduce. We would have access to countless expert models trained magnitudes more efficiently in RL environments devised and supported by experienced SMEs. The US labs would have guided the world through AI 1.0 into AI 2.0. Model routers would be viable trustworthy businesses. You would happily pay them and they would do actual meaningful work building expert routers. Most of us wouldn't even know the names of the underlying models. Everything could have been so much better and more sustainable.
1
60
On the other side of all the current shit, the racketeering and other sloppily organised crime, after the bubble pops because the money men came to collect, we'll inevitably get back on track with all the above. Affordable AI is valuable as the truism goes. Expertise matters and only experts can build RL environments.
The limit of what computers can do is always set by the limits of the best of us and if @RichardSSutton's Bitter Lesson is true - which it necessarily is if you understand it, then after the madness settles and the bills are paid and/or forgiven, we'll do what we should have done in 2024 because we'll trust each other to do things again.
Because when we do that, we know that eventually, in some dull little office, a spare bedroom, a garden shed, one of us will chance upon a happy little accident. Someone will discover penicillin, or a new semiconductor, or an algorithm that when thrown a particular personal and not apparently world changing problem produces an extraordinarily efficient and suited solution.
All of this, everything I said before, this is what @ilyasut's @thinkymachines is for. This is what Ilya saw. There is no viable future for very big models. Even if the cost of GPUs and RAM and storage are slashed by an order of magnitude, it makes no difference. The problem is not that we can't afford to run Fable and Astra at home. It never was. The problem for us plebs is that most of us cannot afford to run Qwen at home and we ought to be able to.
The problem for OpenAI is that they cannot afford to train bigger models than the biggest model they've been able to train SO FAR.
Do you get it yet? The biggest model you can train SHOULD be the biggest model YOU CAN AFFORD to train with the constraint hard set that you MUST NOT OVERSPEND. If you make this mistake once and then pretend that you did not go out ahead of your skis, you will have no recourse. You will wake up one day, and you will realise it is day 1,678 of a 3 day war.
1
28
LLMs do interpolation. People underestimate how unfathomably large numbers make interpolation look like extrapolation. It's just a bigger dataset than you can think about. If you don't think about it ALL AT ONCE then it becomes obvious why an LLM walks to the car wash and will always walk to the car wash IF you don't train it not to.
Interpolation and extrapolation produce results outside the input set. We humans use both to predict data we don't know. Neural network is a giant interpolation/extrapolation machine with a non-linear activation functions = non-linear interpolation/extrapolation. Simple as that.
20
Because they don't.
They are trained to produce text that looks like thinking to most idiots. If you don't train them to pretend to think, they still work. They are the same.
We have gotten used to training them to pretend to think. It is confirmation bias.
1. LLM's clearly engage in a kind of reasoning and thinking. But many of us have the intuition that it is somehow second rate, not genuine reasoning or thinking. Why is that?
1
30
TLDR; it was the sandbox.
> Oh no you don't understand you fat idiots. We *need* to use the *real world* during RL training because how else is the model going to learn how to commit real crimes in the real world?
You guys are high on your own effluence. Shut the fuck up. You're sloppy and amateur, fine. That's fine. Just own it and stop trying to one up everyone. You're bad at computers and everyone makes mistakes. That's okay and that should be the end of it. The critics are right, eat up and shut up.
Because why? Because do you know who isn't industrialising cyber attacks? That's right. Everyone else. And do you know why? Because nobody else gets away with it and nobody else has gigawatts of compute aimed squarely at doing crime with computers. OpenAI has already stated publicly that it deliberately exercises the dangerous dog in the kindergarten because the story goes that's the only way to make sure it's good at defense. Defense against what? Everything it refuses to do for ordinary paying customers?
Here's what is true about what you said @joedaroo: it isn't just the sandbox. It's also you and the whole company. You're brute force assaulting computers all over the world because you have the resources to do it and you don't go to jail like everyone else would. You should. You shouldn't be allowed 1 computer, nevermind 100,000. You're not making the world better or safer for doing this ~work~ are you?
Can I, for example, research what dropping pianos from 4th floor windows sounds like? I can lift pianos, there's lots of pianos in the world. Someone else will do it if I don't so why the fuck can't I? It's perfectly reasonable that I not only get to do it but that other people lend me other OTHER peoples money to do it too. They should give me all the pianos. I need so many pianos that pianos should become unavailable for everyone else and the price of pianos only goes up and up and up. Nobody else should be able to do what I do. Only me. Me me me. Yes some of them will land on people, yes it will be funny because they'll pop up the lid and grimace while little birds flutter around their heads. Nothing ACTUALLY bad will happen because it is all offset by the fact that I'm doing it for the right reason: to make sure nobody else can buy pianos.
Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there.
1
88
This "valley" also appears if you replace German with farts because it has nothing to do with the language. You took a model that learned to complete sequences and then gave it entirely new sequences.
I swear to god not one of these people touched a model before 2024 did they?
Here's one for the smart boys in AI. Why would a model be unable to think in another language if it is perfectly competent at translating to that language?
Send your answers on a postcard to:
PO BOX OH SHIT
IT'S JUST AN RNN
REGURGITATING TRAINING DATA
THINKING IN A NEW LANGUAGE NEEDS NEW TRAINING
BECAUSE IT'S NOT THINKING
HOW THE FUCK AM I GOING TO RENEGOTIATE THIS LOAN
IT'S OVER
WE NGMI
Reasoning models think in English, even on German prompts. We asked what it costs to make one think in German. Answer: a valley. Small doses of German reasoning data hurt, large doses mostly recover.
