singhrac

This article was AI generated and a waste of time. So many obvious LLM patterns that I stopped reading 10% of the way down the page.

I wanted to try NVIDIA’s aistore for our datasets, but I couldn’t figure out how to get a small version up and running so I gave up (a few years ago, today I’d get an LLM to show me how k8s works).

This is great! We need more alternatives to the full calibre stack for working with eBooks. Are you planning to keep it epub focused or would you consider adding support for something like azw3? Also is there a conformance suite for EPUB?

I think people underestimate how valuable these reports are, so I’m very glad that detailed investigation is done here. Every major grid operator around the world is going to study this and make improvements to make sure this doesn’t happen on their grid.

In a lot of ways it’s like investigations into airplane crashes.

It’s prompt processing so prefill - that’s compute bound not memory.

I want to like RustFS, but it feels like there's so much marketing attached to the software it turns me off a little. Even a little rocket emoji and benchmark in the Github about page. Sometimes less is more. Look at the ty Github home page - 1 benchmark on the main page, the description is just "An extremely fast Python type checker and language server, written in Rust.".

I use VoiceInk (needed some patches to get it to compile but Claude figured it out) and the Parakeet V3 model. It’s really good!

To be frank this is a service to you. No company you want to work at has a recruiter that doesn't understand the difference (a fully AI recruiter would be better than this experience).

Does anyone use this? I was setting it up a few months ago but it felt very complicated compared to MinIO (or alternatives). Is there a sort of minikube-like tool I could use here?

They're a bit less bad than they used to be. I'm not exactly happy about what this means to incentives (and rewards) for doing research and writing good content, but sometimes I ask a dumb question out of curiosity and Google overview will give it to me (e.g. "what's in flower food?"). I don't need GPT 5.1 Thinking for that.

One summer in middle school our family computer failed. We bought a new motherboard from Microcenter but it didn’t come with a Windows license, so I proposed we just try Ubuntu for a while.

My mom had no trouble adjusting to it. It was all just computer to her in some ways.

Here you go: ∈ (so f ∈ (...))

Let it go. He made a frustrated remark in a support thread 14 years ago where the OP escalated into calling him deliberately rude. Even if he hadn't changed at all over the years he has been contributing to open source with a product used by nontechnical people for 2 decades and deserves some grace for that.

Who cares if they feel like they are the more valuable person in the relationship? Do you decide your framework based on mental games other people might play? Decide if extending an invite that is declined will cost you something (food, space, etc.) and whether you want the person there.

The 288 core SKU (I believe 6900E) isn't very widely available, I think only to big clouds?

Is it stable? I've been using restic for a while, and I'm interested in rustic, but I have no idea how stable it is overall. Obviously it's still in beta so I won't use it in prod but curious what others experiences have been like.

I think we might just disagree about how much of the GPU spend is on small vs large model (inference or training). I think it’s something like 99.9% of spending interest is on models that don’t fit into 128 GB (remember KV cache matters too). Happy to be proven wrong!

Surprisingly overlapping set with games I’ve played and enjoyed a lot! Dead Cells was another that has a lot in common.

This is almost true but not quite - I don't think much of the (dollar) spend on enterprise GPUs (H100, B200, etc.) would transfer if there was a 128 GB consumer card. The problem is both memory bandwidth (HBM) and networking (NVLink), which NVIDIA definitely uses to segment consumer vs enterprise hardware.

I think your argument is still true overall, though, since there are a lot of "gpu poors" (i.e. grad students) who write/invent in the CUDA ecosystem, and they often work in single card settings.

Fwiw Intel did try this with Arctic Sound / Ponte Vecchio, but it was late out the door and did not really perform (see https://chipsandcheese.com/p/intels-ponte-vecchio-chiplets-g...). It seems like they took on a lot of technical risk; hopefully some of that transfers over to a future project though Falcon Shores was cancelled. They really should should have released some of those chips even at a loss, but I don't know the cost of a tape out.

