vitorsr

Maybe it's an unpopular opinion, but a concise report from one of today's frontier models is not something I would be offended by, even if nobody read it beforehand. I guess the fact that there is a nontrivial cost to it also eases the existential angst of neither being deterministic (Lintian) nor having been produced by a person.

Meant as a comparison given the discrepant pricing points:

I pay the Exa premium for semantic retrieval with far higher SNR, plus the academic index [1].

A field like objective can be up to 4096 characters [2] versus ori_question which is limited to 128 here.

[1] https://exa.ai/blog/publications-search

[2] https://exa.ai/docs/reference/search

> In September 2nd, something like that, the biggest conjecture in our field was proved. Did your life change? My life? No. No. No. I mean, the only people that can actually tell you that it does change something are the mathematician who worked on it. The human understanding of mathematics is the only thing that at the end benefits society.

The Strip Mining of Mathematical Proofs (with Hugo Duminil-Copin) - Numberphile Podcast https://youtu.be/omuJpoV-Ko4

GIS work has very complex inner mechanics that are hard to reason about at a high level without very explicit abstractions to ease the costs and cognitive burden of manual review. Sometimes earlier, longer-lived FOSS projects may have grown organically without the benefit of retrospective appreciation because that path has not been trodden (at least out in the open) at the time.

> u clown

I hate how the Gemini series only responds to outright insult. Same with the Claudes. What is the point of antagonizing the user when he or she offers input. Is it so hard to spend a few "reasoning" tokens to reconcile contradicting evidences. Seems almost like the wrong epistemic posture is altogether imprinted during training. Or undone during alignment, "hardening", whatever. What a time to live in, having to argue with weights.

From memory, the language server appeared to be a second-class citizen, debugging was kind of idiosyncratic, static analysis was poor... but I confess it has been a while since I made a serious effort to use it.

I was particularly frustrated by the promises of composability not translating into practice.

I was also stricken by the same npm/Rust-like pathological dependency trees where I had to pull 200-plus packages for some function.

All I ever wanted from Julia was great a development experience, but it appears to have never quite gotten there.

As a language it has gotten there a long while ago (1.6 LTS series comes to mind).

The landscape has changed substantially since then, and now "ergonomics" has unfortunately taken lower priority due to agent-driven development.

Nevertheless I think everyone should take a page from Go's book, really, language design now should include the default development workflow.

I believe the general phenomenon is metadiscourse and those in specific are signposting constructions.

Most used as a Wikipedia-on-the-fly. Replaces UpToDate and other POC tools NOT professionals operating them.

Nothing wrong with CMOS/APA/MLA dashes, I quite like them but usage in LLM prose has ruined it for me. AP dashes are weird, too much horizontal spacing, looks horrible and can really erode typographic color sometimes. "International" Oxford/Cambridge spaced dashes is nicest.

The funny thing is at both extremes, either none or indiscriminate use of dashes like in TFA, is a canary of "human writing".

People (including comments here) often attribute LLM use to prevalence in training data but it is probably far less common in the wild than we are led to believe, the more likely scenario is that whatever text-to-text normalization tools used ended up inserting whatever their idiosyncratic typographic styles were.

What caught my eye was:

            Model  Lines of Code  File Size  Gzip Size 
      GPT-5.6 Sol          1,264    35.5 KB    10.0 KB 
    GPT-5.6 Terra            827    20.0 KB     6.7 KB

I am not sure it is the management consulting partners or their associates that are doing the vibe coding.

The more likely scenario is that they "consume" the output of designated offshore teams.

Presented at ASAC 2023. See also:

https://dl.acm.org/doi/10.1145/3627106.3627138

There is also the possibility of using Toolbx (formerly [Fedora] Toolbox), distrobox or a container, and the underlying container's package manager. The issue then ends up being about how ergonomic it is to manage a separate guest system (and have to drop into it anytime we wish to use a binary that is unavailable in the host).

Thanks for all the hard work.

We are not many [1], but Homebrew has been a great way to quickly bootstrap an environment in immutable Linux distributions.

Note that certain operating systems such as Universal Blue's Bazzite (1.28%), Bluefin (0.49%) and Aurora (0.28%) default to bundling Homebrew [2].

[1] https://formulae.brew.sh/analytics/os-version/365d/

[2] https://github.com/ublue-os/brew

I am not sure it is recommending more than it is commenting on the current state of developing public-facing APIs in Python.

The downstream users that import the package either have to ignore checking its exported types altogether, manually stub it, or have a subpar development experience to varying degrees.

This is something I saw the other day with some package that provided comprehensive stubs for an untyped library. The .pyi file was littered with comments about quirks from the numerous type checkers (five now).

This is what I do for the former:

    floor( nextafter( 256, 255 ) * value )

I recently installed Fedora Kinoite [1] and I have been very pleasantly surprised by how stable and performant it is. I am afraid I cannot say the same for their new KDE Linux distribution [2] which in my opinion was a bit unpolished at the time. Both are immutable desktop distributions.

The good news, I guess, is Kinoite stands to benefit from KDE Linux development because it mostly depends on Flatpak to install programs which means all of the KDE ecosystem will eventually be available at Flathub [3] as first-class citizens with reasonable maturity.

[1] https://fedoraproject.org/atomic-desktops/kinoite/

[2] https://linux.kde.org/

[3] https://flathub.org/en/apps/collection/developer/KDE/1

Yes. The discount was most likely a "post-market trial" of how efficient the caching works for the new generation models.

Microsoft has already gone down this road some three years ago...

https://blogs.microsoft.com/blog/2023/02/07/reinventing-sear...

Tasks are taken from commit histories in public Git repositories which defeats the purpose.

Or become a major investor on a largely public funded project with commitments set to start at a delayed time in order to benefit from R&D before bearing financial burden. (See [1].)

[1] https://www.ft.com/content/8e75ed31-0c72-4160-b406-1ca6aa36a...

Unsure if this is what the announcement is referring to:

https://resources.wolframcloud.com/PacletRepository/resource...

Seems very promising but then you realize the LLM behind said agent was trained on public but otherwise copyright encumbered proprietary code available as improperly redistributed SDKs and DDKs, as well as source code leaks and friends.

In fact most Windows binaries have public debug symbols available which makes SRE not exactly a hurdle and an agent-driven SRE not exactly a tabula rasa reimplementation.

Please elaborate, I am curious to why would you think WebGPU would meaningfully beat their Metal/DirectX renderers.

Nice although I think the ASCII comma feels wrong as part of a filename even if for purely aesthetic reasons.

If we want to stay within (lowercase) alphabetic Latin characters I think prefixing with the least common letters or bigrams that start a word (x, q, y, z, j) is best.

`y' for instance only autocompletes to `yes' and `ypdomainname' on my path.

Choosing a unique bigram is actually quite easy and a fun exercise.

And we can always use uppercase Latin letters since commands very rarely use never mind start with those.

Zotero is built on top of Firefox ESR.

Thanks for all impressive work on AdGuard.

Any particular reason to adopt Rust for this project instead of Go as many of your other products?

Because I think since you have quite extensive Go codebase I would imagine you had to rewrite possibly a significant amount of code.