verdverm

27 ms·

931.
by verdverm 4mo ago
Bonsai is qwen3.6 based, not 3.5 Likely apples / oranges
932.
by verdverm 4mo ago
I haven't dug into QAT deeply, better recovery is my understanding as well, and also that it is out of reach for most people because you have to train a model to back prop errors based on estimated error under quant. Hopefully more of
933.
by verdverm 4mo ago
yea, it's an experiment in merging multiple fine-tuned models
934.
by verdverm 4mo ago
Did they look at how other smart speakers are used and have fared? It seems unlikely if this is the physical product they intend to release. Hope they have a Spotify integration so we can listen to that sweet Ai generated music /s
935.
by verdverm 4mo ago
Nice history video of 2b2t server: https://www.youtube.com/watch?v=-3eTxoWSo7c
936.
by verdverm 4mo ago
check out geyron-9b, I've only used it a bit, but seems better than orinth on vibe evals huggingface.co/Tivaphraen/Geryon-9B-v1
937.
by verdverm 4mo ago
bigger quant'd harder is not always better than a model of more modest size and quant
938.
by verdverm 4mo ago
4bits is a cutoff point for many model families, but also depends on what parts you quant to 4bits vs alternatives (weights, weight+activation, kv cache). Also depends on model size and task, lots of nuance in quanting I've come to lea
939.
by verdverm 4mo ago
Preliminary analysis via lm-evaluation-harness + vllm model | disk | wikitext | gsm8k (match/error) baseline | 55G | 8.00 | 0.50/0.09 nvfp4-gptq | 27G | 8.25 | 0.47/0.9 nvfp4a16-gpt
940.
by verdverm 4mo ago
Orinth was not impressive in my vibes testing, I just completed my first grid analysis with real evals on qwen 27b. I can now scale that grid analysis and intend to include the qwen 9b ftunes I've seen going around. They were actually
941.
by verdverm 4mo ago
quanting kv cache hurts attention / recall, and long-form tasks by proxy. Model families and sizes have different tolerances to quant ting different parts of the model, same for intended tasks.
942.
by verdverm 4mo ago
That's what happens when you quant too hard. I'm working on quant strats and evals for the same underlying qwen 27b models. When I saw 27b on a phone, I thought not fitting, big phone, or aggressive quant. NVFP4 still takes 27G be
943.
by verdverm 4mo ago
The core idea is hopefully enough when I get a chance to look. What I thought is that I have some related signals low dimension signals, and that graphing them together like this in a higher dimension is pretty interesting as the opposite o
944.
by verdverm 4mo ago
After DOGE, pardon if we remain skeptical
945.
by verdverm 4mo ago
Do you have any docs / readme on the math / methods behind this? I saw a readme_too.docx, but it does not render on GH nor do I have tools that read those files.
946.
by verdverm 4mo ago
> Other computer technologies have seen similarly massive growth without triggering such a large spike in electricity or a shortage of computer components I don't think the author understands the difference in scale from the compari
947.
by verdverm 4mo ago
We don't need general intelligence to create/destroy many things with Ai. Its frustrating that so much of the conversation centers around the variety that concept fissioned language from humans.
948.
by verdverm 4mo ago
They are many already planned/under construction/built/expanding, there are also wildfires and hurricanes to consider. Its easy to offer hypothetical solutions for a far when details are amis. It's another thing to deal
949.
by verdverm 4mo ago
> "totalitarian" for a democratically elected representative government to ban something? Steven Miller has figured it out
950.
by verdverm 4mo ago
Sea water would have to be piped to places like Dallas, whereas much of Finland's people/data centers are closer to the shoreline. A solution that works in certain places does not generally work everywhere.
951.
by verdverm 4mo ago
AlphaFold is a pretty big deal, especially the millions of protein folds they published openly and scientists are just digging into.
952.
by verdverm 4mo ago
Those sticks look less effective today. The hegemony is in recession as the rest of the world smiles, pays omage to dear leader as needed, and works to decouple in the background.
953.
by verdverm 4mo ago
They are switching some lines to building military vehicles aiui
954.
by verdverm 4mo ago
My feelings ever time this comes up on HN lately (-_- )╯ vim vs emacs ( ^_^)σ AP vs AT The truth is that both AP and AT made steps in the right direction but also contain fundamental flaws. The next protocol will hopefully draw on both bu
955.
by verdverm 4mo ago
HN got a new guideline because of it: https://news.ycombinator.com/item?id=47340079 While it was worse earlier this year, I still see a lot of ns;nt in /new, people looking to make money or get users for skinned or vib
956.
by verdverm 4mo ago
The Sleev (the project has been renamed to make a startup) creator was shilling their project in the OpenCode Discord. That person is very convinced they have something that no one has ever built before. They focused on token reduction with
957.
by verdverm 4mo ago
I'm not aware of any low/no code platform which can read an error, delve the code, and create a fix. LLMs are more than just tools for generating code. Case in point, I don't have to wait for open source maintainers to get ba
958.
by verdverm 4mo ago
Steps 7 & 8 seem backwards. Who merges to main before running the lint / test (CI)? Every commit/PR should be running through automated processes
959.
by verdverm 4mo ago
Have we considered the reason might be self (de)selection? Perhaps people who see the world certain ways do (not) desire a life in academics.
960.
by verdverm 4mo ago
I see significant decline in product quality this year among those companies most pilled (Ant, Goog, Msft, etc...). I had been commenting that Atlassian turned a page and Jira was running quite well, but this year... their internal push on

More ›