Small startups will always excel in talent density because of the increase risk (and reward) and alignment along equity, but all of the large companies have enough technically brilliant people. Many in their 30s or 40s want to do excellent work and also have stability to provide for their families. Whether management gets out of their way is another question.

A few projects just in ML: DeepMind has plenty enough for Google alone but Jax and the TPU project are technically ambitious and very strategically important for Google; they hired Adam Paszke away from Meta. Besides Gemini they have Gemma and Veo and they have a reputation in the industry of having extremely high MFU averages. From Facebook there's PyTorch but that's a whole cluster of projects (compiler, abstractions for many accelerators, torchao, torchtitan). They're also famous for DINO and SAM models, as well as many others by FAIR (e.g. Mask R-CNN).

I've always thought that the reason they don't care about gaming is because gaming has always been a race-to-the-bottom in margins business. Gamers care a lot about hardware in ways that make selling gaming PCs always very low margin (compared to enterprise or creative) (very few other customers know about RAM or SSD prices, famously Apple's bread-and-butter margin increaser).

I switched from a 2019 MBP to a new M4 Pro a few weeks ago and I didn’t even know Rosetta wasn’t installed (I assumed on and installed by default) until I had to run a Go binary that hadn’t been updated since 2020.

I use a lot of nonstandard software (not just a browser), not a single piece needed Rosetta.

I agree recent AMD chips are power efficient like the M series (though I don’t have one to compare with) but I thought everyone agreed the comparable chips in 2020 weren’t?

The employees will have to wait 180 days (at least that's the standard) before selling any shares, so they usually feel the effects of a "bounce".

I think we don’t talk enough about tab completion model quality. Recently Copilot’s model got a lot better (probably trying to catch up to Cursor) but I feel like there’s still so much room here (and I assume Zeds is worse from your description).

Smart context / big context is a really interesting question, I’m kind of surprised Google isn’t building here given how much effort they’ve put into big context (they have Jules and Gemini CLI but no tab completion UX).

On further thought I think one of the big 3 (OpenAI, Google, Anthropic) should partner (ideally not buy) with Zed to get a foothold.

I use the Night Owl theme in all my editors (and terminal), including Zed, and the consistency is nice. It looks good, but if there’s a Zed-specific issue I haven’t used Zed in a while.

I just symlink to AGENTS.md, the instructions are all the same (and gitignore the model-specific version).

I've been doing it to Azure (private package repo) since their instructions were ok. I would like some of these to publish instructions for uv as well (instead of just pip/twine).

Production inference is not deterministic because of sharding (i.e. parameter weights on several GPUs on the same machine or MoE), timing-based kernel choices (e.g. torch.backends.cudnn.benchmark), or batched routing in MoEs. Probably best to host a small model yourself.

I spent too long looking into this and settled on restic. I'm satisfied with the performance for our large repo and datasets, though we'll probably supplement it with filesystem-based backups at some point.

Borg has the issue that it is in limbo, i.e. all the new features (including object storage support) are in Borg2, but there's no clear date when that will be stable. I also did not like that it was written in Python, because backups are not always IO blocked (we have some very large directories, etc.).

I really liked borgmatic on Borg, but we found resticprofile which is pretty much the same thing (it is underdiscussed). After some testing I think it is important to set GOGC and read-concurrency parameters, as a tip. All the GUIs are very ugly, but we're fine a CLI.

If rustic matures enough and is worth a switch we might consider it.

Sure, I agree. I’m just saying that most of their defacto disagreements are because of ambiguity in the specs, not because one of them is (intentionally) choosing to fork.

Fwiw I think performance and features are kind of intertwined here, since Pyright’s extra speed makes it possible to infer more things that are not easy for mypy, especially in an LSP implementation.

Pyright is very impressive in that second link (conformance check), very much a product of Eric Traut & others dedicating a lot of energy to this single problem.