mahesh

exploring the matmultiverse

[mahesh @_mnat_]

@xeophon made a "web-search <query>" script that i have my claude code prompted to call in place of its web search which runs o3 with its web tool enabled.. results are night and day.

[mahesh @_mnat_]

@_xjdr @xeophon What does your deslopification workflow look like in cc? A stop hook that kicks off another headless cc to review the most recent changes? Running a slash command every so often to check for certain conventions? Manual review?

[mahesh @_mnat_]

@EcstaticIndian https://t.co/NRLEkTLrBn

[mahesh @_mnat_]

@wangyenchi https://t.co/gUtK50l8Wm

[mahesh @_mnat_]

@sidnbaskaran lebanese blonde better

[mahesh @_mnat_]

@kuberdenis > free s3 inside
> free analytics inside
> free email routing inside
> free waiting room inside
> free image optimization inside
> free ddos protection inside
> free vercel inside

common cloudflare w

[mahesh @_mnat_]

@aidan__pratt

[mahesh @_mnat_]

@dejavucoder for #2: @greptile

[mahesh @_mnat_]

@ConsumerRick zetamac supremacy

[mahesh @_mnat_]

@matthisgarnr maximum viable product

[mahesh @_mnat_]

@dreamingtulpa comfy ui node?

[mahesh @_mnat_]

@teortaxesTex a wise man can learn more from a foolish question than a fool can learn from a wise answer

[mahesh @_mnat_]

Since imo most llm problems are some careful data curation/collection/synthetic generation and gradient descent from a solution, “dataset centric thinking” opens a lot of doors

[mahesh @_mnat_]

surprised people don’t think about data moats more (what models can we build with this data, what data do we need to collect to solve this problem, etc.) it’s always about retrofitting existing general solutions into specific problems

[mahesh @_mnat_]

@darpinian Even just running benchmarks across diff quantization levels would be nice. I should be able to input my GPU and get a spat out chart of inference speed vs benchmark perf, with a highlighted answer like “in order to maintain >100 tok/s and x MMLU, you need to use 4 bit llama 70B”

[mahesh @_mnat_]

@darpinian I’m aware of some huggingface spaces that can at least do “how much vram do I need” for diff models and quantization levels, but a single source of truth to point any newbie in the right direction would be awesome

[mahesh @_mnat_]

@sandkoan Time to cop and redirect to https://t.co/dHQ9lAmg2m until superior website is created

[mahesh @_mnat_]

@darpinian Scraping twitter for a hot papers and sentiment/analysis of opinions on the papers would be cool, though scraping twitter is now slightly more of a pain in the ass now that API is gone (iirc)