singhrac

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.

I think you have 2 issues with content based systems: you have ambiguity (in archiving and retrieval) and you have to use mental effort to resolve that ambiguity.

The Dewey decimal system has less ambiguity in both, and an alphabetical system would be unambiguous for archiving (if not retrieval).

I prefer to organize my notes functionally (eg internal emails, blog posts, links to read, reading notes) and then rely on search for retrieval. It’s not perfect but it lowers the friction, which I think is very important.

Sure, but in this case they are both implementations of a spec defined by PEPs, so a bit more like gcc vs clang (less tightly bound than those, of course, in design decisions). Neither company is trying to invent a new language here.

I think they've mentioned earlier that it's really just because PyCon is this week (so a good time to announce new Python tooling).

Here's a very small piece of I code I generated quickly (i.e. <5 min) for a small task (I generated some data and wanted to check the best way to compress it):

https://gist.github.com/rachtsingh/e3d2e2b495d631b736d24b56e...

Is it correct? Sort of; I don't trust the duration benchmark because benchmarking is hard, but the size should be approximately right. It gave me a pretty clear answer to the question I had and did it quickly. I could have done it myself but it would have taken me longer to type it out.

I don't use it in large codebases (all agentic tools for me choke quickly), but part of our skillset is taking large problems and breaking them into smaller testable ones, and I give that to the agents. It's not frequent (~1/wk).

There’s nothing quite like writing the final squashed commit message and saying “results should be bit-level identical, +25% speedup on average” or the equivalent. Even better if it’s a hot path that matters to the bottom line of the company.

I try to keep a few of these tasks lying around marked “unblocked” in case I need to unwind. It keeps me from doing it right after noticing, which is an extra bonus.

Well, I mean why use a suboptimal tool? I don’t have any ideology to follow, my codebase is just mostly Python and heavily typed. Pyright doesn’t support all the features so I stick with VSCode/Pylance.

At least the alpha milestone is May 12, but I don't think they've publically committed to anything.

In particular for me it would be great to have a better typechecker than Pyright available for Zed (basically why the editor is a nonstarter right now).

Can you choose a model via the Gemini app? I can on the webapp (finally), but on the mobile app it won’t let me choose.

Using Gemini via Google Workspace.

Yes, this is pretty much what Hetzner must have built with their object storage - and they get to 5 EUR/month, so really close to Backblaze pricing.

I'm having a bit of a Gell-Mann moment here - I'm hearing the exact opposite from every residential and C&I installer I talk to, that solar installations are booming right now.

There are very few states in the US where net metering doesn't exist, with California having the NBT (and Arizona and Utah something similar iirc), and Alabama, South Dakota, Tennessee, Idaho, and Texas all have utility-set forms of net metering. Just because California has phased it out and it's a sunny state doesn't mean you can't install a solar roof in Massachusetts and benefit.

The cost of panels is dirt cheap (even with tariffs). You don't need energy storage if you pay $100/month in energy costs, just offset that. Home Depot will install it if you want!

Eh, I write a lot of Python and while it's unnecessary, it can make code easier to read if not overused.

  if (config_fn := base_dir / "config").exists():
      ...

is just a tiny bit shorter than this:

  config_fn = base_dir / "config"
  if config_fn.exists():
      ...

but if you do this enough times (if you are handling a lot of possible errors / missing data, etc.) it can definitely shorten the code.

I want to argue that graphics cards are really 3 markets: integrated, gaming (dedicated), and compute. Not only do these have different hardware (fixed function, ray tracing cores, etc.) but also different programming and (importantly) distribution models. NVIDIA went from 2 to 3. Intel went from 1 to 2, and bought 3 (trying to merge). AMD started with 2 and went to 1 (around Llano) and attempted the same thing as NVIDIA via GCN (please correct me if I'm wrong).

My understanding is that the reason is that the real market for 3 (GPUs for compute) didn't show up until very late, so AMD's GCN bet didn't pay off. Even in 2021, NVIDIA's revenue from gaming was above data center revenue (a segment they basically had no competition in, and 100% of their revenue was from CUDA). AMD meanwhile won the battle for Playstation and Xbox consoles, and was executing a turnaround in data centers with EPYC and CPUs (with Zen). So my guess as to why they might have underinvested is basically: for much of the 2010s they were just trying to survive, so they focused on battles they could win that would bring them revenue.

This high level prioritization would explain a lot of "misexecution", e.g. if they underhired for ROCm, or prioritized APU SDK experience over data center, their testing philosophy ("does this game work ok? great").

That's pretty surprising and breaks my mental model of how these chips perform. Do you think that's because they don't have the raw FLOPs, something inefficient in the Apple/Metal rendering pipeline, something about Civ VI and how it was converted (i.e. x86 or DirectX emulation), or something else entirely?

I've had no trouble with a (base) M4 mini for regular dev work, though I compile remotely and haven't played any games on it.

