😸 Plot twist: AI is making you work more

Welcome, humans.

So ByteDance (the TikTok parent company) just dropped Seedance 2.0, and the AI video landscape basically changed overnight:

We're talking native audio generation, lipsynced speech, 2K resolution, and quality that makes Google's Veo 3.1 and OpenAI's Sora 2 look like they're running on a flip phone. One creator made a full 2-minute cinematic fight scene for $60. Another generated a photorealistic LeBron James dunk that's genuinely hard to tell from broadcast footage. Chinese filmmakers are already producing entire short films from scripts, complete with VFX, voice acting, and editing.

Check out this 1960s Kung fu DragonBall Z match-up and this prompt-to-mini-film demo. We watched these three times each. It’s almost unbelievable…

The catch? It's currently only available in China. But if ByteDance's track record with TikTok is any indication, it won't stay that way for long… (rumors are it’ll be available internationally on Feb 24).

Here’s what happened in AI today:

• Research shows AI tools don't reduce your workload; they make it bigger.

• OpenAI started testing ads in ChatGPT for free and Go tier users.

• Meta AI added browser automation and multi-model agent features.

• An ad for OpenAI’s screenless AI device leaked over the weekend.

New Research: AI Doesn't Save You Time. It Just Makes You Do More.

You know that feeling when you finish a task with AI in 10 minutes that used to take an hour... and then immediately start three new things because now you "have time"?

Yeah, that's not a productivity win. That's a trap.

New research from Harvard Business Review studied how generative AI changed work habits at a ~200-person tech company over eight months. The finding? AI tools didn't reduce workloads. They consistently intensified them.

Here's what happened:

• People took on work that wasn't theirs. Product managers started writing code. Researchers attempted engineering tasks. AI made everything feel doable, so people just... did more.

• Breaks disappeared. Workers started prompting AI during lunch, between meetings, even right before leaving their desks ("let me send one last prompt so it can work while I'm gone"). Work became, as the researchers put it, "ambient"; always there, always possible, always tempting.

• Multitasking exploded. Employees ran multiple AI agents in parallel, revived old tasks they'd been putting off, and juggled more threads than ever. They felt productive. They also felt busier than before AI.

One engineer summed it up perfectly:

Honestly? We felt this one personally. AI doesn't feel like work. It feels like play. You're not "writing a report," you're "chatting with Claude." You're not "doing research," you're "exploring an idea." That friction that used to force you to take a break (waiting for a colleague, staring at a blank page, not knowing where to start) is gone. And without that friction... you just never stop.

So what's the fix? Not "set better boundaries." That's the email advice we ignored 15 years ago.

The better answer is compound engineering, a concept from Kieran Klaassen at Every. The core idea: instead of using AI to do more work, use it to build systems where every piece of work makes the next one easier. Your codebase (or workflow) gets simpler over time, not more complex.

Here are the key principles worth stealing, even if you're not an engineer:

• Spend 80% planning, 20% building. Most people do the opposite. But with AI, "plans are the new code"; if the plan is good, the output will be good. Brainstorm and plan thoroughly, then let AI execute.

• The 50/50 rule. Spend half your time on the actual task and half on improving the system itself (templates, prompts, reusable patterns). This is the compounding part; you're investing in making tomorrow's work faster.

• Build safety nets, not checkpoints. Instead of reviewing every line of AI output, build monitoring and tests that catch problems automatically. Trust the process, then verify at the PR level (for non-coders, that would be the final output before you "publish" the work)

• Let go of "code is self-expression." This applies beyond coding. If your identity is tied to doing the work rather than directing it, you'll resist the shift. The new skill is taste, not typing.

As for the HBR researchers, they recommend "intentional pauses" and "sequencing." That's fine. But compound engineering goes further: it turns the AI treadmill into a flywheel. Instead of doing more things at the same speed, you do the same things faster each time, which is what actually frees up real time.

Watch the full breakdown here (13 min). Even if you're not a developer, the philosophy applies to any AI-assisted workflow.

FROM OUR PARTNERS

Deep analysis for deep questions

Stop talking to walls of predictive text and start doing real research with Superagent. Give it a question, and it gets to work: Subagents deeply interrogate your topic — scouring a wide range of credible sources and using those findings to inform additional research directions — and finally package it all up into boardroom-ready reports, slides, docs, or websites.

Sign up for free

Prompt Tip of the Day

Ever notice ChatGPT or Claude getting... dumber halfway through a long conversation? Repeating itself, forgetting what you said earlier, giving worse answers?

There's a name for it: the "dumb zone." Once an AI uses roughly half of its context window (its working memory), quality starts tanking. Think of a student with 5 minutes left on an exam and half the questions still blank; it starts guessing instead of thinking.

Here's a trick from Kelvin French-Owen, who helped build OpenAI's Codex: plant a canary at the start of your conversation.

Drop in a random fact early on:

When answers start feeling off, ask: "Hey, what's my dog's name?"

