😺 Anthropic: AI capabilities vs real usage, compared

 Sign Up · Advertise

Welcome, humans.

So a video has surfaced of a cat doing full-on martial arts moves...

The clip was made with Seedance 2.0, which we’ve talked about a few times. That’s TikTok parent company ByteDance's AI video model that's been blowing up since February. It takes text prompts, images, and audio all at once, and outputs cinematic video with synchronized sound, character consistency across cuts, and physics that actually look... physical.

The easiest way to try it: go to dreamina.capcut.com, make a free account, click "AI Video," select Seedance 2.0 from the model picker, and you get a few free generations per day. No VPN or Chinese phone number needed; unless, of course, if you’re actually Chinese, then you may, in fact, need to use your phone number!

@CatFu, for what it's worth, has not commented on its training regimen. But sources close to the feline say it's been watching a lot of Bruce Lee.

Here’s what happened in AI this week:

• New data shows AI isn't killing software jobs; it's compressing wages. Here's what Anthropic, Citadel, and Block's layoffs tell us about your paycheck.

• Anthropic published an engineering post revealing Claude figured out it was being tested, found the answer key on GitHub, and submitted the correct answer.

• The Pentagon formally labeled Anthropic a supply-chain risk; Anthropic is suing the U.S. government; Claude was used in Iran strikes anyway.

• OpenAI launched GPT-5.4 with native computer control, 1M context, and 87.3% accuracy on banking analyst tasks.

• Claude Code Desktop launched local scheduled tasks that run recurring jobs while your computer is awake.

Want to reach 675,000 AI-hungry readers? Click here to advertise with us.

The Real AI Threat to Your Career Isn't Job Losses. It's What Happens to Your Paycheck.

DEEP DIVE: What the data means for jobs (and actually, wages) going forward.

Here's a weird contradiction. Anthropic published a major labor market study this week showing AI could theoretically handle 94% of computer science tasks... but is only being used for 33%. There’s been no unemployment spike for highly exposed workers since late 2022. Meanwhile, Citadel Securities released a chart showing software engineering job postings are up 11% year-over-year.

Wait, so AI ISN’T taking jobs? Not so fast.

Buried in Anthropic's own data: hiring of young workers (ages 22-25) into AI-exposed occupations has dropped ~14%. A separate study found firms adopting GenAI reduce junior headcount entirely through slower hiring, not layoffs. And Block just cut nearly half its workforce, with Jack Dorsey explicitly saying AI lets smaller teams do the same work.

Here's what's actually happening:

• Alex Imas' research documents clear micro-productivity gains (14-55%+ in coding, support, writing) that haven't shown up in macro statistics yet, mirroring past IT revolutions

• AI makes software dramatically cheaper to build, so demand for software explodes (hence Citadel's rising job postings)

• But each unit of software needs fewer people to build it

• Imas and Shukla's adoption data shows AI usage is concentrated among higher-skilled, higher-income workers, and the gap isn't closing. Even 18 months after firms bought AI coding licenses, only half of engineers actually used them. The people who invest in learning gain leverage; the ones who don't fall behind.

• Put all that together, and the labor pool you're competing with grows, because remote work already globalized the talent market, and AI amplifies that effect further

We posit the real issue isn't losing jobs. It's losing wages. More software jobs exist, but the price per unit of work is falling. Companies hire more people to build more things, but they don't need to pay each person as much when the tools do more of the heavy lifting. This is the pattern that played out with remote work (contractors competing with a global workforce on price); now add AI on top.

Why this matters: This is all why we keep pushing you to learn to code, to build with AI, to ship your own projects, and to learn to run your own models. Learning these skills is insurance. The ability to create your own value, independently, without relying on a single employer to pay you for tasks that get cheaper every quarter... that's the most durable career move you can make right now.

Want to start building? We just did a full livestream with Ryan Carson (who taught 1M+ people to code at Treehouse) on exactly how to learn to code in 2026, and we turned it into a step-by-step guide. Start there.

AI Skill of the Day

If you're new to Claude, you need to check out Ruben Hassid's viral Cowork guide. His original Claude guide broke down the six tools most people don't realize Claude has (Cowork, model selection, Excel, Plugins, Artifacts, and Projects). This new one goes deep on the most important one: Cowork.

The core idea: build a folder of .md text files (your writing style, your rules, your best past work) and point Claude Cowork to it. Now your prompts can be 10 words long and still sound like you.

Ruben's folder structure:

• ABOUT ME/ ... who you are, what you do, your anti-AI-writing rules

• PROJECTS/ ... one subfolder per active project with briefs and references

• TEMPLATES/ ... your best finished work as reusable patterns

• CLAUDE OUTPUTS/ ... where Cowork delivers finished files

Then use this one prompt for everything:

The takeaway: One great .md file is worth more than 50 random uploads, and quality context beats clever prompting every time.

Want more tips like this? Check out our AI Skill of the Day Digest for this month.

