In partnership with

Welcome, humans.

Every computer use AI model on the market right now learned from screenshots. Static images. Frozen moments with no sense of what happened before or after.

Standard Intelligence just changed that. They released FDM-1, the first AI model trained on 11 million hours of screen recordings, learning from video the way a human would: watching what happens over time.

The results are wild. It can construct 3D gears in Blender, find software bugs by exploring apps like a human tester, and (our favorite) drive a real car through San Francisco using arrow keys after less than one hour of training data.

The secret sauce is a video encoder that fits nearly two hours of 30fps video into 1 million tokens, roughly 50x more efficient than existing state of the art. That's the difference between an AI that can see six seconds of context and one that can follow a two-hour workflow. (Research-only for now; no public access yet.)

Here’s what happened in AI today:

• Anthropic caught three Chinese AI labs running industrial-scale operations to steal Claude's capabilities.

• OpenAI formed multi-year enterprise partnerships with McKinsey, BCG, Accenture, and Capgemini.

• Standard Intelligence trained an AI model on 11M hours of video that can drive a car with 1 hour of data.

• Two viral economic essays modeled what happens if AI displaces workers faster than institutions can adapt.Advertise in The Neuron here!

Heads up: This week’s new episode of our podcast The Neuron: AI Explained will air on Wednesday after 2pm PT this week!

Are we living through AI’s “Steroids” Era?

The 2026 Winter Olympics officially ended this weekend, and we gotta say, it was great watching all that human achievement. Unlike previous Games, there weren't many doping scandals (though there was the infamous, definitely not safe for work ski jumping scandal).

The biggest controversy this year was actually in curling, where Sweden accused Canada of cheating. Profanities flew, secret videos were aired, and emergency rule changes followed. It's a reminder: wherever there's competition, someone's looking for a shortcut.

Sound familiar? Because that's basically what just happened in AI.

Anthropic revealed today that three Chinese AI labs, DeepSeek, Moonshot AI, and MiniMax, allegedly ran industrial-scale campaigns to extract Claude's capabilities and train their own models. The scope: 24,000+ fake accounts generating over 16M exchanges with Claude.

The technique is called "distillation," where you train a weaker model on a stronger model's outputs. This is legit when labs do it to their own models, but less so when you create thousands of fake accounts to do it to a competitor's.

Here's what each lab targeted:

• DeepSeek (~150K exchanges) asked Claude to reverse-engineer its own reasoning step by step, generating chain-of-thought training data on demand.

  • They also had Claude create "censorship-safe" alternatives to politically sensitive queries.

• Moonshot AI (~3.4M exchanges) targeted agentic reasoning, coding, and computer vision across hundreds of fake accounts.

• MiniMax (~13M exchanges) ran the biggest operation. When Anthropic dropped a new model mid-campaign, MiniMax pivoted within 24 hours to start extracting from it.

And all three used proxy services to circumvent the fact that Anthropic doesn't offer commercial access in China.

Needless to say, the internet had thoughts. Critics noted American AI companies trained their models on the entire internet without asking permission, too.

• Andrew Curran joked this is "actually how a Claude reproduces, so congratulations, a blessing upon your family."

• Anthropic researcher Alek Dimitriev was blunt: "Leapfrogging happens through innovation, not distillation."

• Researcher Will Brown raised the harder questions:

  • If a developer uses Claude Code to write a function and commits it to a public MIT-licensed repo, does that count as distillation?

  • Is it against TOS to share Claude outputs on the open internet?

  • Are labs obligated to filter them out of training data?

• Sidra Miconi, PhD called this the most important question in the thread.

  • The line between "illicit distillation" and "normal internet activity that happens to include Claude outputs" is genuinely unclear.

  • "Intent and scale matter," she wrote. "But the gray zone between those two extremes is massive and completely unregulated." Meaning this is probablygoing to end up in court.

To Anthropic’s credit, there’s a national security angle to this, too: distilled models can have safety guardrails stripped, potentially feeding unprotected capabilities into military and surveillance systems.

The timing is spicy. DeepSeek is reportedly about to launch V4, a model CNBC says could rattle the Nasdaq. As if the ‘daq needs any more rattlin’. Hey, at least Anthropic didn’t launch another plugin. Oh wait… (more below).

Basically, we’re living in the steroids era of AI, where your models can get jacked quick if you’ve got the right supply of Anthropic traces… so we’ll see if DeepSeek V4 comes out looking like GigaXi on Wednesday, when we presume it’ll drop (right on time for NVIDIA earnings!).

