querez

I didn't know about the static cache, that was a nice learning for me. The rest is pretty obvious /too lengthy for everyone who ever did profiling/benching. Also, the text has a lot of LLM slop-phrases in it that need cleaning up ("the X is structural", "the Y is real", "the gap is Z")

AI doesn't need to be AGI to be useful. Surely, you've tried Claude Code before and found it more helpful than Clippy?

> The project owner strongly emphasize the no LLM dependency, in a world of AI slope this is so refreshing.

The project owner is talking about LLVM,a compiler toolkit, not an LLM.

That's not how I remember it. Excitement for python strongly predated ML and data science. I remember python being the cool new language in 1997 when I was still in high school. Python 2.4 was already out, and O'Reilly had put several books kn the topic already it. Python was known as this almost pseudo code like language thst used indentation for blocking. MIT was considering switching to it for its introductory classes. It was definitely already hyped back then -- which led to U Toronto picking it for its first ML projects that eventually everyone adopted when deep learning got started.

I think you're underestimating what AI can do in the coding space. It is an extreme paradigm shift. It's not like "we wrote C, but now we switch to C++, so now we think in objects and templates". It's closer to the shift from assembly to a higher level language. Your goal is still the same. But suddenly you're working in a completely newer level of abstraction where a lot of the manual work that used to be your main concern is suddenly automated away.

If you never tried Claude Code, give it s try. It's very easy to get I to. And you'll soon see how powerful it is.

I assume TPU TCO is significantly cheaper than GPU TCO. At the same time, I also assume that market demand for GPUs is higher than TPUs (external tooling is just more suited to GPU -- e.g. I'm not sure what the Pytorch-on-TPU story is these days, but I'd be astounded if it's on par with their GPU support). So moving all your internal teams to TPUs means that all the GPUs can be allocated to GCP.

> No Gemini was not "entirely trained on TPUs". They did hundreds of experiments on GPUs to get to the final training run done entirely on TPUs. GCP literally has millions of GPUs and you bet your ass that the gemini team has access to them and uses them daily.

You are wrong. Gemini was definitely trained entirely on TPU. Of course your point of "you need to count failed experiments, too". Is correct. But you seem to have misconceptions around how deepmind operates and what infra it possess. Deepmind (or barely any of Google internal stuff) runs on Borg, an internal cloud system, which is completely separate (and different) from gcp. Deepmind does not have access to any meaningful gcp resources. And Borg barely has any GPUs. At the time I left deepmind, the amount of tpu compute available was probably 1000x to 10000x larger than the amount of gpu compute. You would never even think of seriously using GPUs for neural net training, it's too limited (in terms of available compute) and expensive (in terms of internal resource allocation units), and frankly less well supported by internal tooling than tpu. Even for small, short experiments, you would always use TPUs.

I grew up in the Internet at that time, and it's certainly not how I type. So you might want to be more specific about which sites or subcultures you think this style is representative of?

The sentence is constructed, weirdly, but it's meant to say that fever is "killing off unknown foreign bodies"

Arguably both Go and Python also have great stdlibs. The only advantage that JVM and .NET have is a default GUI package. Which is fair, but keeps getting less and less relevant as people rely more on web UIs.

I don't think "inference" as a verb has become technical slang. At least not in my bubble.

"developers can prototype, fine-tune, and inference [AI models]"...

shouldn't it be infer?

I think what is distinct in this proposal is that there are n 1:n channels

Why? I fail to see how using chromium as basis for other apps has impact on who has the power to innovate in the browser space?

> The geometric mean of the timings improved from 218 to 12, a ca. 20× improvement.

Why do they use the geometric mean to average execution times?

As someone who worked at both Brain and DeepMind: their cultures were very different. Brain was bottom-up, open-field research whatever you want/care for. DeepMind was much more narrowly focused and massively more top-down.

>> The freedom to waste time. The freedom to waste resources. And the autonomy to decide how.

> As the article notes, several companies (Apple, Google, etc.) could (currently) afford to fund such a lab, but there is no way their management and shareholders would approve.

Google did set up such a lab. The mission of Google Brain was literally to hire smart people and let them do work on whatever they want. ("Google Brain team members set their own research agenda, with the team as a whole maintaining a portfolio of projects across different time horizons and levels of risk." -- https://research.google.com/teams/brain/). Unsurprisingly, Google Brain is the place that originated the Transformer that powers the current AI craze (and many, many, many other AI innovations).

I guess it depends on use-cases. I use ChatGPT a lot for "trivial" questions a la "how do I uncommit a specific file in my last git commit" or "how do I paste from one PIL.Image into another one". In the past I would have to search google, click on the StackOverflow link, and then parse that whole page. Asking ChatGPT to give me just the snippet is faster, so doesn't get me out of my flow as much.

I use a very simple custom system prompt (not on my work machine at the moment, but essentially something along the lines of "for technical questions, please be concise and to the point, and when asked for code, omit explanations and emit just the code itself unless I ask for explanations"), and it does wonders.

So, with regard to Greenland, he specifically refused to rule out military action [0]. Which in diplomatic protocols comes very darn close to threatening a military attacks.

As far as Canada goes, Trump's national security advisors is on the record [4] saying he doesn't think there's any plans about this. Which is also an extremely uncertain way of talking about military action about your closest neighbor. Trump has repeatedly talked about making Canada the 51st state [1], and called Trudeau a "Governor" [2]. Here's a New York time article summing up the types of threats/attacks Trump made wrt. Canada[3].