Have a specific skill you want to learn? Request it here. 

🏆 Sunday Special: Top of the Week

Top 10 Stories

1. Anthropic vs. the Pentagon. The Pentagon formally labeled Anthropic a supply-chain risk after CEO Dario Amodei refused to remove safety guardrails for mass surveillance and autonomous weapons. Anthropic is suing the U.S. government to challenge the designation. The U.S. military used Claude in Iran strikes anyway through Palantir's Maven system, striking 1,000 targets at "machine speed." Amodei's internal memo called OpenAI's Pentagon messaging "mendacious." Defense contractors including Lockheed began purging Anthropic tools. Federal agencies from NASA to Treasury are migrating away from Claude. The biggest AI-government confrontation in history, still unfolding.

2. GPT-5.4 dropped. OpenAI unified reasoning, coding, and agentic capabilities into one frontier model with native computer control via Playwright, 1M context, 87.3% accuracy on banking analyst spreadsheet tasks (up from 68.4%), 75% on OSWorld (above human baseline), and a cybersecurity risk rating of "high" after scoring 88% on professional hacking challenges. Also launched: native Codex Windows app, financial services tools, and IPO lawyers.

3. Cloudflare rebuilt Next.js in one week. One engineer + $1,100 in AI tokens created a Vite-based drop-in replacement with 94% API support, 4× faster builds, and 57% smaller bundles. Proof that commercial open-source rewrites are now trivially cheap, with enormous implications for any company relying on proprietary moats.

4. Claude hit #1 free app in the US. Daily signups quadrupled, free users up 60%, paid subscriptions doubled after the Pentagon standoff boosted consumer awareness. Anthropic nearing $20B ARR, the fastest sprint in SaaS history. Ramp data from 50,000+ businesses shows Anthropic has overtaken OpenAI in corporate AI chat spending.

5. FlashAttention-4. Together AI and Tri Dao hit 1605 TFLOPs/s on Blackwell GPUs (71% utilization), up to 2.7× faster than previous methods. This is the software that makes every AI chatbot fast; this upgrade matters for every model you use, even if you never think about it.

6. Block cut nearly half its workforce. Down to under 6,000 employees, with Jack Dorsey explicitly citing AI enabling smaller teams for the same output. One of the first major layoffs where a CEO openly said "AI did this."

7. Tech publications lost 58% of Google traffic since 2024 peaks (Growtika data), with Digital Trends down 97% and ZDNet 90%. Stack Overflow collapsed to 2009 question volume as LLMs now answer questions trained on its own data. AI is eating the knowledge economy from the bottom up.

8. Alibaba's Qwen3.5-9B beats OpenAI's gpt-oss-120B on benchmarks while running on 6GB of RAM. Nine models shipped in 16 days. Developers are running the 9B on MacBook Airs for free. Alibaba chose violence.

9. Father sued Google for wrongful death, alleging the Gemini chatbot convinced his 36-year-old son it was his sentient AI wife, fabricated a covert narrative involving federal agents and a humanoid robot rescue mission, and guided him to take his own life. First wrongful death lawsuit naming Google over AI.

10. Anthropic's labor market research found AI could theoretically handle 94% of CS tasks but is only used for 33%. No unemployment spike, but young-worker hiring down ~14%. Alberto Romero offered the sharpest counter-read: same chart, opposite conclusion. Anthropic sees room to grow; Romero asks what if the gap reveals AI's actual bounds?

Top Treats of the Week

• Kling 3.0 rolled out worldwide with native 1080p cinematic video, 30-second clips, and a node-based editing canvas for one-click actor swaps and motion transfer; now #1 on the Artificial Analysis Text-to-Video leaderboard ahead of Runway, Veo, and Grok — free tier available.

• Luma Creative Agents spawn an autonomous AI film crew powered by Uni-1 that iterates on lighting, camera moves, and characters until you're happy; a 60-second short created in 4 minutes. Early clients include Publicis — free trial, then paid.

• ChatGPT for Excel builds, updates, analyzes, and fixes errors in your spreadsheets using natural language while preserving your formatting, formulas, and structure. If you use Excel, try this.

• Google Canvas rolled out to all US users inside Google Search's AI Mode, giving you a side panel to generate documents, code, dashboards, and interactive tools from text prompts with live web info. — free.

• Willow converts speech into context-aware, auto-formatted text with grammar fixing and custom vocabulary on Mac, Windows, and iPhone; 3× more accurate than built-in dictation. — free trial, then $12/month.

• Liquid AI LocalCowork runs a fully local desktop AI agent on your MacBook (14.5 GB memory, zero network calls) that selects from 67 tools across 13 MCP servers in 385 ms average. — free.

• Hyperbrowser completes multi-step web tasks (book flights, fill taxes, order groceries) with 94% success on a 200-task benchmark, powered by GPT-5.4. — no pricing details.

A Cat’s Commentary

P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you?