It is really nice, I'm very sad that Apple's RDP implementations seem fragmented and difficult to use. I've tried a few times with my Mac Mini, and it just doesn't seem to work that well.

I have spent a lot of hours looking at `watch "xsv select ... | xsv table"`.

Do you have a roadmap? I'm interested (love Linear, hate Quickbooks/Xero, looking for something better, finances are not that complicated) but I want to know what is planned and what is supported now.

I use fd all the time. I admit it, I couldn't get the muscle memory right for find. One day I'll be on a bare Linux installation where I can't cargo install whatever I want and I'll have to ask ChatGPT to generate my find commands.

I like `fd -o {user}` and `fd --newer {time}` to find recently changed files and owned by others.

Sigh, there is a lot of misinformation about this and higher education tuition.

* It’s not a new policy, when I went to school the same threshold was ~$150k

* It’s not a step function, it’s a sliding scale (with <0 tuition, i.e. cost-of-living assistance below that number, and some (but not full) tuition above $200k)

* It’s not just for athletes or minorities - I was neither (by Harvard standards, Indians were fairly common and not considered a minority in admissions processes). Lots of smart kids whose parents did not make a lot of money fell into this bucket.

* It’s not a small subset of the student body (as tzs said); despite what you might perceive, a large fraction of students came from modest backgrounds. It was balanced by a sizable fraction that came from incredibly well-off backgrounds. But the total number that had donated a building to get in were pretty small (there are only so many buildings).

Higher education has many, many flaws, but please examine the facts before you respond with knee-jerk cynicism.

I mean, it is covered by mainstream media. There’s like 4 articles on it in the last year in the NYT. Here’s one: https://www.nytimes.com/2024/08/29/opinion/el-salvador-safet...

Ghost as an alternative? They’ll let you sign up paying subscribers as well.

There's some rumors of an Arc Pro, which would be a B580 in clamshell configuration with 24 GB of VRAM (which iiuc would be the same memory bandwidth unfortunately). Unless the price is absurd it would be the cheapest dollar/VRAM card at 24 GB.

This works on my personal Google account, but not on my workspace one. So I guess there's no access to 2.0 Pro then? I'm ok trying out Flash for now and see if it fixes the mistakes I ran into yesterday.

Edit: it does not. It continues to miss the fact that I'm (incorrectly) passing in a scaled query tensor to scaled_dot_product_attention. o3-mini-high gets this right.

I just tried it but nothing happens when I click on that. You're talking about the thing on the upper left next to the open/close menu button?

What is the model I get at gemini.google.com (i.e. through my Workspace subscription)? It says "Gemini Advanced" but there are no other details. No model selection option.

I find the lack of clarity very frustrating. If I want to try Google's "best" model, should I be purchasing something? AI Studio seems focused around building an LLM wrapper app, but I just want something to answer my questions.

Edit: what I've learned through Googling: (1) if you search "is gemini advanced included with workspace" you get an AI overview answer that seems to be incorrect, since they now include Gemini Advanced (?) with every workspace subscription.(2) a page exists telling you to buy the add-on (Gemini for Google Workspace), but clicking on it says this is no longer available because of the above. (3) gemini.google.com says "Gemini Advanced" (no idea which model) at the top, but gemini.google.com/advanced redirects me to what I have deduced is the consumer site (?) which tells me that Gemini Advanced is another $20/month

The problem, Google PMs if you're reading this, is that the gemini.google.com page does not have ANY information about what is going on. What model is this? What are the limits? Do I get access to "Deep Research"? Does this subscription give me something in aistudio? What about code artifacts? The settings option tells me I can change to dark mode (thanks!).

Edit 2: I decided to use aistudio.google.com since it has a dropdown for me on my workspace plan.

You can decorate a function with @torch.compiler.disable() to disable a specific function (and anything further down the call stack).

I was a longtime user of Sublime Text and switched to VSCode because of Remote Development and Pylance. Everything else is pretty much fine and Sublime has always had better performance, though these days having support for AI integration (a la Copilot) is also necessary for me.

I tried Zed recently because it has remote support and AI integration, but the Python integration is limited to Pyright so I gave up temporarily (I guess I can recreate Pylance using this doc: https://github.com/microsoft/pylance-release/blob/main/USING...).

This is a good place to start for sure. There is a DOE equivalent here: https://www.eia.gov/outlooks/aeo/pdf/electricity_generation.... .

One simple issue with LCOE estimates is that their revenues are not all equal. For example, rooftop residential solar seems very expensive (no economies of scale) but since residential electricity rates are high and you don't have to pay (as much) for transmission, they can be very financially viable, at least recently. Solar PV + storage looks more expensive, but that's because it shifts the supply to more valuable times (i.e. cuts the duck curve). Gas is more valuable than it seems because it is reliable on rainy days.

In my estimation the fact that wind and solar are often anticorrelated makes onshore wind basically financially viable without subsidies in some places (I think $27/MWh is basically viable).