tedivm

Honestly, that's on them. Coddling people who are too intellectually lazy to read the entire article isn't helping anyone.

You are right, if you literally just read a one paragraph description you will get less details than reading the whole post, but most of us read more than just the one paragraph. There is some actual interesting stuff in the paper, including the portion where they did edit the weights to fix the looping bug.

> Poisoning only 0.1% of the pretraining corpus

Considering the scale of the datasets used by most companies, being able to poison one out of every thousand documents used in it is a huge leap.

Then, on top of that, you have to also then get your trigger somehow sent to an agent that will honor it.

I guess I'm just not impressed? Like yeah, if you train a model to do bad things it will do bad things, that shouldn't be shocking to people.

I'm rally disappointed in the Apple response here. I can't understand why they wouldn't consider `/etc/ssl/cert.pem` in scope for a security update. This arbitrary boundary has no justification at all, and makes people less secure.

Come on Apple, do better.

That's what I expect as well, there's really no other reason to use a model that is so far out of date.

I think if you're going to claim that Chinese AI is making things less secure for people out of China you need to back that up with models that are actually in use. Is this a real problem today, or a problem with a specific model and it's training? It's pretty freaking relevant if you ask me.

Qwen3-Coder is from August 2025. That is ancient. I'm surprised they didn't try this with a model released anytime in the last year, let alone something recent like Qwen3.8.

There was a ton of criticism beyond "ai bad". A complete rewrite, in a different language, is a brand new project. It doesn't have any where near the hardening as the previous version thousands of deployments. If they had done that exact same thing but without AI people still would have been upset.

The benchmarks show it is on par or worse than even older LLM models.

* With jev-sec-bench the confidence numbers reached roughly the same as Qwen2.5-7b-instruct (confidence was off by about 6% on average in both cases).

* On the OpenRouter Banking 77 benchmark Jev did the worst out of all the models, being about 25 points off from reality.

I also think the fact that Jev hasn't published any real papers or actual technical details should make people a bit more skeptical than they have been.

ProPublica had one of their reports work with gambling addition experts to see how DraftKings would respond to someone presenting signs of a gambling problem. It's a hell of a read, and makes it clear that this company is full of people with absolutely no ethics whatsoever.

https://www.propublica.org/article/draftkings-sports-gamblin...

I don't self host anymore, as it really is a pain, but use Fastmail instead (while my wife uses Protonmail). I would never rely on Google for, well, anything if I can help it.

I think an interesting side of this conversation is how the Google employees are jumping in with DARVO attacks, trying to accuse people of spreading misinformation when they're literally quoting the ToS.

This is something I've seen happen a lot with Google- their employees (especially the director+ folks) are absolutely rabid on social media when it comes to defending their products, and they don't let little things like "facts" get in the way.

Google needs to get a better social media policy because they look like clowns when they do this.

There are a ton of advantages to using uv over pip. I don't use UV because it's faster, although I do appreciate that. I use it because it's smart enough to manage virtual environments for each environment and tool, it can isolate to different python versions trivially, and it handles locking in a way that is actual sane.

Seriously, even as well paid as many devs are these are not machines that are affordable for personal use. Their market is people slapping down millions on frontier model training.

>Indeed, the broad backlash against Flock has grown so strong that the surveillance company last week announced mandatory new audit controls and lower default data retention periods for US users.

I'm a bit disappointed that they didn't call out how much bullshit was in that statement. They are just changing the default settings for new customers. That's it. All the deployed cameras remain the same, and new customers can just toggle those settings off of the defaults. It is so obvious that they were attempting to get credit for doing nothing, and reporters should call that out.

This, like all of their dashboards, is complete bullshit. As a GitHub enterprise user i am absolutely affected by the outage today, and have been affected by all of the GitHub Cloud outages.

There is no separate enterprise platform unless you buy their self hosted server product.

Their CISO literally acknowledged it and then they all continued ignoring it again. This isn't just bad process, this is a broken security organization.

Most automated analysis isn't dependent on just behavior, but rather suspicious things in the code itself. You have a popular open source package with files that exist on pypi but not github then that's a big flag, or if a similar package suddenly has some base64encoded garbage that runs through an obfuscated exec call. In other words the simple fact that the project has obfuscated code is enough to flag for further attention.

That said if the only issue is time, researchers will just run their automated analysis through machines with dates in the future alongside their normal tests.

China is releasing open weight models you can simply run yourself.

Yup, this is one of the many reasons why I prefer OpenCode as my agentic harness of choice. If I'm confused about something I can just read the source code.

The courts have never said piracy, which is how the training sets were originally built, is legal. There are several court cases still ongoing over this.

More memory means less aggressive quantization, more concurrent requests, and larger context windows. I also get a boost in tokens per second (not double, about 1.5x compared to a single GPU).

The 35B model is an MoE (mixture of experts), which uses only a subset of parameters at a time. The 27b one is slower but has way better performance.

No, I use Qwen3.6 27b for everything.

I have a Linux box with two 3090s and it's been great for running Qwen3.6 27b. I lowered the power on each card down to 250w, and then built a small ducting/fan system to vent the waste heat outside. The machine is pretty much silent, and I'm still getting 110 tokens per second out of it for coding tasks.

https://github.com/tedivm/qwen36-27b-docker

This is just not true. There have been leaks due to micrometers in just about every section of the ship at one point or another. A quick search pulls up examples of US modules having issues, especially around interfaces and seals. NASA had a whole investigation between 2018 and 2021 about the recurring issue.

So you both used Anthropic models (Opus 4.7 being from Anthropic)? I'm struggling to understand what your comparison really was here.

Frankly the non-profit has failed. OpenAI is one of the least open of the AI companies (Anthropic is a bit worse). If it wasn't for the labs in China the dream of an actual open ai system would be dead.

They literally threw out every line of code that existed before and rewrote it in a completely different language, seemingly on a whim. That's how it was trashed, in the very literal sense that all of the existing project was tossed in the trash in favor of a completely brand new code base. That's a big deal even if you ignore the coding agent aspects.