Unsupervised Learning NO. 517

Hey! Hope you all are doing well out there!

—

One of my main thoughts right now is that people are focusing way too much on AI model advancement. Yes, they’re getting really good. But that’s not where the impact to the enterprise (and jobs) will come from.

Let’s say, theoretically, we’re now at 80% of where our AI model progress is capable of (it’s not). And let’s say we’re increasing really fast, at like 5% a year (whatever).

That’s not the number to watch. The real question is:

Those numbers are tiny. Businesses are JUST STARTING all of this. And they aren’t even sure yet what the maturity levels are!

So we’re at like (roughly):

Those are obviously made-up/estimate numbers, but you get the point. It’s barely started. The A.I. could be frozen at today's levels and it wouldn’t matter.

The impact is going to come from the integration number, not the AI getting better and better (although that will help, and especially if it hockey-sticks again)

Anyway, just thought that was a useful perspective.

—

I like this mental model for thinking about getting ready for the change that is coming.

The number one question I get asked is: What do I study? What do I learn? What do I tell my kids? What do we tell young people? So here are my three:

READ THE POST

I have another framework which is.

These go together well. The first set is more like things to do, and the second set is more like things you should have. But in both cases it's a matter of constant effort and nurturing. This will be a big focus of mine in the Human 3.0 courses. Like, what do you actually do? What do you actually read? What do you actually study? Etc.

—

I think General Verifiability is the biggest thing in AI that nobody is talking about yet. I'm expecting this to catch fire as a concept within the next few weeks or months.

The idea is that everything is about verification. If you can't verify whether or not you are improving, you can't make progress. And the most important component of this is being able to articulate an ideal state.

But how do you capture an ideal state for a painting, or for a grocery list, or for a career, or for a relationship? Extremely difficult.

But if you can do it even a little bit, and then climb towards that, you will have made a whole lot of progress that AI is currently not able to do because it has nothing to hill climb towards. GENERAL HILL-CLIMBING

This is the entire premise of The Algorithm in our PAI project.

China leads cyber espionage volume against defense contractors, Russian groups target Ukraine drones, North Korean IT workers infiltrate firms

Another brilliant report by Google's Threat Intelligence Group.

Every time I read a report like this, I get really frustrated that we are not focusing on the right things. The country should be building an automation system that threat models all of this continuously. GOOGLE THREAT INTELLIGENCE REPORT | APT44 SANDWORM REPORT | MANDIANT M-TRENDS 2024 | APT1 REPORT

DEFCON bans three hackers tied to Epstein TECHCRUNCH ARTICLE

Palo Alto acquires Koi Security to protect AI agents on endpoints CYBERSECURITY NEWS ARTICLE

Cline AI tool supply chain attack installed malware on 4,000 machines I really think it's time to start signing our AI harnesses. Skills need digital signatures. Same with MCPs and other ingresses into our AI workflows. STEPSECURITY REPORT

DARPA's missile-launching drone nears flight tests this year. DARPA LONGSHOT UPDATE

China will keep funding Russia's Ukraine war because it weakens everybody. POWER CHANGES ARTICLE

Canadians now see America as a bigger threat than Russia Nearly half of Canadians view the U.S. as the biggest threat to global peace, ahead of Russia, and most would rather depend on China than the current administration’s America. POLITICO POLL ARTICLE | POLITICO POLL RESULTS

Claude Code autonomous runs nearly doubled in three months, now up to 45 minutes ANTHROPIC RESEARCH

AI-generated apps are about to murder subscription pricing AI GOING TO KILL APP SUBSCRIPTIONS

AI progress is accelerating faster than ever, with costs dropping 300x in a year In an essay, Tomas Pueyo breaks down how frontier models have jumped from below 20% to 55% on the hardest reasoning tests in one year. TOMAS PUEYO'S ARTICLE | TOMAS'S COMPUTE ARTICLE | TOMAS'S AGI ARTICLE | ARC-AGI LEADERBOARD | NVIDIA RUBIN ANNOUNCEMENT | MILLION STEP TASK PAPER

Indians use AI deepfakes to bring dead relatives to weddings I find this really endearing but also nightmare fuel if done poorly. but I guess that applies to a whole lot of everything I see these days. And especially to anything AI-related. REST OF WORLD ARTICLE

Letter to a Young Person Worrying About AI Basically, an argument that things take a very long time to change, so don't worry too much about it. Focus on the thing right in front of you.. Not a perspective I generally hold to at this particular moment, but I like presenting alternative ideas. CEDRIC'S LETTER

The left thinks AI is fake autocomplete while missing the revolution I hadn't thought about it this way, but it really is turning into something quite partisan. It's not quite perfect though, because the working class and middle America is also relatively anti-AI, and they tend to be more right/conservative. But I would say there is some truth to this. THE LEFT IS MISSING OUT ON AI

Custom Everything = More Separation Between All of Us
I think everyone's about to have custom software: custom interfaces, custom media, custom books, basically completely customized interfaces to the world. I don't see how this doesn't result in more fragmentation, which is something that Robert Putnam wrote a lot about in books like Bowling Alone. I think this is problematic for multiple reasons. As a country, it's problematic because we no longer believe that America stands for anything, and those people aren't willing to advocate for it. Putting the present administration aside, this is a problem. But the bigger problem I worry about is people living in their own worlds and caring less about the worlds of others. We need a shared experience substrate. At least to some degree.

Taste is not a moat TASTE IS NOT A MOAT

AI gives you a motorcycle for the mind but someone has to drive. MOTORCYCLE FOR THE MIND

The only moat left is money ELLIOT'S ESSAY

A Pascal's Wager for free will says bet on its existence THE DEONTIC WAGER PAPER

Some theodicies accidentally suggest God should make us NPCs FULL PAPER

AI's making code more like labor than capital DANCO ESSAY

AI makes you boring AI MAKES YOU BORING

Influence diagrams help you regain control over complicated systems THINKIE INFLUENCE DIAGRAM

Norwegian scientist zaps himself with Havana syndrome weapon, gets Havana syndrome He didn't believe it, but now he does. WASHINGTON POST ARTICLE

Being more public and online lets your future friends find you FERNANDO BORRETTI ESSAY

.plan files keep you organized and improve technical writing. MATTEO'S PLAN FILES POST

Building a TUI is easy now HATCHET BLOG POST

You die every night when you sleep and a new consciousness wakes up In a short story by Erik Hoel, aliens realize humans die every sleep cycle, rebooting consciousness each morning with the same memories but entirely new awareness. THEY DIE EVERY DAY | ERIK HOEL'S SUBSTACK | THEY'RE MADE OUT OF MEAT

Prompt injection taxonomy should focus on intent over text techniques This is a pretty cool methodology by LASSO SECURITY, but I prefer my buddy Jason Haddix's, which I think is simply the best anywhere on the internet. THE ARCANUM PROMPT INJECTION TAXONOMY | LASSO SECURITY BLOG

Beautiful interactive explainers generated with Claude Code EXPLAINERS SITE

Agents are making filesystems cool again 1PASSWORD BLOG POST

Maxxing and mogging explained by internet manosphere MORNING BREW EXPLAINER

Someone built a wearable AI narrator that describes your life in real-time like a movie A Raspberry Pi Zero with a camera captures images every 20 seconds, sends them to OpenAI's vision API, turns the descriptions into third-person narration via ChatGPT, and plays it through wireless earbuds. LAMPYSECURITY ARTICLE

Consider the anti-fragility framework of:

DO:

BE: