tedivm

Yeah, this was going to happen regardless of the US.

> The European Union Agency for Cybersecurity (ENISA) first announced the project in June 2024 under a mandate from the EU's Network and Information Security 2 Directive, and quietly rolled out a limited-access beta version last month during a period of uncertainty surrounding the United States' Common Vulnerabilities and Exposures (CVE) program.

No one said that TOS's override the law.

If you shout something in public, someone might hear you. If you shout in a locked house they won't hear you unless they break in.

If you post something publicly on the internet, then people can look at it. If you post something privately on the internet, people have to agree to your terms to view it.

That's it, that's the logic the government uses to allow website scraping.

I think calling this "scraping" is hiding an important point. Scraping tends to mean crawling the web and pulling down stuff people put up. In the US (and yes, I know that this is a UK article) it is legal to scrape, with the idea being that if people responded to the request that's on them. This is why lots of sites have login gates to access data, since to log in you have to agree to a TOS.

AI companies haven't just been scraping, they've been pirating. I think it's important because people consent to put their comments on websites, but authors don't consent to have their work stolen. Mixing the two together as if they are the same is an attempt by AI companies to muddy the ethical waters a bit more than I'm comfortable with.

Using MCP you can send "notifications" to the server, and the server can send back notifications including the availability of new tools.

So this isn't the same as saying "this user agent gets X, this gets Y". It's more like "this client requested access to X set of tools, so we sent back a notification with the list of those additional tools".

This is why I do think websockets make more sense in a lot of ways here, as there's a lot more two way communication here than you'd expect in a typically API. This communication also is very session based, which is another thing that doesn't make sense for most OpenAPI specs which assume a more REST-like stateless setup.

I'm not saying MCP is perfect, but it's better than OpenAPI for LLMs for a few reasons.

* MCP tools can be described simply and without a lot of text. OpenAPI specs are often huge. This is important because the more context you provide an LLM the more expensive it is to run, and the larger model you need to use to be effective. If you provide a lot of tools then using OpenAPI specs could take up way too much for context, while the same tools for MCP will use much less.

* LLMs aren't actually making the calls, it's the engine driving it. What happens when an LLM wants to make a call is it responds directly with a block of text that the engine catches and uses to run the command. This allows LLMs to work like they're used to: figuring out text to output. This has a lot of benefits: less tokens to output than a big JSON blob is going to be cheaper.

* OpenAPI specs are static, but MCP allows for more dynamic tool usage. This can mean that different clients can get different specs, or that tools can be added after the client has connected (possibly in response to something the client sent). OpenAPI specs aren't nearly that flexible.

This isn't to say there aren't problems. I think the transport layer can use some work, as OP sent, but if you play around in their repo you can see websocket examples so I wouldn't be surprised if that was coming. Also the idea that "interns" are the ones making the libraries is an absolute joke, as the FastMCP implementation (which was turned into the official spec) is pretty solid. The mixture of hyperbole with some reasonable points really ruins this article.

To be president you have to be a resident for the previous 14 years, so he wouldn't be eligible unless he moved here today and waited 14 years. He'd be 83 at that point.

They'll train a new version to fix the problem.

Once you have the models on local storage you can move pretty quickly from there to VRAM, I've never found that to be the biggest bottleneck. The problem is provisioning itself, especially if you have to actually move models locally. Some of this can be avoided with extremely expensive networking (infiniband to a NAS with model weights), but that's not something you're going to have fun dealing with in a cloud environment.

It might help to remember that the training process is essentially a game of "how fast can we shove data into these GPUs", and having a GPU sit idle because the data can't get into it fast enough is a challenge people have been tackling since at least the P100 series. This has resulted in improvements on the GPUs as well as all the hardware around them. Getting data into the chips is one of the most efficient processes at this point.

Even Alibaba is releasing some amazing models these days. Qwen 3 is pretty remarkable, especially considering the variety of hardware the variants of it can run on.