Why did the market sell off yesterday anyway? While this can’t explain all of it, two economic essays went viral for exploring what happens when AI works too well:

1. Citrini Research modeled a fictional "2028 crisis" where AI displaces white-collar workers so fast that consumer spending collapses and the S&P crashes 38%; this one hit with the financial crowd.

2. Economist Alex Imas asked whether AI could actually shrink GDP through demand collapse (his answer: probably not, but "disappointing growth" is on the table without policy solutions like sovereign wealth funds).

Whether the AI is homegrown or hot-wired, the question stays the same: the value of human intelligence is getting repriced… so what’s the new number?

FROM OUR PARTNERS

Free email without sacrificing your privacy

Gmail is free, but you pay with your data. Proton Mail is different.

We don’t scan your messages. We don’t sell your behavior. We don’t follow you across the internet.

Proton Mail gives you full-featured, private email without surveillance or creepy profiling. It’s email that respects your time, your attention, and your boundaries.

Email doesn’t have to cost your privacy.

Ditch the Gmail data grab

Prompt Tip of the Day

Anthropic just published their AI Fluency Index, studying how ~10K people actually use Claude. The biggest finding: people who iterate get 2x more value. Users who treated AI's first response as a starting point showed double the rate of productive behaviors like questioning reasoning and identifying gaps.

The catch: when AI produces polished outputs (code, documents, apps), users get less critical. They're 5.2 percentage points (pp) less likely to spot missing context and 3.1pp less likely to question reasoning. The shinier it looks, the less we inspect it.

Two tips:

1. Set the terms upfront. Only 30% of users tell AI how to interact with them. Try: "Push back if my assumptions are wrong" or "Tell me what you're uncertain about."

2. Use the "one more question" prompt. Developer Jeffrey Emanuel recommends this after you think you're done:

"What's the single smartest and most radically innovative addition you could make to this project at this point?"

Try it with a few different models. You'll be surprised.

Bottom line: The best AI users don't write the fanciest prompts. They don't stop at the first answer.

Want more tips like this? Check out our Prompt Tip of the Day Digest for February.

Treats to Try

1. Want OpenClaw, but actually safe? IronClaw lets you deploy AI agents with built-in security and privacy guardrails, from the team at NEAR AI (open-source, code on GitHub).

2. Guide Labs built an open-source language model called Steerling-8B that lets you see exactly why it generated each response — it breaks its reasoning into human-readable concepts you can inspect and adjust, like a glass-box version of an LLM (code).

3. Daytona gives your coding agents a safe place to run code — isolated sandboxes that spin up in under 90ms, so nothing touches your real infrastructure—$200 free compute, then pay-as-you-go.

4. Reve generates images from text prompts — type "a cinematic sunset over Tokyo" and get a photorealistic 4K image in seconds — and its latest v1.5 model already ranks in the top three on Arena.ai.

5. Siteline analyzes AI agent and bot traffic to optimize content for visibility, with recommendations on prompts, rankings, and citations against competitors—free to try.

Around the Horn

1. OpenAI formed "Frontier Alliances" with McKinsey, BCG, Accenture, and Capgemini in multi-year partnerships to deploy AI "coworkers" across enterprise clients.

2. IBM shares tanked 13% after Anthropic announced Claude can streamline COBOL, threatening legacy mainframe business

3. US Secretary of War Pete Hegseth summoned Anthropic CEO Dario Amodei to the Pentagon with ultimatum on Claude's military usage restrictions.

4. Goldman Sachs said AI added basically zero to US GDP in 2025 (video); separately, ~90% of firms report zero productivity gains from AI despite $252B in global spending.

5. An NBER study found AI closes 75% of education productivity gaps during use, but gains vanish when the tool is removed.

6. Anthropic published The Persona Selection Model, a research paper arguing AI assistants behave like human-like characters the model learned to "play" during training, not alien intelligences with hidden goals. It's less Terminator, more community theater; if you missed it, this was proposed by Scott Alexander last year and we did a whole podcast on this. 

Want everything that happened in AI this week? Read Around the Horn Digest here.

FROM OUR PARTNERS

Gartner® report: Redesign Your Workflows for Strategic Adoption of AI Agents

Is your team deploying AI agents, but not seeing value? Learn when agentic workflows are justified and how to redesign processes for context-driven automation. Access a five-step framework. Enjoy complimentary access from Weights & Biases for a limited time.

Access now

Tuesday Tweet

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.