Unsupervised Learning NO. 533

Hey! Hope you all are doing well!

—

I had a debate on Friday with Zach Korman about the Morality of Dario Amodei. Here’s the video:

Some quick post-debate thoughts:

I think Zach is very wrong about the tech itself and its capabilities, and the speed with which it’s improving. And I think this is a big part of our disconnect.

I think his core argument is bad, and I’m looking for a great analogy for it. But it does hinge on the point above. It’s something like,

Basically he’s putting a high-level principal (which I also agree with) above real-world practical risk, and he’s doing so because he doesn’t believe that risk is real. But then, even if he did, his solution for that (complete and immediate Nationalization) isn’t good either. That’s my core summarization of the debate, but you should watch it.

I forgot to ask my favorite thing in these types of discussions, which is, “What would make you change your mind?”

Here’s my answer: If I saw Dario moving away from a consortium with competitors and democracies, and more towards this should all be up to me/Anthropic, that would sway me. Or if I saw him dropping his values and saw him releasing too fast in order to maximize profits. I should have asked Zach what would change his mind.

WATCH THE VIDEO

—

My new favorite joke:

“If alcohol can shorten your memory, imagine what alcohol can do.”

I love these kinds of short pun jokes. Some other favorites of mine:

—

I love the prompts I have in here already, but will keep adding to the list. It’s really not just for Fable, but for whatever the new hotness is.

—

My improved (in my opinion) definitions of AGI vs. ASI. Highly simplified, basically, to enable better conversations.

—

My latest piece on who will get the new jobs created by AI, and how to make sure we can be in that group.

—

Had a notably great sponsored conversation with the people over at Scanner.dev. I don’t know what I was expecting, but it wasn’t that. One of the coolest pieces of tech I’ve seen in years. THE CONVERSATION

—

OpenAI says China-linked accounts tried to steer the data center fight OpenAI banned fake accounts that used ChatGPT to push posts, comments, and cartoons around US data centers and tariffs. TECHRADAR ARTICLE

Novo Nordisk lost AI models and patient data in a breach Attackers stole training code, model checkpoints, and some de-identified patient records. Now they’re trying to get ransom money. SECURITYONLINE.INFO ARTICLE

China used fake consulting sites to hunt clearance holders The FBI and DOJ seized 13 fake consulting sites. They were part of a long-running Chinese targeting campaign against officials with access to classified info. DEFENSE ONE ARTICLE

Cursor changes the result more than the model Same model, different harness, and Cursor gets a much better security score on real coding fixes. Can’t stop won’t stop saying the harness matters way more than the model, at least for now. ENDOR LABS BLOG

UV adds built-in vulnerability and malware checks uv now scans dependencies for known vulnerabilities and deprecated packages, and it can stop installs when OSV flags malware in your lockfile. ASTRAL BLOG POST

How threat intelligence feeds help automate socs to reduce mttr AI helps here, but the real win is cleaner data flowing into SIEM, SOAR, IDS, and EDR so analysts move faster. CYBER SECURITY NEWS ARTICLE

The warfare of the future is already here AI and drones are turning war into a fast system fight. Some of this is good, but it seems obviously bad to make it easier to kill from a distance. NOEMA MAGAZINE ESSAY

GEOINT is becoming machine-readable for autonomous war GEOINT is shifting from mapmaking for humans to encoded geographic logic for autonomous systems. MODERN WAR INSTITUTE ARTICLE

U.S. tests drone swarms to stop drone attacks AFRICOM’s CURTAIN CALL system uses defensive drone swarms to spot and hit incoming UAVs. The test looked promising, but it’s still experimental. NV.UA ARTICLE

Space Force buys two more GPS satellites The Space Force just handed Lockheed Martin $514 million for two more GPS 3F birds. These are built to be tougher against jamming and spoofing. SPACENEWS ARTICLE

