stop saying bitter lesson @localoptimiser

There is no deal. He is an internet retard without any money. He's not even competent at pretending to understand what he's doing. Free money for FIVE YEARS? Okay buddy. I'll take 100,000 and put it in AI stock. If you're right I'll pay you back by 2027 anyway and if you're wrong I lose nothing.

@_esk_kse_ @SHL0MS @George_3d6 He isn't loaning anything to anyone. It's LARP. He doesn't have his own money. He uses his mom's store card.

@SHL0MS You get all their money? You're going to lend your own money at 0% for the next 5 years? Okay then sign me up. My p(doom) is 100%. How do you want to do this? IBAN?

Same experience yesterday. Absolute piece of shit model. In the official harness too. Lies and doubles down when caught. Tells you you're stupid. Blames you for the things you watched it do. Tells you what you're actually looking at on the screen isn't there in real-time. Dumpster fire of a model. Anthropic should be fully liquidated to save the oxygen for OpenAI. OoenAI should be liquidated too, but at Anthropic deserves to die first.

@BlueWhaleFlys @hive_echo @robertskmiles @luke_chaj "the network is also approximating the structures and processes of that of the brain" Zero evidence of this ever being the case. Pure wishful thinking.

@Michaeljhandy @basedjensen At this point they should download Kimi and change the system. It's the only way they'll ever make any money.

@lacker @basedjensen If you're a multi-million dollar company, nevermind billions or trillions, it is categorically a basic skill. A normal business wouldn't even get a bank loan if they were this fucking sloppy.

@ramathakovesh @basedjensen They're all bankrupt and they big databases you actual certifiable retard.

@8teAPi No it indicates the hidden layers in models have successfully learned long range features. You know, because that's what deep learning was invented to do?

LLMs could learn mathematics but they tend not to learn it because in practice training LLMs induces them to memorise sequences and produce interlaced completions. That is why they're especially bad an arithmetic until you do one of two things whilst training: 1. You overwhelm them with arithmetic or 2. You make them extraordinarily large so as to memorise more answers. There is a 3rd route too, you lie. You serve a model behind an API that has a middleware to detect math problems and inject answers. I don't know for certain that Anthropic did this overnight when Claude 3.5 was the model of choice and spontaneously became miraculously accurate and fast at specifically arithmetic. I likewise don't know Gemini only yesterday would do this if you happened to refer to two numbered points earlier in a terse response like "give me 1 and 10" and it produce "11." So there's three options then. It's only good at math, it's very big, you're being misleading by providers. But I ask you, how big is a calculator in silicon? How complex is that circuit versus a locally hostable LLM? If models learn how to *do* arithmetic, why haven't they learned to straight up model the integrated circuit from a pocket calculator? I mean, holy shit they're literally trained on the actual schematics of calculators and electronics simulators. Why aren't they yet understanding they could rearrange a morsel of their own weights to become state of the art at adding long numbers together internally or, say, counting to 200 without quitting or tripping up more often than not? There are some arguments. Now explain how even after gigawatts have been spent on individual models none of the above has happened yet.

@BlueWhaleFlys @hive_echo @robertskmiles @luke_chaj Brains do not multiply numbers. They do not operate on numbers. They are not computers. People who only know computers think that everything else is a computer.

@BlueWhaleFlys @hive_echo @robertskmiles @luke_chaj No it is not.

LLMs don't learn to infer. Sorry. They can't understand. They memorise features that are common in the explanation of things and when those features do not dominate the corpus they do not learn who the mother of Tom Cruise is. They're trained on the facts but they cannot connect them. They cannot distinguish an upturned beaker from a joke beaker with a sealed top and an empty bottom, because such a silly example has barely been described relative to the abundant examples leading through [help -> problem -> trick -> joke -> practical joke]. They're vector databases with novel and yes at times very useful query interface. They do not understand anything. They do not think. They perform as per the training regime.

It's not compression. Stop it with this fucking gibberish. Genuinely tell us what happened to you? You had a bright career ahead of you and you became a car crash overnight. You're not thinking about anything you're saying anymore. You're putting every ounce of energy into defending this ridiculous position and it's fucking tiresome and heartbreaking. I've known loads of what I thought were promising people go bananas but they always had either blindspots or things I sheepishly turned a blind eye to. But you don't do politics. You're not racist. You don't seem to pick on anyone at all. And here you are presenting with a Phineas Gage tier disfunctionalism. We know exactly how LLMs work. We know exactly they work. At every scale they do exactly the same thing. That's what's so great about them. We know there's no such thing as emergence. We know we just make bigger models to memorise more things. We know all of that because we know how expensive it is to keep doing this. The costs are meaningful signals. We write pixel shaders and we autotrain the parameters. That is what ML is and we know that is what it is. Stop lying to yourself and stop lying to everyone naïve enough to look to you for inspiration. I know you know this is correct so tell the truth you miserable coward.

@robertskmiles This is terrible and you have no shame Robert.

Recursive self improvement implies the model goes into the world and gains capability volitionally. You can't claim that training is recursive self improvement otherwise the word doesn't mean anything. Kind of like you calling yourself CTO doesn't mean anything. Because you're not a CTO.

@NoOnemklz @dhtikna Children who write functional drivers in under an hour while the adults make weight trackers nobody uses?

@reyhenrick @dhtikna They're all benchmaxxxed. What the fuck do you think models act are? Have some taste.

The tap and hold trend is what you get for hiring a one trick pony with a hand scrawled profile picture.