Github does give free copilot access to open source developers it considers important enough (which is a pretty low bar). While not the same as actually paying, it's the only example I can think of where the company that used people's copyrighted material actually gave something back to those people.

The reason the ValKey fork happened so fast is that AWS was already sponsoring developers for the open source redis project.

IBM did it with their Granite models.

It's not just about reproducibility (although I do think that's important), it's about analysis of the model. With traditional software you have a pretty well defined "this code does this", but with machine learning models one of the only ways to validate that bias or propaganda hasn't been inserted during training.

They allow a major component of the model, the data, to be withheld.

According to the story you linked, they claim they had a warrant but the judge and other staff say the warrant was never presented.

Further, ICE has a habit of lying about this. They refer to the documents they write as "administrative warrants", which are not real judicial warrants and have no legal standing at all. So when ICE says they presented a warrant it's important to dig in and see if it was one of their fake ones.

Here's an article about that last point from the same source you used: https://www.jsonline.com/story/news/local/2025/04/23/what-is...

The ICE agents didn't have a warrant, so the judge was under no legal obligation to say anything to them at all.

The Open Source AI Definition (OSAID) is a slap in the face to anyone who has been part of the open source community. Allowing companies to redefine "Open" to allow closed components is a complete betrayal of everything the OSI should stand for, and it was done purely so large companies can pretend their closed models are open.

What you're proposing, updating the license file to list the authors, is a pretty common way to do this. It does mean that the code is mixed a bit, so it would be hard to split who owns what, but this is only relevant if one of the copyright owners wants to change the license (as they can legally only do that to the code they own).

The "notice" is the literal license file. It is illegal to strip someone else's license from their work. It doesn't matter that they replaced MIT with MIT, because they stripped the author and attribution from it.

https://github.com/Azure/peerd/blob/main/LICENSE

If you read that file you'd think that Microsoft was the copyright holder, but they very clearly aren't.

I'm afraid you didn't read my response. I explicitly said I can't see a case where it isn't needed for some services. I never said it was required for every service. Once you've got it setup for one thing it's pretty easy to set it up everywhere (unless you're manually deploying, which is an obvious problem).

I really don't see many scenarios where HTTPS isn't needed for at least some internal services.

Previous studies I've read did account for this, and put the number at about 3%. That said CT manufacturers have been working on a bunch of technology to make this better, either by doing a better job at not aiming at sensitive areas or by reducing the amount of radiation needed to get the same resolution in the images.

There are already a ton of high-end gaming consoles you can buy today with nvidia chips. In fact, this is exactly what I did for my own AI machine (dual 3090s).

I've made two big purchases from Amazon in the last year and had to return both because the supposedly new item was obviously a return (one had missing pieces, the other had visible damage). It really is at the point where you can't trust Amazon to ship you what you actually purchased due to how they manage inventory.

I don't think we need to pick between incompetence and malice here. I think both apply to most of their actions.

Seriously, I absolutely can't stand the azure ui. I don't think the AWS console is great, but it is definitely better than the Azure one to me.

The kernel is doing fine today, but I don't think that's sustainable. The average age of the maintainers seems to be rising, and plenty of skilled people completely avoid kernel development explicitly because of the hostility and, frankly, assholish behavior that comes from the folks on these mailing lists.

Eventually a lot of these people are going to age out, retire, or otherwise move on. I do think there will be a crisis moment at some point in the future.

They literally cited LeCun in their GPT papers.

I was more referring to this paper from 2015:

https://scholar.google.com/citations?view_op=view_citation&h...

Basically all LLM can trace their origin back to that paper.

This was just a single example though. The whole point is that people build on the work from the past, and that this is normal.

OpenAI has been hiding their datasets, and certainly haven't credited me for the data they stole from my website and github repositories. If OpenAI doesn't think they should give attribution to the data they used, it seems weird to require that of others.

Edit: Responding to your edit, Deepseek only claimed that the final training run was $5m, not that the whole process caught that (they even call this out). I think it's important to acknowledge that, even if they did get some training data from OpenAI, this is a remarkable achievement.