anonymoushn

I couldn't reproduce this behavior with Sonnet 4, and Sonnet 3.7 has been deprecated since I messed with this stuff. You can try tokenizing the string "<hello> </hello>"

I think the correct tokenization of the string will not have any tokens that contain mixed punctuation and letters, but the result of this approach does contain such claimed tokens.

That's " 'd ".strip(), an english contraction suffix. it's 1 token, but using this echo approach you will be served the apostrophe and the subsequent letter for the first time in different steps.

You can't reliably obtain correct token boundaries with this method. For example, "'d" is 1 token, but the API will return "d" stuck to the next token. Weirdly this seems to be specific to the letter "d". Similar stuff happens around "<". About all caps words, some words are in the vocab in all caps, such as MERCHANTABILITY.

their old tokenizer performed some space collapsing that allowed them to use the same token id for a word with and without the leading space (in cases where the context usually implies a space and one is not present, a "no space" symbol is used).

Is this the wrong URL? this seems to be a blog post from October 2025 called "Introducing: Local Browser AI"

How do you guys decide which settings should be configurable via environment variables but not settings files and which settings should be configurable via settings files but not environment variables?

On of our product principles is to avoid changing settings on users' behalf

Ideally there wouldn't be silent changes that greatly reduce the utility of the user's session files until they set a newly introduced flag.

I happen to think this is just true in general, but another reason it might be true is that the experience the user has is identical to the experience they would have had if you first introduced the setting, defaulting it to the existing behavior, and then subsequently changed it on users' behalf.

Oh, can you post some benchmarks? I didn't know that parser throughput per core would change with the amount of data like that.

why

are those tools known for their fast json parsers?

ideally users could be banned for posting LLM outputs as if they were authored by humans https://www.pangram.com/history/49335ddf-118d-43e4-9340-a58a...

Hello, the part about canonical filtering in https://openreview.net/pdf?id=DFybOGeGDS doesn't seem to try to account for pretokenization. For example, if you receive " 天天中彩票APP" in o200k, it means there has to be a lowercase letter within the span of letters, and while tokens like (4 spaces) may be pairwise compatible with tokens like "123" according to the BPE merge rules, the pretokenizer would split the span of spaces to give (3 spaces), " ", "123" instead. Are you aware of any work that does actual canonical generation for models with this kind of pretokenization regex?

use claude code if you want to use opus

what does "logprobs look off" mean

Hello, a couple years ago I participated in a contest to count word frequencies and generate a sorted histogram. There's a cool post about it featuring a video discussing the tricks used by some participants. https://easyperf.net/blog/2022/05/28/Performance-analysis-an...

Some other participants said that they measured 0 difference in runtime between pshufb+eq and eqx3+orx2, but i think your problem has more classes of whitespace, and for the histogram problem, considerations about how to hash all the words in a chunk of the input dominate considerations about how to obtain the bitmasks of word-start or word-end positions.

requires fully deterministic inference, which turns out to be unusual, but for this sort of thing it's probably fine if you do really slow inference on cpu. cool idea.

please write your own posts from now on

i love stemming, i love searching for "anime" and getting "animal"

so sad to hear that about Streaming SIMD Extensions

This is true economically but in reality if you have much larger cost savings than that for sale then these companies mostly say "we would be happy to buy that for $0 while we pay you a million a year to move to the united states"

If they had, they would know that it involves many weeks of arguing with support, of course

The framing in this post is really weird. Automated evals can be much more informative than unit tests because the results can be much more fine grained. A/B testing in production is not suitable for determining whether all of one's internal experiments are successful or not.

I don't doubt that Raindrop's product is worthwhile to model vendors, but the post seems like its audience is C suite folks who have no clue how anything works. Do their most important customers even have any of these?

it is frustrating that the post opens by describing latency and then saying that it is called throughput.

fwiw i thought the previous async based on whole-program analysis and transformation to stackless coroutines was pretty sweet, and similar sorts of features ship in rust and C++ as well

Arrays of length 0 also exist in json?

Hello, it's probably better to add leading spaces before all of the words rather than none of them

I don't really know what you mean by "preserve semantic meaning at the word level." The significant misunderstanding about tokenization present elsewhere in the article is concerning, given that the proposed path forward is to do with replacing tokenization somehow.

hello. it would be great because it celebrates the total destruction of the search capabilities of the tool, just like the article does.

well, most of the users are aarch64

you can have some vector registers n_acc, ns, idx_acc, idxs, then you can do

  // (initialize ns and idxs by reading from the array
  //  and adding the apropriate constant to the old value of idxs.)
  n_acc = min(n_acc, ns);
  const is_new_min = eq(n_acc, ns);
  idx_acc = blend(idx_acc, idxs, is_new_min);

you can consult the list of toys here: https://www.intel.com/content/www/us/en/docs/intrinsics-guid...