jbellis

It's roughly equal on price and intelligence as GLM 5.2 while being ~8x faster.

I guess you didn't read ellie's reply directly underneath that?

It tops benchmarks because it uses them in its training data. https://x.com/eliebakouch/status/2077425801633427919

This just isn't true. On balance, data centers are turning out to be more like the "anchor tenant" of the power grid, financing improvements for everyone.

Overview article with links to actual studies: https://cityjournal.substack.com/p/data-centers-arent-raisin...

voyage 4 nano is sota at the next size up

and if you really want the best teacher models it's probably the voyage commercial APIs

FWIW -- Granite r2 small is a 30M model, still small enough to run on CPU, and a good baseline for fine tunes.

Yes, disappointing given that Chronicle actually does have legit expertise here. (Chronicle Map is not very well-known even in the Java space but it's by far the best larger-than-memory Map available.)

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 less) and decode (bw-bound) will be slower.

It's no mystery. https://en.wikipedia.org/wiki/Anders_Hejlsberg

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 unless you're actually vetting bug reports you're just spreading FUD.

3. Again, do you have any reason to believe that the thousands of devs using rtk are silently tanking their performance without noticing? here's a thought: instead of reporting that SOMEONE SHOULD MEASURE THIS, you could, you know, measure it yourself.

4. Good lord, what is this doing in a purportedly technical article?

5. Yes, this is inherent in the problem domain, again, nobody is being "fooled".

Yes, I'm grumpy; reading this article was a waste of time.

Bias: had my first RTK pr accepted today, so I guess I probably know more about it than this guy who got offended by "gain" and spit out the first thoughts that came to mind.

man, Tom Lane has hated query hints for literally decades

did he finally come around?

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 1T model is new-ish in the sense that the dflash researchers don't have the hw budget to prove that out - persistent engine kernel: this is like CUDA 101 - warp specialization: I think this just means "keep different gpu resources all busy w/ pipelining" which is CUDA 201, some of it is even baked into pytorch now - MXFP4 QAT: not new - TileRT: hard to tell what this actually does, there's a PyPi wheel with support for DS 3.2 and GLM 5 but binary only

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.

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?

Because it gets you a minimal amount of abuse prevention for free.

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 gemini.)

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 overall.

If you think that dsv4 behaves differently enough from the aggregate of other models, submit a PR with a patch to special case that to your harness of choice with evidence. Just blindly assuming "append only all the time because cache" is a waste of everyone's time.

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

https://archive.is/kO6Ph

Yes, but not diffusion based, it's still doing token-at-a-time speculation.

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, and it only matches 16% when you include the O term.

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?

Your calibration is wildly off. Asking people for a spot is totally normal at any gym with free weights.

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)?

Probably the best single resource is https://github.com/pmcfadin/awesome-accord

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 list parameter, instead of hoping the model makes appropriately parallel tool calls. (This matches my experience btw, models are reluctant to make a large number of parallel calls simultaneously, and this seems more pronounced with weaker models.)

Branimir is an engineer's engineer. Excellent choice.

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 languages.

what are you using for web search?

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.)