mhjkl

11 ms·

1.
by mhjkl 1mo ago
If you adjust for writing speed, this model is already more efficient than humans. It can run at 233 tps on a 5090 which is roughly 10k words per minute compared to human typing speed of 80 wpm. If you duty cycle it down from 575 watts of 5
2.
by mhjkl 1mo ago
This will make more money for Claude from individuals unofficially acting as meat puppets in more inefficient workflows
3.
by mhjkl 1mo ago
All big LLMs already visibly watermark all their text with easy to detect annoying phrases and turns of speech that everyone is already sick of hearing. Why do AI companies keep making their products worse to appease anti-AI, it’s not like
4.
by mhjkl 1mo ago
I’ve seen Chinese open weights models say “can’t use this if you’re in Europe” in their licenses, so I doubt they would invest too much into complying
5.
by mhjkl 1mo ago
You must not use any open source software, because much of it comes with a whole paragraph of ”IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY“
6.
by mhjkl 2mo ago
If you’re on AMD, use llamacpp with the vulkan backend. For the 35B I get around 115 tps without mtp and 150-170tps with it. Also the pcie slot you use affects it
7.
by mhjkl 2mo ago
I'd say Anki has more friction that writing physical cards though
8.
by mhjkl 2mo ago
How do you make these compressed prompts like that
9.
by mhjkl 3mo ago
This comment would go hard in a sci-fi novel 5 years ago
10.
by mhjkl 4mo ago
During the dot com boom they said data centers consumed 8% of US electricity. It was later proven to be a lie, created with fake studies by coal companies to trick the US into building more coal power plants using environmental activist out
11.
by mhjkl 4mo ago
Reminds me of the “Has The Turing Test Been Passed” website. It says no, but if you read on they cite “The relatively minimal funding allocated to AI research” as one of the reasons AI hasn’t been achieved “yet”. Website stopped being updat
12.
by mhjkl 4mo ago
Nim
13.
by mhjkl 5mo ago
Antis: "If AI is so useful, where are the AI shovelwares? Where are the AI open source contributions? It's all hype" 6 months later: Matz used Claude and now Ruby runs 86 times faster after 1 month of work At this point it&#x
14.
by mhjkl 5mo ago
Reminds me of the Enron dataset
15.
by mhjkl 5mo ago
They didn’t need to wrap it because it’s modular arithmetic so the result after casting to int is the same regardless of wrapping behavior. 4294990000 after wrapping is 22704 and 4294960000 - 22704 = 4294937296 which is -30000 after uint to
16.
by mhjkl 5mo ago
But the TSTMP_GEQ macro casts the difference to int, so any number above the signed integer limit (about 2 billion) becomes negative and the comparison returns false as they said
17.
by mhjkl 5mo ago
OP had no problem with pointers prior to trying C++. I think there is a case to be made that C(++) makes pointers unnecessarily confusing and there is no real disconnect between understanding pointers in theory and in practice otherwise
18.
by mhjkl 6mo ago
You can remove and replace the switches too
19.
by mhjkl 6mo ago
What exactly broke? I had this problem recently and thought the keyboard was done for, but turns out you can also replace the switches (not keycaps) and that solved the issue in my case. You have to be very careful though since it is very e
20.
by mhjkl 6mo ago
Couldn't he just subcontract it out to whoever he needs to make his chips? Doesn't need to do anything himself
21.
by mhjkl 6mo ago
Afaik sodium batteries are much safer than li-ion and already in mass production. Unless Donut scales up really quick I think it will be more viable to just use sodium batteries for safety in the medium-term future
22.
by mhjkl 6mo ago
There's a big difference between things like Metaverse, social media and AI. The former are mainly entertainment/communication products that rely on network effects, so if people don't like them they're bust. AI is a cap
23.
by mhjkl 6mo ago
What RAM? OpenAI booked the silicon wafers, they can print anything they want on them. I wouldn't call them "far behind" on hardware when OpenAI are actively buying Cerebras chips.
24.
by mhjkl 7mo ago
Because as we all know the EU would never try using AI for mass surveillance /s
25.
by mhjkl 7mo ago
Might as well just encode the Unix timestamp at creation into proquints.
26.
by mhjkl 8mo ago
The reason is that the web empire is just better at operating systems than Microsoft. If they just had less bad development tools for native UI this would not be a problem. Look at what Google does with Android, or Apple.
27.
Open Source Historical LLM trained exclusively on 19th century text(github.com)
12 points by mhjkl 9mo ago|0 comments
28.
by mhjkl 9mo ago
There’s no guarantee the toolchains will support WASM “preview” forever and make the bloat optional, and even if they do you could still end up in an ecosystem where it would be unviable. At some point you’re probably better off just compil
29.
by mhjkl 10mo ago
Afaik most LLM datasets use FastText or something similar to detect the language of the data and if it's spam, and some additional small language models to detect if text is "educational" or desirable in some other way. Often