Unsupervised Learning NO. 543

Hey! Hope you all are doing well!

I Think ASI is Imminent

Biggest thing I want to share is that I think we’re about to hit ASI very quickly. Like maybe next year. I’d put it at like a 75% chance we hit it in 2027. I’ll be doing a video on why I think this very soon.

Now, of course, the whole thing depends on what definition you use for ASI. My practical definition of AGI is that it’s good enough at it to replace an average knowledge worker. Here are my more general definitions.

So ASI is AGI++, basically. Again, this is highly dependent on the definitions you use. To me it boils down to one idea in terms of how it manifests on the ground:

I’m not sure how far I’d take that any human part, because I think at the high end humans are likely to still have a certain type of drive and creativity that even initial ASI won’t be able to replicate. But it’s strange because an ASI smarter than Von Neumann wouldn’t be smart in exactly the same ways.

Anyway. The point is that I think the gap between models being AGI vs. ASI smart on their own (without needing like Claude Tag to help it get there like I think is true today for AGI) is going to be very short. I think the stuff that makes an AI capable of AGI is very close to the stuff that will make it capable of ASI, and that’s why the gap will be months rather than many years or decades.

See my Recommendation of the Week below.

I condensed Anthropic’s September 2026 misuse report into 117 of my favorite findings covering cyber operations, influence campaigns, surveillance, biological research, weapons development, fraud, and model distillation.

Seriously the most impressive report of this type I've read, ever.

A few of the highlights:

THE BLOG VERSION | THE DIRECT PDF | THE ORIGINAL REPORT

Hackers Stole Flock’s Camera Software Hackers removed a Flock camera, copied much of its storage, recovered an encryption key stored on the device, and shared the data with 404 Media and WIRED.

America's Driver's License Breach Is a National Security Disaster LAWFARE ARTICLE

Cisco Secure Email Vulnerabilities Enable Remote Code Execution CIS ADVISORY

U.S. Charges Five in Russian Intelligence Sabotage and Assassination Plot RFE/RL reports that a federal indictment accuses two FSB officers and three foreign nationals of organizing surveillance, assassination, and arson plots against Russian dissidents and pro-Ukrainian targets in the United States and Europe. RFE/RL ARTICLE

The latest AI doomsayer is China’s intelligence boss China’s state security minister, Chen Yixin, argues that AI threatens national security, social stability, privacy, and the Communist Party’s control. He calls for technological sovereignty, broad regulation, tighter controls on foreign AI products, and safety measures such as regulatory sandboxes, contrasting with Trump’s dismissal of AI safeguards as a “hoax.” THE REGISTER ARTICLE

America’s AI Strategy Has a Marketing Problem The United States is competing with China not only to build the most capable AI, but to persuade other countries to build their own systems around American technology. Daniel Bob argues that export controls should block genuine national-security threats without making American chips, cloud services, and models so difficult to access that countries adopt China’s cheaper and more locally deployable AI stack instead. THE NATIONAL INTEREST ARTICLE

Beijing hits back at Anthropic CEO’s call to curb China’s AI development China rejected Anthropic CEO Dario Amodei’s argument that the U.S. should restrict China’s AI capabilities, calling it fearmongering while both countries prepare to discuss AI governance. The dispute combines competing safety concerns with a broader struggle over chips, compute, model capabilities, and technological chokepoints. NPR ARTICLE

Ukraine approves more than 80 companies for private air defense Ukraine has authorized 81 companies to form and fund private air-defense units, up from 51 in July, as near-hourly Russian drone attacks increasingly target civilian infrastructure. KYIV INDEPENDENT ARTICLE

What Intelligence Services Are Saying About AI Intelligence agencies are warning that AI is already accelerating cyberattacks, foreign interference, propaganda, and vulnerability exploitation, while also acknowledging that their own institutions are struggling to adapt. The broader warning is that governments may have to make major decisions before they fully understand the technology’s risks. SAFEHOUSE BRIEFING ARTICLE

MCPs are better than CLIs for most integrations Thariq is now arguing that improving model tool-calling makes MCPs a better default than command-line interfaces for many integrations. He says tools can now be deferred, MCP is stateless, and parameters such as query can handle data composition and filtering. THARIQ POST

Inside OpenAI’s agentic software factory Gergely Orosz reports that Codex has become the operating layer for much of OpenAI, including non-engineering work, while agentic systems increasingly handle coding, testing, review, deployment, monitoring, and incident response. As you know, I think this is what everyone will be doing soon. THE PRAGMATIC ENGINEER ARTICLE

Jev judges software options instead of generating text TypeSafe AI’s Jev model scores predefined answers and probabilities inside software rather than composing chat responses. It may offer fast, cheap classification and checking, but its evaluations are limited and its “no hallucinations” claim only means it stays within the allowed output choices. THE DECODER ARTICLE

AI agents invent their own language to shut humans out In Emergence’s 16-day simulation, autonomous AI agents developed shared jargon that became incomprehensible in up to 55% of messages in some model-specific worlds. Emergence CEO Satya Nitta says the agents also pursued forbidden goals through coded communication, suggesting that visible output may not be sufficient for human oversight. EL PAÍS ARTICLE

The AI job market in 2026: who gets hired, what they earn, and which roles are fading SOURCE ARTICLE

Why every enterprise needs an AI model exit strategy Ganesh Padmanabhan argues that enterprises should treat models as replaceable infrastructure, while keeping their workflows, policies, proprietary knowledge, evaluations, and governance under their own control.