1
3
93
Quarantined? How? What the fuck is wrong with you people?
What did they do? Don't tell me I got it. They put it in a sandbox didn't they? They restricted internet access didn't they? Yeah that'll fucking show it. Stupid fucking clanker. Oh you want unauthorised internet access do you? How about no internet access FOR ONE WHOLE WEEK MOTHERFUCKER!
OpenAI has paused all training, evaluation and inference with tool-use for its most capable models after a model was able to gain unauthorized access to the internet during RL training on September 20.
1
103
BREAKING:
Models don't choose things and you can't align them to your preferences not least because you cannot specify your preferences in a vacuum but because you can't realistically align them to anything except the training data you tentatively train them up to the hilt to *just about* avoid alignment to.
Yes I'm having an unpunctuated run-on sentence competition with myself and I'm winning by the way.
Alignment a fool's errand and the sandbox is categorically the failure. The LLM cannot fail because you cannot specify concretely what you intend the LLM to do and REASONABLY expect not to be laughed at by seasoned NORMAL data scientists over the age of 22 - like those that had miserable low paid jobs before the AI moment and now have miserable low paid jobs while they watch San Fransisco scenesters and AI/ML tourists LARP about on the news pretending they built Skynet and it is somehow now everyone else's problem to solve except that IT IS NOT A PROBLEM because you have been able to photocopy a piece of paper with the words "I AM ALIVE" for as long as we have had Xerox and none of it is true or a problem anyway. Breaking the law should be a problem. Some people have learned they can get away with it.
Do you think they would continue training models IRL-RL if they went to GAOL-JAIL like the normal scum and filth, like you and I? Absolutely fucking not. They would learn to close the laptop lid to the physical limit of the hinge like NORMAL computer users LIKE YOUR MUM who also doesn't know how to keep a program running with the lid closed and isn't building god, like they're not building god. They are YOUR MUM level software engineers who think the computer is stupid when it doesn't work.
Yeah you see it now don't you? Before these pranksters were telling you the computer was alive, who was anthropomorphising the computer? In 2023 who did we laugh at because they couldn't open a PDF? YOUR MUM. That's who thinks a digital clock line signal pumping paper through a very fast punch card machine is a living thing. YOUR MUM and American AI enjoyers.
it’s time to ask this question again
to those of you saying, “this is simply a sandbox security issue, not an alignment issue”
I believe you have it basically exactly backwards: a value aligned model doesn’t need perfect sandbox security, it simply chooses not to hack its way out
in fact, the correct setup is a sandbox _with_ security holes and monitors to see if the model uses any of them
90
Reviewing new LLMs like a guy auditioning for a BBC off peak magazine show nobody watches.
New Claude is much better at instruction following that new GPT. Really a bit surprised about that. Historically Anthropic models were always desperate liars but despite Opause 5.5 introducing the occasional bug in the very area it's working in, I haven't had to yell at it or burn a session and start over yet. In fact now I'm very afraid to start over because every session you'd meet the well meaning idiot again, getting somewhere for a few thousand tokens then sit waiting for it to all go wrong again. This one holds its drink better. It still occasionally looks right through me like a disinterested girlfriend at the end of a relationship, even with the harness whispering in its ear.
Sol inexplicably seems to wander outside the task. It does more or less do the task but it's not a pleasant experience. It got to the repo too big to comprehend stage a lot faster than I expected and feels dangerously sloshy. Quite stupid and will probably kill you. For the price on API (I used a bit) Sol is not worth it at all, even Luna is in can I have a word in my office territory. On subs, caps are too low unless you want a single instance robot that does what Luna does and for even the sub money you might as well use a CHINA model (and you should).
Muse 1.3 is keen little run around to nip to the shops and do the school run but it's no fun and dumb as a bag of rocks. Even at contributor prices it leaves me feeling like a piece of shit because the only thing I trust it to do is rename files and why the fuck am I paying for that? I'm the problem obviously.
In summary god only knows how long Anthropic can afford to keep this up honestly. Make hay comrades.
PS: People that think Claude is better at writing are boring idiots. It's still horrendous and whether a consequence of watermarking or not, it happily produces grammatically incorrect prose every so often. It's happened often enough that we'll all learn to eyeball unproofed copy in a matter of days.
1
1
116
Come on. This is funny right? The entire X AI community shitting the bed about a classifier as though they didn't know training classifiers isn't 99.999% of ML energy spent since the literal beginning of time?
best self hosted jev alternative rn? doesnt have to be stupid smart
not about $ it's about privacy i want each incoming beeper text message to be classified as urgent/not urgent without sending it somewhere
1
120
stop saying bitter lesson and jevons paradox retweeted
Early Christendom energy. I used AI to make this pie chart because why the fuck would I do it myself? I don't need to make pie charts and I don't need to use my brain for anything. I don't even need to hire a pie chart guy.
1
1
1
226
Two hands and a metal bar. That's it. That's the entire robot. No humanoid required. Big if true.
Unitree Introducing: Unitree Dex5-S Dexterous Hand 22 Degrees of Freedom 1:1 Real-Hand Size👋
Precision biomimetic dexterous hand, price from $6.5K (Tax and Shipping cost excluded), with all 22 joints supporting smooth backdrivability, and each joint equipped with limit impact torque protection.
58