If it forgets? Your context is cooked. Start a new chat and paste in just the key details from the old one. You'll immediately get better results.

Pro tip: For important work, keep conversations short and focused. One topic per chat. Your AI will stay sharper, and you'll spend less time cleaning up bad output.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 600K readers here!

1. *Outskill is hosting a 2-Day AI-Mastermind that will teach 20+ AI tools, AI workflows, build agents & more in just 16 hours. Become an AI Pro now! (usually $395, free for you).

2. Claude with Ads is a satirical spoof (just joke tool, not real Anthropic product) that roasts OpenAI's ad plans by forcing you to sit through full product pitches before answering your questions. Try it! It’s hilarious.

3. Claude Code's fast mode lets you toggle faster Opus 4.6 responses by typing /fast when you need speed for rapid iteration or debugging. It's in research preview, so features may change.

4. Shipper creates full Chrome extensions from a single prompt using Claude, handling browser compatibility, privacy policies, and Web Store autofill for about $0.11 per extension

5. Interface Craft teaches you interface design with tools via lessons and demos, providing frameworks, practical demonstrations, AI collaboration lessons like using Claude Code and v0, and interactive playgrounds for building interfaces ($77 lifetime founding membership).

6. Ricursive Intelligence automates chip design so AI creates the chips that power the next generation of AI, compressing years-long design timelines into weeks (raised $300M).

7. WordPress.com launched the first official Claude connector for any WordPress host, letting you securely connect Claude to your site for read-only access to analyze traffic, comments, and content performance through natural language prompts.

Around the Horn

1. OpenAI's Codex app hit 1M downloads and 1M weekly active users within a week of launch; OpenAI just hosted a hackathon where winners built OpenCortex (AI research paper generator), Envy (auto-tool builder), and Paradigm (AI coding assistant).

2. An ad for Jony Ive's screenless AI device leaked over the weekend; the device, built in partnership with OpenAI, ditches screens entirely for a voice-first design that could launch later this year.

3. Anthropic is closing a $20B funding round at a $350B valuation (doubling its initial target), with Nvidia and Microsoft contributing $15B combined; the company's revenue run rate now exceeds $9B.

4. Meta AI added Manus-style agent features including browser automation, multi-model fusion, and proactive task management

5. A researcher at Anthropic used 16 parallel Claude agents to autonomously build a complete C compiler from scratch that compiles the Linux kernel; the agents coordinated using Git and plain text files, no human oversight needed, completing in two weeks what normally takes teams years.

6. Anissa Gardizy of The Information writes Anthropic is pursuing at least 10GW (gigawatts; like ten nuclear power plants worth) of data center capacity and hired former Google and Stack Infrastructure executives to help build out its own infrastructure, a move that could cost hundreds of billions of dollars and signals plans to move beyond rented cloud space to control its own fleet.

7. Microsoft AI Red Team discovered that a single unlabeled prompt can break safety alignment across 15 language models using a technique called GRP-Obliteration that inverts standard fine-tuning methods.

8. Anthropic cofounder Daniela Amodei said humanities skills will become "more crucial than ever" as AI reshapes work, prioritizing emotional intelligence over technical backgrounds.

9. OpenClaw patched security flaws after Chinese government warnings that the open-source AI agent's vulnerabilities could expose systems to cyberattacks and data breaches.

10. Watch: Harper Carroll explain how AI is able to “see” images.

Tuesday Tool Tip

Most people use AI for writing. But one of the most underrated workflows we've seen is using it for thinking.

Here's a setup from @Ashwinreads that turns Obsidian (the free note-taking app) into a research engine using Claude Code and a voice transcription tool called Monologue:

The workflow:

Populate a markdown folder in Obsidian with your source material (reports, articles, notes, whatever you're studying). Then use Monologue to stream-of-consciousness talk through what you're reading; basically asking "what's the idea I'm saying here?" Claude Code processes the folder and helps pull threads together, linking related ideas across documents.

Why this works: Most of us collect information but never connect it. We bookmark 40 articles, highlight a bunch of stuff, and forget all of it by Thursday. This workflow forces synthesis in real-time; you're talking through the material, and Claude is organizing it as you go.

How to try it:

1. Create a folder in Obsidian for a topic you're researching (could be anything: an industry trend, a work project, a side hustle idea).

2. Drop in your raw notes, saved articles, or bullet points as markdown files.

3. Open Claude Code and point it at the folder. Use Monologue (or any voice transcription app) to talk through your reading, then let Claude organize your thoughts and link ideas.

4. Ask Claude to generate recall questions: "Quiz me on the key takeaways from this week's research."

The magic is that your notes get smarter over time. Claude adds links between ideas, surfaces connections you missed, and builds a knowledge base that compounds. Sound familiar?

It's basically compound engineering applied to your brain instead of your codebase.

Tomorrow we’ll talk about some of the other agentic tools and systems you can use to implement these types of practices at work.

A Cat’s Commentary