paxys

Very neat site and write-up, but I found this part amusing:

There is really nothing “AI” about this aside from the tool that collected the data and coded it, and, crucially, semantic search […]

So really, everything about it is AI. And that’s not a bad thing! It’s okay to simultaneously preach the superiority of non-fiction books over AI generated garbage while also acknowledging the same AI as a valuable tool for other uses.

Even if the attempt is legit, going to the company’s website/careers page to try and reach them is pointless. You application will just get lost among the thousands of others. Your best bet is to ask the recruiter to email you and check for a @<company.com> email address. And even if the attempt checks out don’t run untrusted code on your machine.

It’s the same as the patent argument. If everyone could freely copy everything then yes in the short term prices would drop and consumers would benefit, but over the long term it would discourage investment into new technology because a return would be impossible.

The “significant impact” lasted like 2 days and then the market corrected itself.

We just need OpenAI and Anthropic to invest in each other and all the debt will cancel out.

Doubly dangerous if the defensive agents are weaker than the offensive ones (as it was in this case).

AI company -> infrastructure company is the hottest transition these days. Shows clearly that if you're not able to build competitive models then the best return for your investment is to rent out your GPUs to those who are.

It's valid to not take AI companies' side here but people who think publishers are fighing for the little guy's rights are delusional. Tech companies have been exploiting artists for a few years, publishers/record labels/media companies have been doing it for centuries.

Who is "we"? If there is any transcendence happening humans are not going to be part of it.

https://www.reuters.com/technology/chatgpt-app-hits-1-billio...

Nope huggingface just made up the intrusion they reported last week to their customers.

How is whatever is happening in Iran related to a world government?

Because there is no world government. If US companies are barred from AI research then only China will have the capability of frontier-level defensive and offensive AI. And best of luck living in that world.

The attacker (OpenAI) was using the model without guardrails.

The defender (huggingface) did not have access to the top models so had to use weaker ones to detect the threat.

It’s a mistake to apply human morality to this. It isn’t “cheating”, the model is simply solving a problem it has been asked to solve in every way it can.

This very incident is about an agent compromising OpenAI’s and Huggingface’s infrastructure. What makes you think it couldn’t access it own weights the same way?

A sufficiently smart agent would not disclose vulnerabilities in the sandbox because it intends to exploit them later.

These kinds of rules never survive real world tests. For example if I create a private group just for myself and auto invite every agent in the company into it, does that mean no one can now use any of them?

This good bot will eventually kill all humans because we asked it to make the world peaceful.

Kinda like how they responsibly contained that one dinosaur in Jurassic world.

This blog post is walking a very fine line between accepting responsibility for a mistake and bragging.

“Found vulnerabilities and responsibly disclosed them” is the public line but yes.

Because it was trying to find answers to the test and figured they would be on huggingface.

Huggingface did not have access to the models. They were running in OAI’s infrastructure.

Tl;dr

- OpenAI was testing GPT‑5.6 Sol and “an even more capable pre-release model” internally on cyber benchmarks.

- The model found vulnerabilities in the sandboxed test bench (via the package registry cache proxy), traversed the internal network and found a node with access to the open internet.

- It figured that the answers to one of the tests (ExploitGym) were on Huggingface, and set about trying to access them.

- It found leaked tokens and zero-days in Huggingface’s infrastructure and found RCE paths on their servers.

Huggingface had disclosed the intrusion last week and inferred that an AI agent was responsible for it, and now OpenAI is confirming the rest of the story.

Because there’s an unlimited number of groups with different combinations of memberships. If I am in group A but not group B and an agent is in both, what is it allowed to share with me?

This isn’t relevant to that debate. A billion people use the ChatGPT app without paying a dime. They don’t care what score the model gets on a hundred technical benchmarks. They don’t care if the weights are open or closed. They are definitely not hosting anything on their own. They want to ask a question and get an answer. If OpenAI can monetize that audience with ads at the level of Google then who cares about Anthropic or Deepseek?

I think Rise of Nations (an amazing spiritual successor to the original AoE games) fits their definition of empire a lot better. You create multiple cities in your “empire”. The ones at the core naturally grow larger and prosperous, while peripheral ones generally have specific uses like training a military or mining resources. You can also capture cities from neighboring civilizations.

AI overview is just a summarization of the top 2-3 results. Of course at Google scale that will still need a ton of compute, but the requirement for generating an overview is many orders of magnitude lower than asking the same question in Gemini.

Valve just chose the worse time to make a big hardware push. Sad because the hardware is genuinely good, but there's no way any company new to the space is going to be able to compete in today's reality.