ZeroCool2u

This looks almost identical to Dash. Any meaningful difference here? https://dash.plotly.com/

Interestingly, only 1 test from the PDS-H1 benchmark at scale factor 80 experiences a performance regression. Looks like it's test 6 and comes in at 0.8 of the CPU execution performance the standard polars engine.

Seems like there's a more concrete user guide here: https://docs.pola.rs/user-guide/gpu-support/#whats-supported...

It seems like they've done a great job of supporting most ops on the GPU already. The main one that is a really bummer is the categorical datatypes. Those are so common in ML workflows and using the native categorical datatype saves a lot of boiler plate as well as prevents you from messing something up during training!

The only big one I know of is Cloud Run on GCP.

https://cloud.google.com/run/docs/configuring/services/gpu

Damn, GeoParquet and R-Tree for spatial indexes is huge!!! ESRI better watch their back!

Ray seems to be a popular choice now if you need multi node. But Polars is really nice if you're okay with just vertically scaling on one machine.

XBox and Epic absolutely do fall in that category. People, myself included, detest using either of those stores, because of their poor implementations. They're slow and unwieldy. The Xbox store on PC is actually so bad. I bought Forza as a chill on the couch and relax kind of game and it is one of the most regrettable purchasing experiences I've ever had. The Epic store is just outright unpleasant software.

Mojo is interesting, because I get to keep all my existing Python code and libraries for free. Then when I need to speed things up I can use Mojo syntax.

Got it. eBPF module run as part of the kernel, but they're still user space programs.

I would would consider using a larger model for demonstrating inference performance as I have 7B models deployed on CPU at work, but GPU is still important training BERT size models.

How do you do that exactly? Are you using eBPF or something else?

Also, for my ML workloads the most common bottleneck is GPU VRAM <-> RAM copies. Doesn't this dramatically increase latency? Or is it more like it increases latency on first data transfer, but as long as you dump everything into VRAM all at once at the beginning you're fine? I'd expect this wouldn't play super well with stuff like PyTorch data loaders, but would be curious to hear how you've faired when testing.

There's an entire app called Google TV already. https://tv.google

" Android TV has expanded to 220 million devices worldwide and we are continuing to bring Google Cast to other TV devices, like LG TVs."

This line in particular puts a bad test in my mouth, because my $2k LG G2 OLED has the worst support for casting I've ever experienced. In fact the software in general is so bad I was excited to pre-order the new Google TV Streamer this morning, so I don't have to deal with it again.

I'm forced to operate in AWS GovCloud for my work and it's the same thing, but even worse. Old instance types and there's barely any of them in there. Mind you, there's two GovCloud regions and the East one is even worse! There's basically nothing outside general compute instance types, so you're really stuck with just the West region. P4's are officially available in only 1 AZ, come in exactly 1 size, and are one of the most expensive instances in the region. P3's (initially released in 2018!) are so hard to come by it's infuriating. Meanwhile we have a horde of AWS reps and they all claim there's availability.

Really feels like if you need accelerated compute GCP is the better option these days. At least there you can rewrite in Jax if it comes down to it and opt for TPU's.

Seems like this is a sales funnel for Azures OpenAI/LLM gateway with GitHub as a proxy(?). It's a bit unclear. Regardless, I'd be pretty wary of adding either Azure or GitHub as a core dependency to any of my apps at this point with how poor uptime seems to be at both lately.

Also, the pricing is pretty shady. It seems like your GitHub PAT gives you free limited access, but if you ever want to move to a paid model, you have to transition to Azure. The shady part is that it's pretty hidden. You have to go to a specific model in the Marketplace, (for example: https://github.com/marketplace/models/azureml-mistral/Mistra...), then go to the bottom of the "Chapters" to the "Going beyond rate limits" section. There it just directs you straight to the Azure portal.

This is one of the first times I've heard of people using Daft in the wild. Would you be able to elaborate on where Daft came in handy?

Edit: Nvm, I kept reading! Thanks for the interesting post!

Is there a rough equivalent to Arduino in the RISC-V ecosystem?

Pay walled, so here's the article:

In a major reversal, Google is ending a plan to eliminate cookies in its Chrome browser after four years of efforts, delays and disagreements with the advertising industry.

The decision to keep the pervasive tracking technology known as "cookies" in Chrome comes after a series of setbacks, as both digital-advertising companies and regulators objected to the plan and to Google's proposed replacement technologies.

Chrome users can already choose to block cookies in the browser's settings. Now, instead of eliminating them, Google will present users with a prompt to decide whether to turn cookies on or off, said the U.K. privacy regulator, which has been overseeing Google's plan to block cookies.

"We recognize this transition requires significant work by many participants and will have an impact on publishers, advertisers, and everyone involved in online advertising," Anthony Chavez, vice president of Google's Privacy Sandbox, the company's initiative to replace cookies, wrote in a blog post Monday. "In light of this, we are proposing an updated approach that elevates user choice...We're discussing this new path with regulators, and will engage with the industry as we roll this out."

Google first announced the plan to kill cookies in 2020, saying it would do so within two years to help protect users' privacy when surfing the linternet. Advertisers objected, saying that Google's plan to replace cookies would force them to shift spending to the search giant's digital-ad products.

