Did you know that we have LinkedIn, Instagram and X accounts
that you can follow?

Hi {{first_name|everyone}},

Here’s a way to know which AI actually answered you - Anthropic is working on cryptographic proof for it. After that, we’ve got a few fresh tools to try, a couple of quick updates from Netflix and Apple, and a robot octopus built from a single picture. Let’s see what’s in the pile.

🤿 DEEP DIVE

Anthropic is working on a way to prove who's answering.

Here's a weird thing about using AI. You don't actually know whether the model answering you is the model you asked for.

You might select a particular model, pay for access to it, and get a perfectly reasonable response. But there's no reliable way to tell from that response alone whether it came from the right model, a cheaper substitute, or even something an attacker slipped into the system.

And apparently, Anthropic has been working on a way to fix that.

It's called FlashProve, and the basic idea is to make AI responses come with cryptographic proof that they were actually produced by the model they're supposed to come from.

Not a badge saying "trust us." Actual mathematical proof.

Okay, this gets a little technical, but the idea is surprisingly straightforward.

First, the AI provider creates a kind of digital fingerprint of the model's exact weights. Those are the internal numbers that determine how the model behaves.

Then, when the model produces a response, FlashProve generates a separate proof showing that the response matches what that specific model would produce.

Someone checking the proof doesn't need access to the model's secret weights. They don't even need a powerful GPU.

And that's the clever part. The provider can prove which model was used without giving away the thing that makes the model proprietary.

Anthropic tested FlashProve on Llama 4 Maverick, a model with 400 billion parameters.

Using a single H100 GPU, the system generated proofs at roughly 110 tokens per second. Each proof was about 19 MB, and checking one took around 17 seconds on a 24-core CPU.

For comparison, an earlier open-source system proving the same model needed roughly 14 hours for 1,000 tokens.

That's a pretty enormous difference.

But there's a catch. FlashProve still uses around 70 to 95 times the GPU time of ordinary inference, based on Anthropic's comparisons. And the proofs are generated after the responses, rather than alongside them.

So this isn't something ready to quietly run behind every AI conversation.

One example Anthropic gives is independent AI testing.

Say someone is evaluating whether a particular AI model is safe. How do they know they're testing the same model that customers will actually use?

Or imagine a company claiming its software was checked for security vulnerabilities by a powerful AI model. How would anyone verify that the check really happened?

FlashProve offers a way to verify those claims without revealing the model itself.

There are still limitations. The current prototype handles contexts of up to 1,024 tokens, and its exact guarantees apply to a particular integer-based version of the model using greedy decoding.

But I think the underlying idea is the part worth remembering.

Right now, we mostly judge AI by what it says. Anthropic is exploring how to verify which AI actually said it.

🤝 FROM OUR FRIENDS

A quick share we thought you might find useful

AI is stepping out of the screen and into the physical world. Robotics VC deals hit record highs in early 2026: nearly $16B across 450+ deals.

Labor shortages, aging workforces and breakthroughs from NVIDIA, Tesla and specialized suppliers are accelerating automation in factories and warehouses worldwide.

Our free guide breaks down the 6 investable robotics subsectors driving the buildout, from AI compute and machine vision to humanoid robots. Enter your email and we’ll send it straight to your inbox.

Get the Free Robotics Guide

🤖 AI MADE THIS

Interesting and inspiring creations made with AI

A Robot Octopus Built From One Picture

Fokki gave Claude Opus 5.5 a single picture of a mechanical octopus and asked it to build the model. The AI wrote the 3D code for a robot with eight intricate arms that opens, falls apart, and reassembles itself.

AI used
Claude Opus 5.5

⚡ QUICK HITS

The latest in AI, tech, and productivity worth knowing

⚒ TOOL SNAPSHOTS

Futuristic tools within AI, no-code, and productivity

ℹ️ ABOUT US

The Intelligent Worker helps you to be more productive at work with AI, automation, no-code, and other technologies.

We like real, practical, and tangible use-cases and hate hand-wavy, theoretical, and abstract concepts that don’t drive real-world outcomes.

Our mission is to empower individuals, boost their productivity, and future-proof their careers.

We read all your comments - please provide your feedback!