Nvidia's open-source model surprise, Gemini does your errands, and the AI royalty fraud that just got a guilty plea

Something quietly shifted in the open-source AI landscape over the weekend - and it came from a company most people associate with hardware, not model weights. Let's get into it.

Nvidia's Nemotron-Cascade 2 uses a Mixture-of-Experts (MoE) architecture. In an MoE model, what does "active parameters" actually mean?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Nvidia just dropped Nemotron-Cascade 2, an open-weight 30B Mixture-of-Experts model that only activates 3 billion parameters per inference pass. That gap matters enormously - you get the reasoning quality of a large model at a fraction of the compute cost.

According to Nvidia, Nemotron-Cascade 2 is the second open-weight LLM to achieve Gold Medal-level performance on the 2025 international reasoning benchmarks. The focus is on what Nvidia calls "intelligence density" - packing frontier-level capability into a model that doesn't require frontier-level hardware to run.

For developers building agentic applications who need strong reasoning without API costs, this is a genuinely useful release. It's fully open-weight, so you can download and run it locally or deploy it on your own infrastructure.

Read the full story →

The Verge got hands-on time with Gemini's new task automation feature on the Pixel 10 Pro and Galaxy S26 Ultra, and the honest verdict is: it works. Right now, it's limited to a handful of food delivery and rideshare services, it runs in beta, and it's slower than doing things yourself. But it actually completes tasks end-to-end without you touching the screen.

This is Google's first meaningful step toward letting Google Gemini fully operate apps on your behalf. The reviewer noted it doesn't solve any obvious problem most people have right now - but it's early. The category of AI that takes actions for you, rather than just answering questions, is where every major lab is pointing.

The clunkiness is expected for a beta. What's notable is that it exists at all on consumer hardware. If the app library expands significantly over the next few months, this could be the feature that finally makes on-device AI agents feel real rather than theoretical.

Read the full story →

Michael Smith, 52, of North Carolina, has pleaded guilty to defrauding music streaming platforms and fellow musicians out of millions in royalties. His method: flood platforms with thousands of AI-generated songs, then use automated bots to stream those tracks on repeat - generating royalty payouts that came directly out of the pool shared with real artists.

Streaming royalties aren't paid at a fixed rate per play - they come from a shared revenue pool. Every fake stream Smith generated diluted the earnings of every other musician on the platform. At scale, with thousands of AI tracks running around the clock, the damage to real artists compounds fast.

This case sits at the intersection of AI-generated content, music industry fraud, and platform accountability. Expect streaming services to accelerate their AI detection efforts following this verdict - the legal precedent is now set.

Read the full story →

When the UK government signed its memorandum of understanding with OpenAI last year, ministers positioned it as a cornerstone of the country's AI-led public service reform. A Freedom of Information request now reveals there is no evidence the government has tested any OpenAI technology at all in the eight months since that announcement.

The story is a useful reminder that government AI partnerships are often more about signaling than substance - at least initially. The UK technology policy space has been particularly active in making headline-friendly agreements, but the gap between announcement and actual deployment in public services remains large.

The government pushed back, saying it "delivers timely, high-quality work to meet public needs" - but produced no evidence of any actual testing. If you're building tools for the public sector and wondering why procurement cycles feel glacial, this story has your answer.

Read the full story →

TechCrunch is asking a question that's starting to circulate seriously in tech hiring circles: are AI token credits the new signing bonus? Some companies are already bundling API credits into compensation offers for engineers, and on the surface it looks like a perk. But the framing matters enormously.

If tokens are treated as a fourth pillar of compensation alongside salary, equity, and benefits, engineers gain genuine value. If they replace existing perks or shift the cost of AI tooling from employer to employee, that's a different story entirely. The article notes engineers may want to hold the line before accepting token credits as a win - the value of those credits depends entirely on how heavily they end up using the tools in practice.

If you're thinking about token costs and how to calculate them properly, our Token Calculator can help you model actual usage before agreeing to any package.

Read the full story →

Speaking of building fast - if you need to spin up a web presence without the usual overhead, 60sec.site lets you build and launch an AI-powered website in under a minute. Worth a look if you've been putting off that project.

The answer is B - the number of parameters activated and used for each individual token during inference. In a Mixture-of-Experts model, only a small subset of "expert" layers activates for any given input. Nemotron-Cascade 2 has 30 billion total parameters but only routes each token through 3 billion of them - which is why it's so computationally efficient while still delivering strong reasoning performance.

Gemini's task automation is live in beta and doing things like ordering Uber rides on your behalf. Would you actually use an AI to control apps for you - or does that feel like one step too far into your phone? Hit reply and let me know. I read every response.

That's all for today. More AI news tomorrow - find everything in our archive if you want to catch up on anything you missed. See you then.