[0] https://apnews.com/article/trump-biden-offshore-drilling-gul... [1] https://apnews.com/article/how-canada-could-become-us-state-... [2] https://www.nbcnews.com/politics/donald-trump/trump-mocks-pr... [3] https://www.nytimes.com/2025/03/11/us/politics/canada-trump-... [4] https://www.nbcnews.com/politics/donald-trump/trump-national...

You didn't read the article, did you? Hint: it's in the last third.

It isn't obvious at a first glance, but this is a fork of ocis (i.e., ownCloud's Go rewrite). It is hosted on Github(https://github.com/opencloud-eu/opencloud). They don't really provide any info on how/what is different from ocis that I could find.

I see twoo reasons:

1. Because they're researchers, not devops experts. They release the model in the way that they are most familiar with, because it's easiest for them. And I say that as someone who's released/open-sourced a lot of AI models: I can see how Docker is useful and all that, but why would I invest the time to do package up my code? It took long enough to cut through the red tape (e.g. my company's release process), clean up the code, document stuff. I did that mostly because I had to (red tape) or because it also benefits me (refactorings & docs). But docker is something that is not immediately useful for myself. If people find my stuff useful, let them do it and repackage it.

2. most people using these model don't use them in docker files. Sure, end users might do that. But that's not the primary target for the research labs pushing these models out. They want to reach other researchers. And researchers want to use these models in their own research: They take them and plug them into python scripts and hack away: to label data, to finetune, to investigate. And all of those tasks are much harder if the model is hidden away in a container.

The paper was published in Nature *Scientific Reports*, not in Nature. Nature SR is a low-quality journal due to its very limited peer review (SR peer reviews only focus on scientific validity of a submitted paper, rather than its perceived importance, significance, or impact -- ie., they check if you followed proper scientific procedures, but won't check if what you're saying makes sense / fits with the rest of the scientific literature). It's a known dumping ground for papers that aren't good enough to be published in a real scientific journal, and it survives/thrives thanks to its association with the prestigious Nature journal.

Two questions:

1) Going by the Runtime vs GLUE graph, the ModernBERT-Base is roughly as fast as BERT-BAse. Given its architecture (especially Alternating Attention), I'm curious why the model not considerably faster than its predecessor. Any insight you could share on that?

2) Most modern LLMs are Encoder+Decoder model. Why not chop of the decoder of one of these (e.g. a small Llama or Mistral or other liberally-licensed model) and train a short head on top?

Some very weird things in this.

1. The title makes it sound like the author spent a lot of time on this project. But really, this mostly consisted of noting down a couple of URLs per day. So maybe 5 min / day = ~130h spent on the project. Let's say 200h to be on the safe side.

2. "Get first analyses results out quickly based on a small dataset and don’t just collect data up front to “analyse it later”" => I think this actually killed the project. Collecting data for several years w/o actually doing anything doesn't with it is not a sound project.

3. "If I would have finished the project, this dataset would then have been released" ==> There is literally nothing stopping OP from still doing this. It costs maybe 2h of work and would potentially give a substantial benefit to others, i.e., turn this project into a win after all. I'm very puzzled why OP didn't do this.

Ok, so you have a review article about PC networking equipment, sponsored by Cisco. That's an ad for you. Now you go to the homepage of the site that wrote that review. They link to that review from their home page. Is that an ad? Probably, because it could contain an ad-like image as its banner (showing Cisco equipment) and maybe even a slightly ad-ish title ("Cisco outperforms other networking equipment in our recent review of x articles").

So, is another page, that links to that page also an ad? Because maybe the homepage that runs Cisco-sponsored reviews might also have a cisco-sponsored banner. And even if not, maybe over time they become associated as "the thing that reviews things were cisco always wins"...

Oh, and what if it wasn't a sponsorship. What if instead Cisco just gave them the hardware to test at a discount? Or for free? Would you not write a slightly more favourable review for nvidia if they gifted you their top-tier GPU? What if instead of gifting you the GPU, they just give it to you at a discount? Or earlier than normal customers? Or they just have nvidia-stickers for your kids in the box?

Ads are an extremly slipperly slope, they're part of human nature; everyone I ever met is trying to tell me about (=market) themselves, their beliefs, their current hobbies, that new burger place down the road, ....

It does sound like you're in a particularly bad mood, so yes, maybe our outlook does change. Maybe it helps to think of a darker timeline where Google would have kept all of these advances to itself and improved its ad revenue. Instead it shared the research freely with the world. And call me naive, but I use LLMs almost daily, so there definitely _is_ something of value that came out of all this progress. But YMMV, of course.

> honestly though Google would be that - but apparently it's easier to fund R&D on "selling copying machines" than "selling ads". Maybe "selling ads" earn _too much_ money ? I don't know.

I'm pretty sure Google Brain was exactly what you are looking for: People like to think of DeepMind, but honestly, Brain pretty much had Bell Labs/PARCs strategy: they hired a bunch of brilliant people and told them to just "research whatever is you think is cool". And think all the AI innovations that came out of Brain and were given to the world for free: Transformers, Vision Transformers, Diffusion Models, BERT (I'd consider that the first public LLM), Adam, and a gazillion of other cool stuff I can't think of right now.... Essentially, all of the current AI/LLM craze started at Brain.

I'm not sure what you're trying to get at: where's the irony in "Germany started a war to get some oil. Now someone else starts a war and Germany doesn't want to do business with them because they had enough of warmongering people"? What am I missing?