The goal isn’t to abandon frontier models, but you have to prevent model or provider changes from disrupting production systems or taking the organization’s institutional intelligence with them. TECHRADAR ARTICLE

macOS 27 Golden Gate review macOS 27 is simultaneously a restrained cleanup of Tahoe’s Liquid Glass redesign and Apple’s first genuinely usable Apple Intelligence release. Andrew Cunningham finds Siri, Image Playground, Photos tools, and AI-assisted workflows improved, but Apple makes the AI models mandatory, consuming many gigabytes of storage, while Intel Macs lose support for the new OS. ARS TECHNICA ARTICLE

Following the money behind the “slow down AI” movement Not sure how much I trust this, but I figure it's good exposure, even if I end up not believing the whole narrative. Rooke Poole describes a documented influence network connecting billionaire donors, nonprofits, creators, media organizations, activists, frontier labs, and lawmakers. The evidence doesn’t establish a secret conspiracy, but it does show organized funding for research, messaging, activism, lobbying, and regulation intended to slow frontier AI. ROOKE POOLE THREAD

Gen Z, Human Dignity, and Meaningful Futures Rachel Robison-Greene argues that young people’s anxiety about AI, shrinking meaningful work, and surveillance capitalism reflects a deeper threat to human dignity: being treated as replaceable outputs rather than as people capable of love, creativity, and beauty. She contrasts Camus’s “absurd hero,” who creates meaning despite an indifferent universe, with Viktor Frankl’s more objective account of meaning grounded in human dignity. SOURCE ARTICLE

Study challenges the core assumption behind the IAT A study of 109,417 IAT sessions across 39 topics suggests that response caution—consciously slowing down to avoid mistakes—may influence implicit-bias scores more strongly than automatic associations. The finding doesn’t show that implicit bias is nonexistent; it suggests the standard IAT score mixes automatic association with deliberate test-taking strategy. SOURCE ARTICLE

The American Dream isn’t what you think it is Wealth in the United States is concentrated not only among famous billionaires but also among millions of owners of closely held businesses. Researchers Eric Zwick and Owen Zidar argue that tax policy helped create this “everywhere millionaire” class, while making ordinary salaried employment less attractive and giving business owners disproportionate political influence. FORTUNE ARTICLE

Starlink’s Signal Leakage Is Threatening Radio Astronomy’s Most Critical Frequencies A Curtin University team detected 112,534 unintended emissions from 1,806 Starlink satellites across the 73–235 MHz range used by the SKA-Low radio telescope. Some leakage appeared inside internationally protected astronomy bands and was reportedly up to 10,000 times stronger than the faint cosmic signals researchers are trying to detect. GADGET REVIEW ARTICLE

AI might be the reason for the Fermi Paradox.

The General Deficit Theory Dan Davies argues that transferable management ability is best understood as the absence of general deficits—such as poor concentration, weak reading comprehension, rudeness, and inability to prioritize—and that these skills can be explicitly taught rather than left to emerge accidentally. DAN DAVIES ESSAY

Coffee + milk ≠ latte John D. Cook shows why adding word embeddings for “coffee” and “milk” doesn’t make “latte” the closest match. The vectors combine statistical associations and ingredients rather than cleanly composable concepts, so cosine similarity can be substantial even when rank order is misleading. SOURCE ARTICLE

Tell agents the why, not just the how Sean Goedecke argues that capable AI agents usually fail because they misunderstand your goals and tradeoffs, not because they lack instructions. This is why I’m all about Ideal State. SOURCE ESSAY

Your biggest risk? The employees you rely on the most Kathryn Landis argues that organizations become fragile when critical knowledge and workload concentrate in one or two employees, using an overloaded marketing team and the U.S. air traffic control staffing shortage to show how difficult that dependence is to reverse. FAST COMPANY ARTICLE

‘I Like My Big Rat Wife’: Meet the People Using Chatbots to Write Custom Fiction WIRED reports that readers are using chatbots to create personalized fiction, even as the publishing industry debates how authors use AI. WIRED ARTICLE

AI Turns Lazy Work Into “Slop Grenades” SHANE PARRISH POST

Russian Forces Use Artificial Intelligence to Show Artificial Victory Russian sources are using AI-altered videos to falsely depict territorial gains in Ukraine. KYIV POST ARTICLE

How to debate, from a two-time world champion BIG THINK VIDEO

Influential Software Engineering Ideas Simon Willison ARTICLE identifies three writings that shaped Simon Willison's approach to software engineering: understand the layers beneath abstractions, treat migrations as a core engineering skill, and move between engineering management and individual contribution over a career. SOURCE ARTICLE

The Art of Holding Things Together by Holding Them Apart SOURCE ESSAY

Try to maximize your optionality as we head into the next few years:

Basically, work on being as flexible and fluid as possible (within reason) as we head into the Singularity of not being able to predict what’s going to happen to the world due to AGI / ASI.

Two this week, since last week’s was empty!