Ukraine’s drone war is getting more serious Ukraine is using bigger middle-strike drones to hit bridges, air defenses, and logistics nodes across occupied territory. It also points to autonomous interceptors and drone deals as the next step. UKRAINE'S ARMS MONITOR POST

Your AI lead should be an engineer This argument says you don’t want a vague “AI person.” You want someone who can actually build the data pipeline behind the thing. I mostly agree, but as that skill becomes easier the taste will be increasingly more important. ABIGAIL HADDAD POST

GLM-5.2 tops Artificial Analysis’ open-weights chart Z AI’s GLM-5.2 jumps ahead of other open models on Artificial Analysis, but it burns more tokens to get there. Tons of excitement around this model right now. But it works best via cloud, and I’d rather self-host. ARTIFICIAL ANALYSIS ARTICLE

Claude Code feels like a strategy game now Claude Code turns coding into a loop of moves, feedback, and micromanagement that feels more like RTS play than chatting. PROVI BLOG

AWS is turning context into the thing agents run on This is promising. AWS is stitching together catalogs, graphs, S3 metadata, and permissions so agents can reason over governed context instead of guessing. AWS BLOG ARTICLE

McKinsey’s latest AI survey says agents are still early Most companies use AI now, but they’re still stuck in pilots and only a few are turning it into real enterprise value. Agree with this. I think the difference between companies/people doing this well and poorly is vast. It all comes down to this: Most Companies Aren’t Anywhere Near Ready for AI. MCKINSEY REPORT

The harness is the business now An argument I’m sympathetic to, obviously, that OpenAI’s IPO story isn’t really about the model. It’s about the work layer around it, and who owns that layer. ACAST PODCAST

The brain might learn backprop by using time instead of error cells This thread says cortex may approximate backprop through 200 ms phases, and that idea matters for neuromorphic AI hardware. THREAD READER APP POST

AI won’t fix a company that can’t ship Nikhil Suresh says most AI hype just covers up weak teams, bad processes, and managers who already don’t ship. MURRAY ROBINSON PODCAST

The repo should become the wiki now Teams using agents need truth beside the code, not buried in Confluence. Tony Karrer says docs drift becomes a real production risk once agents start trusting it. I really love this idea of the “wiki” being constantly maintained by agents like a code repo, which I think is going to be critical when agents start relying on corporate knowledge to make decisions TECHEMPOWER BLOG POST

Midjourney wants to turn body scans into a spa habit Midjourney is building a water-based MRI-like scanner and a spa around it, so people can get fast body data without making it feel clinical. MIDJOURNEY MEDICAL BLOGPOST

Meta is unwinding a giant Manus deal after Beijing stepped in Meta is backing out of its $2 billion Manus acquisition after China forced a breakup over national security concerns. The startup’s founders may try to buy it back. TECHCRUNCH ARTICLE

Design is quietly getting a bigger seat at the table AI is making companies notice that shipping fast isn't enough anymore. The real fight is over judgment, trust, and what gets built. UX COLLECTIVE ARTICLE

Technical literacy is beating the mba in some jobs Fast Company says managers now need to read software systems, AI flows, and vendor tradeoffs, not just business school slides. FAST COMPANY ARTICLE

Elon Musk is wiring his companies into one AI stack Musk is stitching xAI, X, Tesla, Neuralink, and SpaceX into one system for data, compute, robots, and interfaces. TECHREPUBLIC ARTICLE

Ubiquiti is shipping a license-free enterprise NAS Tom Hildebrand says ENAS gives businesses ZFS storage, local backups, and shared block storage without the usual licensing mess. UBIQUITI BLOG ARTICLE

Estonia wants to give AI agents government IDs Estonia is trying to stop agents from borrowing a human's whole identity just to do one task. The bigger question is who gets blamed when the agent messes up. DECRYPT ARTICLE

Sound lets coffee skip hot water and cut energy use Researchers at UNSW brewed espresso with room-temperature water and sound waves, and people couldn’t tell the difference. DIGITAL TRENDS ARTICLE

