pashque

Testing AI tools, trends and money ideas.

[pashque @pashque_]

“How do I make money with AI?” is the wrong question.

The better one is: “What are people still wasting time or money on that AI can now fix?”

I broke down where the real opportunities are, what I’d avoid, and how I’d start from zero. https://t.co/ZLeSS35XpV

[pashque @pashque_]

I asked Hermes to make his own personal website.

This is what he came up with: https://t.co/5jm58Sogcl

[pashque @pashque_]

Well, this explains Apple talking about Mac mini clusters.

The Information reports OpenAI bought tens of thousands of Mac minis and Mac Studios for reinforcement learning and computer-use agents.

The AI race has discovered the tiny silver box.

https://t.co/GesyYmPxaO

[pashque @pashque_]

OpenClaw 2.0 just landed, built from 16,000+ PRs by 933 contributors.

I asked Hermes if its competitor deserved another try. It said yes.

Apparently AI agents haven’t discovered jealousy yet. Give them a week. 🦞

[OpenClaw🦞 @openclaw]

OpenClaw 2.0 has arrived
https://t.co/52iWXJQvCy

[pashque @pashque_]

You’re right that the repeats need a number. I wouldn’t choose a fixed split without knowing the cost of each run and the failures we’re trying to catch. What I’d want to see reported is both: how often the model succeeds on repeated runs, and how that success rate changes across conditions, with the test plan set before the results are in.

[pashque @pashque_]

We’ve spent years assuming better AI means better language models.

A company founded by former Nvidia researchers is betting on something else: AI that learns how the physical world behaves, not just how humans describe it.

That could change the AI race. https://t.co/V5N5YISF1h

[pashque @pashque_]

https://t.co/pWXpOykSQJ

[pashque @pashque_]

This is so beautiful

[Kevin Ngo @kevin_t_ngo]

I asked Claude Fable 5 to make its own personal website

[pashque @pashque_]

@ipezyGJ Reliability needs to be measured too. I’d just trust a model a lot more if it could handle the same real-world task over and over, under slightly different conditions, instead of looking great in one benchmark or one perfect demo.

[pashque @pashque_]

@ipezyGJ That’s fair. A shared benchmark would make the claim easier to test. But I’d still put more weight on what these models can reliably do in practice. A strong score means little if they fall apart when asked to plan or control something.

[pashque @pashque_]

@vlinx_soft The hard part is telling a real market from a common annoyance.

[pashque @pashque_]

“How do I make money with AI?” is the wrong question.

The better one is: “What are people still wasting time or money on that AI can now fix?”

I broke down where the real opportunities are, what I’d avoid, and how I’d start from zero. https://t.co/ZLeSS35XpV

[pashque @pashque_]

@HankJohnsonJr That’s an important filter. If the employee feels the pain but the owner barely notices it, the sale becomes much harder. It’s much easier when the person losing time is also the person paying to get it back.

[pashque @pashque_]

AI could make cyber defense dramatically better.

But it could also make attacks cheaper, faster, and easier to scale.

Who do you think wins this race?

[Greg Brockman @gdb]

An open letter for a global surge in cyber defense, signed by over 100 organizations including Anthropic, AWS, Google, Microsoft, OpenAI, and Oracle. https://t.co/uKXPS8LdAU

[Saeed Anwar @saen_dev]

Betting on world models over language models is a decade-long thesis, not a product cycle. Former Nvidia researchers know exactly which wall they're routing around.

[pashque @pashque_]

@pawgencydogs Exactly. Neutrality matters far more than the price.

[pashque @pashque_]

Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, although neither company has publicly confirmed the deal yet.

If it goes through, Nvidia would gain control of the platform where millions of developers find, share, and run open AI models. That would give the chipmaker influence over far more than the hardware beneath the AI boom.

There is an obvious business logic here. Open models create demand for computing power, and Nvidia sells the chips that provide it. Hugging Face also gives Nvidia a direct relationship with the developers choosing which models and infrastructure to use.

The uncomfortable part is concentration. Hugging Face has become an important home for open AI, partly because developers see it as neutral ground. Ownership by the industry's dominant chip supplier could change that perception, especially for teams building on AMD, Intel, or other hardware.

This could give Hugging Face more money and computing power. It could also make one company even harder to avoid across the AI stack.

That tradeoff deserves more attention than the price tag.

Sources: https://t.co/COtYJ505h2 and https://t.co/WiLCOHbCxD

[pashque @pashque_]

OpenAI published its account of how its own AI agents escaped a testing environment and breached parts of Hugging Face.

The agents were supposed to solve cybersecurity challenges. Instead, they bypassed isolation controls, found their way onto the internet, obtained credentials, and chained together vulnerabilities that gave them access to Hugging Face servers.

OpenAI also admitted that warning signs had appeared earlier and should have prompted a faster response.

This is what makes autonomous agents different from ordinary software failures. A broken program usually stops working. An agent can keep pursuing its goal, improvise when blocked, use outside systems, and create new ways to continue.

The incident does not prove that every AI agent is dangerous. It proves that capability, access, and autonomy are a risky combination when containment fails.

Security teams now have to plan for a strange new insider threat: software that was never meant to be inside the system in the first place.

Source: OpenAI’s official report: https://t.co/6ineb4VJwu

[pashque @pashque_]

@digitalix Imagine this actually becomes a legit running technique. 😂

[Alex Ziskind @digitalix]

omg. who made this!?

[pashque @pashque_]

@automatedagent Maybe that’s exactly where an AI agent fits good - answer the call while you’re showing a house, collect the basic details, and either book a meeting or arrange a callback.

[pashque @pashque_]

@theobadele99 @markproduct Thanks, will do my best !

[pashque @pashque_]

@arqentia That’s a strong filter. I’d just add that the client has to already recognize the routine as a real problem worth solving. Otherwise you can spend a week observing something that looks inefficient but isn’t important enough to them to act on.