jbellis

4 ms·

31.
by jbellis 3mo ago
That's a reasonable option, just be aware that you get about 1/3 as much memory bandwidth with the M5 Pro, or 2/3 with the M5 Max [now you're at $4100 for the lowest-end]. So both your prefill (flops-bound, M5 has a lot
32.
by jbellis 4mo ago
It's no mystery. https://en.wikipedia.org/wiki/Anders_Hejlsberg
33.
by jbellis 4mo ago
I feel bad that I wasted my time reading this. On the points in the article: 1. Yes, "gain" is a vanity metric but it's harmless, nobody is being "fooled" here. 2. This could be a problem in principle, sure, but unl
34.
by jbellis 4mo ago
man, Tom Lane has hated query hints for literally decades did he finally come around?
35.
by jbellis 4mo ago
it is hard to understand what the actually meaningful innovations are here / what TileRT is bringing to the table. - dflash: new-ish but February is ancient by the standards of the pace of AI innovation lately, I guess applying it to a
36.
by jbellis 4mo ago
Because you need kv proportional to context length during inference of a single token to avoid quadratic recomputation. So compressing the kv lets you handle longer contexts in the same amount of vram.
37.
by jbellis 4mo ago
> the association with flexibility or set shift did not remain significant after false discovery rate correction this is right there in the abstract, isn't that the entire game?
38.
by jbellis 5mo ago
Because it gets you a minimal amount of abuse prevention for free.
39.
by jbellis 5mo ago
Generically, I would say, just start building it and ask your favorite coding agent for advice when you get stuck. This is the first technology that can teach you how to use it! (But do ask a model with a recent knowledge cutoff, i.e. not g
40.
by jbellis 5mo ago
As someone who has been writing harnesses for a year: the people at opencode etc aren't stupid, when they decide to break the prefix cache [usually partially] it's always because they've tested it and it gives better results
41.
by jbellis 5mo ago
Developers: stop doing whiteboard interviews, they don't measure anything relevant to the real job Also devs: stop giving us real world problems to solve
42.
by jbellis 5mo ago
https://archive.is/kO6Ph
43.
by jbellis 5mo ago
Yes, but not diffusion based, it's still doing token-at-a-time speculation.
44.
by jbellis 5mo ago
BTW the paper says > Since only (Qdiff,Kdiff,Vdiff) are updated during training, the total number of trainable parameters is approximately 16% of the full model. But the code defines q_proj_diff, k_proj_diff, v_proj_diff, and o_proj_diff
45.
by jbellis 5mo ago
Really cool work! Does the training data budget scale with model size? How would you compare the Gemma 4 draft model which is also integrated with the base kv cache?
46.
by jbellis 5mo ago
Your calibration is wildly off. Asking people for a spot is totally normal at any gym with free weights.
47.
by jbellis 6mo ago
How should I update my simplistic understanding that decode is bw-bound with these results that show the B70 decoding faster than a 4090 (about 50% more bw)?
48.
by jbellis 6mo ago
Probably the best single resource is https://github.com/pmcfadin/awesome-accord
49.
by jbellis 6mo ago
> Batches all operations. Does large number of reads/edits simultaneously... I wasn't sure what this meant, so I looked at the source. It seems to be referring to tool APIs being designed around taking multiple targets as a lis
50.
by jbellis 6mo ago
Branimir is an engineer's engineer. Excellent choice.
51.
by jbellis 6mo ago
The newest structural diff tool is RefactoringMiner, there's a paper and a Github repo that works out of the box which is rare for this space. Excellent results but mainline is limited to Java IIRC with a couple ports for other languag
52.
by jbellis 6mo ago
what are you using for web search?
53.
by jbellis 6mo ago
Flash 2 isn't even at EOL until June but we started seeing ~90% error rates getting 429s over the weekend. (So we switched to GPT 5.4 nano.)
54.
by jbellis 6mo ago
Isn't the "KV Compression Strategies (FAIR)" chart showing that the fancy complex algorithm only barely beats simple topk? The commentary says that topk "degrades rapidly at low ratios" but the same can be seen for
55.
by jbellis 6mo ago
For coding, qwen 3.6 35b a3b solved 11/98 of the Power Ranking tasks (best-of-two), compared to 10/98 for the same size qwen 3.5. So it's at best very slightly improved and not at all in the class of qwen 3.5 27b dense (26 so
56.
by jbellis 6mo ago
Wonderful, another "my preexisting biases are all correct, because AI" piece.
57.
by jbellis 6mo ago
> the treatment for methanol poisoning is… ethanol! I looked this up, it is directionally correct but if you are in a hospital setting they have better options https://www.ncbi.nlm.nih.gov/books/NBK482121/
58.
by jbellis 6mo ago
I've had some close calls already and with AI making it cheap to tailor scams to individuals it's probably only a matter of time. For my parents in their 70s, even more so. No amount of reminding them to read URLs first is going t
59.
by jbellis 7mo ago
So MacOS supports only a handful of low dpi resolutions and high dpi must be an integer multiple of one of those?
60.
by jbellis 7mo ago
TIL. And good to see you on HN!

More ›