AI with Kyle Daily Update 160

Andrej Karpathy - co-creator of ChatGPT - scored 342 occupations from 0 to 10 on AI exposure. He published the whole thing on GitHub. It went viral. Elon replied. News outlets picked it up.

Then he deleted it within 10 hours…

Fortunately it had been cloned and reposted. I found the tool a little hard to search. Well, impossible because there’s no search feature.

So made you a searchable version:

Here’s the link:

The average AI exposure across the entire US economy? 5.3 out of 10. Gulp.

$3.7 trillion in annual wages sits in high-exposure occupations.

Some specific scores:

• Software developers: 9/10 (median pay $130k, 1.9M jobs)

• Customer service reps: 9/10

• Travel agents: 9/10

• Lawyers: 8/10

• Secretaries: 8/10

• Network administrators: 8/10

• High school teachers: 7/10

• Physicians: 5/10

• Heavy truck drivers: 5/10

• Delivery drivers: 3/10

• Electricians: 2/10

• Plumbers: 2/10

• Construction labourers: 1/10

Highly recommend you go find your score here AI jobs exposure and think:

• does this look right? (it may not be!)

• if it does look right what does this mean?

• what steps are you personally taking to combat this?

Kyle's take: Three things keep you safe - physicality, unpredictability, and human contact. If your job is primarily sitting at a computer moving data, text, and information around, your job is massively at risk. If you're physically operating in the real world, doing unpredictable work, and interacting with real humans - you're in a much better spot.

My advice? Don't quit your job. Hold onto it as long as possible. Jobs may become something of a rarity this decade… I fear that a couple of years from now we'll look back and think "remember when I had that nine-to-five and they just paid me every month?"

But start something on the side for optionality. Diversify your income.

A Man Used ChatGPT to Create a Cancer Vaccine for His Dog. It's Working.

This one hit different.

Paul Conyngham - a bloke in Australia with a computer science background, NOT a medical professional - used ChatGPT and Google DeepMind's AlphaFold to create a personalised mRNA cancer vaccine for his dog Rosie.

And…it worked.

The process: he asked ChatGPT how to go about the task. Then based on ChatGPT’s advice he got Rosie's DNA sequenced (healthy blood DNA plus tumour DNA then find the difference) for $3,000 at the University of New South Wales. Compared them to find the exact mutations. Ran the data through AlphaFold to identify mutated proteins and potential drug targets.

Getting ethics approval took longer than creating the vaccine. Three months, two hours of work every night, typing up a 100-page document. Basically he has to get the go-ahead to run a one-dog drug trial. Thankfully the University of Queensland School of Veterinary Science eventually provided the coverage.

The result? Rosie had a tennis-ball-sized tumour. It's shrunk by half. She's had one injection plus a booster, with another due next week. A researcher at UNSW (very professionally!) summed it up as : "I was like, holy crap. It worked."

Kyle's take: This is NOT just a cute dog story. This is the dawn of personalised medicine. One person, with no medical training, using AI tools that exist right now, created a tailored mRNA vaccine for an individual patient's specific cancer.

If we can do this for a dog, the question is obvious - why aren't we rolling this out for humans? When people tell you AI is just going to destroy the world, point them to this. The cost of sequencing DNA keeps dropping. The AI tools keep improving. Within a few years, personalised cancer vaccines could be routine. That's worth getting excited about.

If you want to understand the AI behind this, watch "The Thinking Game" on YouTube (free) - it's a documentary about Google DeepMind and AlphaFold (the tool used to solve protein folding). Genuinely brilliant film.

Source: The Australian (no-paywall)

Meta Is Falling Out of the AI Race

Meta has delayed their new frontier model "Avocado" until at least May. It underperformed on internal evaluations. Per the New York Times, Meta is now considering licensing Google Gemini as a temporary solution. A little embarrassing…

Avocado outperformed Meta's previous Llama models but hasn't matched Gemini 3 from November 2025. Not good enough at reasoning, coding, or writing. This despite Zuckerberg spending $100M+ on compensation packages to poach engineers. Meta hasn't released anything competitive for roughly a year…

Kyle's take: They're out. We're down from five Western AI labs to effectively three - OpenAI, Anthropic, and Google. The Chinese open-source models have overtaken Meta's too. Every new model release has to beat ALL current top models or it's a collective shrug. Meta can't clear that bar.

Here's the thing to watch though. Google is quietly winning by being the default AI everywhere. Gemini coming to iPhone through Siri. Gemini on Android (Google owns it). Gemini in Chrome (70% browser market share). Now potentially licensing to Meta. Same playbook as Google Search becoming the default 20 years ago. Slow, silent, everywhere.

Is Gemini the best model? No not really. Does that matter?

Great distribution often beats great product. I shall point you to Exhibit A:

Claude Skills - 88,000 Experts At Your Fingertips

I showed skills in action last week. Today I'm giving you the resource links so you can actually go get them.

Quick recap: a skill is a folder with a markdown file that gives Claude domain expertise on demand. Like Neo downloading kung fu in The Matrix. They only load when needed - unlike MCP tools which sit in memory permanently.

I’ve put together a resource page for you with my recommendations: Claude Skills Recommendations.

It includes:

1. skills.sh - Directory of 88,000 skills with popularity ranking. Top skill: "Find Skills" - a meta-skill that helps you discover and install others from the whole library. Do NOT download all 88k. Pull in what you need.

2. Official Anthropic repo - The ones Anthropic themselves built. PDF creation, PowerPoint, etc.

3. Awesome Claude Skills (my #2 pick after official) - Word docs, Excel, generative art, canvas design, GIF creator, frontend design, MCP builder, web testing, brand guidelines, skill creator, iOS simulator, Playwright automation, social media images.

4. Curiosity Tech's Claude Skills - Quality collection of productivity skills.

5. Alireza Rezavani's repo - 192 skills and agent plugins including a full marketing team bundle: SEO, analytics, app store optimisation, brand guidelines, competitor analysis, cold email, churn prevention.

All here: Claude Skills Recommendations.

I'll do a full walkthrough later this week showing these in action. And I’ll keep updating the resource.

Want the full unfiltered discussion? Join me tomorrow for the daily AI news livestream where we dig into the stories and you can ask questions directly.

Until then, keep PROMPTING!

Kyle