Sam Harris talks debt, AI, and the economy Sam Harris and Noah Smith go from national debt to AI, then land on inflation, inequality, and culture. WAKING UP PODCAST

The mom who runs a household with AI agents Jesse Genet has agents buying groceries, buying books, and handling homework and paperwork while she stays busy. THE CUT ARTICLE

WhisperQuest turns world building into a playable AI RPG WhisperQuest is a tiny AI role-playing game where you make your own worlds and run solo or group adventures inside them. I love stuff like this because I think humans creating experiences for other humans will be a huge part of a positive AI/tech outcome. WHISPERQUEST APP

AI can spot Alzheimer’s risk in a retina photo Cheap eye photos now seem good at flagging the blood pressure, sleep, and habits that pile up into Alzheimer’s risk. NEUROSCIENCE NEWS ARTICLE

AI watches aging parents so families feel safer Wired follows a son whose father wants to stay home, while sensor companies sell peace of mind to families and stretched care agencies. WIRED ARTICLE

AI is making answers cheap and curiosity priceless AI speeds up answers, but it also tempts leaders to stop asking better questions. The real failure is mistaking a polished response for actual understanding. FAST COMPANY ARTICLE

Beware of star trek managers, especially when bearing MBAs Jussi Pakkanen says Star Trek bosses make terrible real-world managers because they confuse command with competence. JUSSI PAKKANEN POST

The ai debate is really about free will AI won’t wipe out randomness, so the real fight is whether humans live inside a deterministic universe. JUILEE BHOSALE SUBSTACK POST

America's infrastructure is still flying blind Gregg Herrin says the U.S. spent huge money on pipes and grids, but still can't see what they're doing in real time. Digital twins and better metering would catch failures earlier and make the whole system less fragile. This is a huge part of my TRIOT book, basically sensors on everything to get real-time state, and then AI to help us manage it according to our human preferences. FORTUNE COMMENTARY ARTICLE

Your raise used to go offshore, then to buybacks, now to data centers FORTUNE ARTICLE

People finish broken stories with the version they already want People don't wait for full evidence when a story has gaps. They snap it shut with the explanation that already fits their frame. PRESTON LOCKE POST

The work AI can’t do Companies are pouring money into AI people analytics while the human side of management gets thinner and less capable. FAST COMPANY ARTICLE

How to be irreplaceable AI only eats what you can explain. Alberto says the stuff that makes you hard to describe is the stuff that keeps you safe. Cool idea but this reminds me of encryption: if your security relies on the secrecy of the mechanism, you’re screwed. ALBERTO ROMERO POST

The people who know too much Dr. Dinesh Kumar Jangra rounds up stories about hobbyists whose fixation turns into real expertise and public value. LONGREADS ARTICLE

CyCognito turns pentesting into a live AI attack simulation TNW ARTICLE

Claude Code feels like a strategy game now Claude Code turns coding into a loop of moves, feedback, and micromanagement that feels more like RTS play than chatting. PROVI BLOG

Bubbles is a small web front page for indie blogs It pulls in independent personal blogs, ranks them by votes and freshness, and lets the crowd shape what rises. BUBBLES TOWN

Dostoyevsky got a fake execution and came back obsessed with life Maria Popova pulls one letter out of that night and shows how close death made him talk like a different person. MARIA POPOVA ESSAY

File systems are the new primitive for AI agents Carter Rabasa says AI agents are moving from databases to file systems because that’s where the real work and state live now. CARTER RABASA POST

Currently tripping out about the idea of eternal, or evergreen, questions.

Like what’s my running list of questions that nobody on Earth is smart enough to answer?

This can be big things like physics or biology or whatever. But also things like,

What’s your list?

(Email me if you have any good ideas for inclusion, I’ve got a list started here but these are all AI so far. I’m still working on mine)