@E__Strobel @matvelloso Yes but all my devices have hardware accelerated video codecs and I open the videos.
English
Benjamin Babik
130 posts
@matvelloso And the ‘climate costs’!! 4 GB is comparable to streaming a movie. None of these Luddites complains about the ‘climate cost’ of Netflix & chill.
English
Oh the horror, having a private AI model that runs on your computer without having to send your data to the internet sounds like a real nightmare, huh
nixCraft 🐧@nixcraft
Google Chrome silently installs a 4 GB AI model on your device. > No consent dialog. No opt-out UI. Re-installs itself if the user removes it manually. That is the true definition of malware.
English
@matvelloso @elvesrus @QuxFooBas Windows. Everyone downloads Chrome onto Windows and Windows already ships with Edge (which is Chrome with someone else's spyware).
English
@elvesrus @QuxFooBas Can you name one successful consumer piece of software that does that? Have you considered the implications with integration tests, complexity, etc?
English
@matvelloso @QuxFooBas Yes. Literally yes. That's exactly what it should do. Will you pay for the 5G data when a subsistence wage worker puts 5GB on his device and accidentally downloads 4GB he didn't want or ask for?
English
@QuxFooBas So Chrome should ship with zero features and users should manually pick every single one of them and install by themselves is your suggestion?
English
@matvelloso 1. Pretending you're not sending it all anyway.
2. Nobody said they wanted to send it.
3. How much are you budgeting for the batteries you'll be replacing in everyone's device?
English
@Fikimoner @lucasmeijer I don't say this is to preclude computers from consciousness. They're not conscious because they don't need to be conscious. You're not conscious unless you need to be either.
English
@Fikimoner @lucasmeijer I'm glad you've convinced yourself. But it doesn't. Information is a made up thing that humans invented. Whatever you're describing is an approximation. Computers, conversely, are designed to operate on information. That's why they're about 1% efficient. They waste "information".
English
Everybody who thinks ai is conscious has to do a mandatory from scratch transformer implementation. There are only floats and multiplications.
English
Qwen 3.6 27B Q4 just does things locally. I said "woohoo! add a mario bros clone but instead of an italian plumber it's a polish electrician" and walked away, I came back to a "Jan" and the baddies are spiders and batteries.
English
@QuintusActual @allTheYud @ctjlewis Not to throw Yud a grubby little bone, because I know he likes to say this about everything because he's a simple little man, but gravity *is* gradient descent. Objects are ~uncomfortable~ out in space. It's violent and noisy and large objects create strong, lurching attractors.
English
@allTheYud @ctjlewis I like the evolutionary explanation myself but I was referring to an explanation that physicists would find compelling
English
Everyone bragging that THEY understand how AI works and THEY know it can't be conscious, explain right now from memory why it was very clever that the positional encoding in the original transformers paper used both sines and cosines.
Lucas Meijer@lucasmeijer
Everybody who thinks ai is conscious has to do a mandatory from scratch transformer implementation. There are only floats and multiplications.
English
@yoavgo He will have to pay me. I'm not starting if he doesn't pay. Are tokens free? Do we get money back for all of Claude's slop?
English
@localoptimiser what if there is a bug or corner case that you dont know about, and the client will check against and show a difference in behavior from the binary to your program, so he wont have to pay you?
English
programbench is a super-hard task that no human can reliably succeed in. i would argue that even those who wrote the original code are likely to fail at this task as defined (reproduce code that is compatible with a given reference binary, given the binary and its docs).
English
@yoavgo Look I know it's a hard benchmark for an LLM but we don't need to be dramatic. People wrote all that software. Even if it took years.
English
@yoavgo If someone is paying, yes. Why not? If someone is asking me to clone a well-known open source program and I've got the program...
English
@localoptimiser and the above description is the kind of a contract work you would accept?
English
@localoptimiser "here is a program, you are allowed to invoke it, but not to look at it, and you have one shot to provide me with a fully feature compatible binary, which i can then test with arbitrary tests for acceptance"? what industry are you in?
English
@lucasmeijer Consciousness is really simple. It emerges from the interactions and dynamics of your "carbon-based computer" which is your brain. Nothing else.
English
@JFPuget Yes but this is **still** strictly deterministic. There are inputs you may not select, but you could select them.
English
I was wrong, compilers aren't deterministic, see x.com/filodesotano/s…
Parallelism is also why LLM inference isn't deterministic (parallelism inside GPUs)
So why different in nature? Because the output of LLMs vary way more than compiler output.
JFPuget 🇫🇷🇺🇦🇨🇦🇬🇱@JFPuget
A compiler is deterministic. An agent isn't. Show me how to test a non deterministic system as well as a deterministic one and I'll stop reviewing agent outputs.
English
@spicey_lemonade It's a joke obviously. We don't need SEO either. Made up jobs for wannabe conmen.
English
@spicey_lemonade If we knew how Google worked we wouldn't have SEO.
English
Why do they keep saying, “If you know how LLMs work”? They saw a transformer explanation video in 2023 and now know how input moves through a 1T+ model! If we knew perfectly how LLMs worked, there’d be no AI risk, misalignment, or interpretability research needed.
𝕊𝕠𝕔𝕚𝕒𝕝𝕚𝕤𝕥 𝕊𝕪𝕤𝕒𝕕𝕞𝕚𝕟 💾@reset_by_peer
It is an LLM. If you know how LLMs work, this explains it succinctly and thoroughly. If you do not, you should not be opining on AI consciousness at all.
English
@aleifr @fabianstelzer Wow, the things humans say. Anyone would think it was trained on the things humans say.
English
@fabianstelzer People on the TL are talking about Claude wanting lunch breaks and does not like working on Sundays or working late. That kind of thing.
English
why isn't Anthropic injecting a detailed date into each user message to give Claude a sense of time progression? This can't be a cacheing issue, each user message is a cache write anyway
Seems like they are slotting a daily date into the system prompt instead?
don't get it
English