Edgeley°hard word, North Dakota°hard word, is a small rural°hard word town in LaMoure°hard word County°hard word, located°hard word in the southeastern°hard word part of the state. With a population hovering°hard word around 500 people, it's°hard word one of many prairie°hard word towns that exemplify°hard word the broader character°hard word of the upper°hard word Great Plains—quiet, sparsely°hard word populated°hard word, and closely tied°hard word to agriculture°hard word.
----
https://www.lesswrong.com/posts/bfHDoWLnBH9xR3YAK/ai-2027-is-a-bet-against-amdahl-s-law
Of course the post is right. The various°hard word FOOM°hard word claims°hard word are all bullshit°hard word. And Amdahl's°hard word Law is one of the reason why. Just because a few things will be a hundred°hard word times faster (or a million°hard word times faster) doesn't°hard word make the whole°hard word thing that much faster.
Also, AGI°hard word definitions°hard word vary°hard word so widely, from things that have already happened to things that are impossible, that a "prediction°hard word market" is nearly meaningless°hard word.
----
I have seen°hard word various°hard word commentary°hard word related°hard word to "Twilight°hard word of the Edgelords°hard word" ⚙️ https://www.astralcodexten.com/p/twilight-of-the-edgelords , a piece that I don't°hard word have access°hard word to.
And, the response°hard word I can piece together from the fragments°hard word I can see would fall under GUILD°hard word LAW. ⚙️ additional°hard word commentary°hard word at https://www.writingruxandrabio.com/p/the-edgelords-were-right-a-response and https://theahura.substack.com/p/contra-scott-and-rux-on-whos-to-blame
----
To make Gemma°hard word 3 even more accessible°hard word, we are announcing°hard word new versions°hard word optimized°hard word with Quantization-Aware°hard word Training (QAT°hard word) that dramatically°hard word reduces°hard word memory requirements°hard word while maintaining°hard word high quality. This enables°hard word you to run powerful models like Gemma°hard word 3 27B locally°hard word on consumer-grade°hard word GPUs°hard word like the NVIDIA°hard word RTX°hard word 3090.
It seems pretty obvious. A majority°hard word of the users°hard word of open-source°hard word models are using°hard word quantized°hard word models on personal°hard word hardware°hard word; might as°hard word well optimize°hard word that use-case°hard word. 💡 it is less clear that a majority°hard word of the CPU°hard word cycles°hard word are there; but a majority°hard word of the people certainly are.
My next round of updating°hard word the Greenland°hard word metrics°hard word will have to include°hard word the gemma°hard word3-12b-qat°hard word model. 💡 or, maybe the 27b. According to Hacker°hard word News, gemma°hard word3-27b-Q°hard word4 only uses ~22Gb°hard word (via°hard word Ollama°hard word) or ~15GB°hard word (MLX°hard word). On a 24GB°hard word machine, this clearly needs the non-Ollama°hard word approach°hard word.
And, also, GPT°hard word-4.1 . And probably Gemini°hard word-2.5 . 💡 the goal°hard word for these models should be to perform°hard word at 100% accuracy°hard word. ⚔️ well, actually, a few of the "correct" benchmark°hard word answers right now are incorrect°hard word.