In 2021, U.K. regulators opened an investigation into whether the plan would hurt competition in digital advertising. Google pledged to collaborate with the regulator and committed to give the agency at least 60 days notice before removing cookies to review any plan, and potentially impose changes to it.

As that investigation dragged on, Google's schedule to kill cookies by 2022 slipped.

In April, The Wall Street Journal reported that the British government's Information Commissioner's Office would issue a report criticizing Google's proposed replacement technologies as deeply flawed. A few days later Google said it would delay cookies' demise beyond last announced target date of the end of this year.

I mean I've personally given our Nvidia rep some light hearted shit for it. Told him I'd appreciate if he passed the feedback up the chain. Can't hurt to provide feedback!

Yeah, this seems like a weird way to interpret what I said. I just meant we got in front of the right people to get permission, which we were already waiting to do for quite a while before the pandemic.

There was a time at the beginning of the pandemic where my team was asked to build a full text search engine on top of a bunch of SharePoint sites in under 2 weeks and with frustratingly severe infrastructure constraints, (No cloud services, single box on prem for processing, among other things), and we did and it served its purpose for a few years. Absolutely no one should emulate what we built, but it was an interesting puzzle to work on and we were able to cut through a lot of bureaucracy quickly that had held us back for a few years wrt accessing the sensitive data they needed to search.

But I was always looking for other options for rebuilding the service within those constraints and found Quickwit when it was under active development. I really admire their work ethic and their engineering. Beautifully simple software that tends to Just Work™. It's also one of the first projects that made me really understand people's appreciation for Rust as well outside of just loving Cargo.

>All units and projectiles are simulated in real-time. The game offers fully simulated projectile ballistics, explosion physics and terrain deformation.

This actually looks really cool!

KDB is the epitome of finance software. Cool in theory, and a decade ago probably the right choice if you needed to separate compute/storage and just get crazy high performance, but god damn I hate actually having to deal with it. Would rather DuckDB or even just Polars every time these days.

I mean TPU's are generally the only off the shelf viable and reliable alternative to Nvidia GPU's for training and there are a lot of former Google staff at Apple that have experience with TPU's, so I wouldn't be surprised.

If you were a Google PM trying to drive ad revenue, wouldn’t it be strategically in your best interest to make it look like uBlock Origin Lite worked equally well until uBlock Origin standard was no longer available and only then start introducing anti-ad blocking mitigations that uBlock Origin Lite couldn’t manage due to MV3 constraints?

Yeah, this is why I mentioned Gemini with the context caching. It's not out yet, but supposedly launching soon. You pay a lower rate for storing the system prompt or whatever you dump in before the user query, plus you don't have to wait the full minute or so for all your research to be ingested every time.

https://ai.google.dev/gemini-api/docs/caching#get-started

This was the first thing I thought of when I read the headline. I'd be curious to hear his thoughts on this student design.

I've been disappointed in some of these services before, because in the early days and I suspect now still as well they used RAG for obvious reasons. I feel like for research it's rare that RAG works really well.

What I'm really excited about is a tool like Elicit using the new Google Gemini 1.5 Pro/Ultra models with the 2 Million token window sizes, filtering down the papers using traditional search and high quality meta-data, then critically, prompt/activation caching to make the tool economically viable.

Maybe it won't work better, but I'm willing to bet it'll find those really specific ideas/needles in the haystack a lot more often than vanilla RAG will.

That is a very cool setup!

My org would never allow that as we're in a highly regulated and security conscious space.

Totally agree about the BQ costs. The free tier is great and I think pretty generous, but if you're not very careful with enforcing table creation only with partitioning and clustering as much as possible, and don't enforce some training for devs on how to deal with columnar DB's if they're not familiar, the bills can get pretty crazy quickly.

Frankly I think for long term hoarding BQ is hard to beat. The storage costs are pretty reasonable and you never pay for compute until you actually run a query, so if you're mostly just hoarding, well, you're probably going to save a lot of time, money, and effort in the long run.

Good luck! They have some really great tutorials on how to get started with BQ and geospatial data. One other nuance of BigQuery that doesn't seem to apply to many other tools in this space is that you can enable partitioning on your tables in addition to clustering on the geometry (Geography in BQ) column.

https://cloud.google.com/bigquery/docs/geospatial-data#parti...

It's FEMA's NFHL. I can't recall the specific layer of the GDB file, but you could probably figure it out. Try loading up Iowa into redshift and if that works for you I'd be quite surprised.

My org has a very large AWS spend and we got to have a chat with some of their SWE's that work on the geospatial processing features for Redshift and Athena. We described what we needed and they said our only option was to aggregate the data first or drop the offending rows. Obviously we're not interested in compromising our work just to use a specific tool, so we opted for better tools.

The crux of the issue was that the large problem column was the geometry itself. Specifically, MultiPolygon. You need to use the geometry datatype for this[1]. However, our MultiPolygon column was 10's to 100's of MB's. Well outside the max size for the Super datatype from what I can tell as it looks like that's 16 MB.

[1]: https://docs.aws.amazon.com/redshift/latest/dg/GeometryType-...