
UPDATES
Hey! Hope you all are doing well!
Just did an AI workshop at the most excellent Swiss Cyberstorm Conference in Bern, Switzerland! My second time being here and I just love the energy here, and the people. Some of the smartest people I’ve ever seen gathered.

The Swiss alps from the hotel, including Eiger, Monch, and Jungfrau. Image doesn’t remotely capture it.
—
New essay out about how both models and humans experience model collapse… READ IT
The most powerful troll I’ve ever seen…I could barely resist…

I will not reply that I have told them. I will not reply that…
—
Not saying I agree with this, but I find it worth thinking about.

I’m all about some AI, but I really have been wondering about this exact shell game. I am unqualified to know/say whether it’s an actual problem or not, but my guess is that it’s not nothing.
—
Sponsor
Identity is now the primary attack surface, not your network
Infostealers have quietly become the backbone of cybercrime—fueling ransomware, fraud, and massive data breaches through stolen identities.
Flashpoint’s Proactive Defender’s Guide to Infostealers explains:
How today’s leading infostealer strains operate
How to operationalize intelligence for proactive defense
Get the full breakdown in the guide to strengthen your defenses against the next wave of identity-based attacks.
CYBERSECURITY
Local LLMs might be less secure than cloud ones because you're basically running untrusted code on your machine A Quesma post argues local models create attack surface since they're essentially executing weights from the internet, while cloud providers can sandbox better. This really depends on the models you’re using, on which platform. Some are more risky than others, but in general they’re fairly safe. It is worth noting that there’s some risk though, which most people aren’t aware of at all. QUESMA BLOG POST | REDDIT DISCUSSION
China says the US hacked a government agency via text-messaging vulnerabilities China claims the NSA exploited text-messaging system flaws to control mobile devices of government staff between 2022 and 2024. THE INFORMATION BRIEFING
183 million Gmail passwords leaked in massive data breach Forbes reports a huge Gmail credential dump hit the dark web, though it's likely a compilation of old breaches rather than a fresh Google hack. FORBES ARTICLE | HN DISCUSSION
Chinese and Russian spies are using honeypots and pitch competitions to steal Silicon Valley secrets Some counterintelligence people are saying attractive women on LinkedIn and startup contests are going after influential tech people—one Russian "crypto expert" married an aerospace guy and had kids while running a lifelong collection op. Sounds like the best spy show ever, The Americans. THE TIMES ARTICLE
Sponsor
Can GPT-5 pass a real security audit?
Sonar analyzed GPT-5’s AI-generated code with large-scale static security testing across 4,000+ Java tasks. The model wrote cleaner code than earlier LLMs yet still produced injection, path traversal, and concurrency flaws. The findings reveal where AI coding progress causes a bottleneck in verification.
It was a DNS race condition in DynamoDB that brought down AWS for 15 hours A software bug in Amazon's DNS management system cascaded through their network, causing one of the largest internet outages ever recorded. ARS TECHNICA ARTICLE | AWS POST-MORTEM | OOKLA ANALYSIS
L3Harris cyber exec allegedly sold $1.3M in secrets to Russia Peter Williams, who ran the company's offensive cyber weapons division, got busted and feds are seizing his Rolexes and crypto. COURT DOCUMENTS | L3HARRIS TRENCHANT | THE REGISTER STORY
North Korea hacks European drone makers with fake job offers THE RECORD ARTICLE
Top security researcher finds bugs by following curiosity not methodology GitHub spotlights @dev-bio, who discovers injection flaws by getting sidetracked exploring how systems handle edge cases rather than following rigid processes. I’m sure they do both though. Balancing both is key. DEV-BIO'S PAGE | DEV-BIO LINKEDIN | GITHUB BUG BOUNTY PROGRAM | GITHUB VIP PROGRAM POST | SUBMIT BUGS ON HACKERONE
NATIONAL SECURITY
China controls the world's medicine supply and could weaponize it China dominates global active pharmaceutical ingredients production, giving them leverage beyond rare earths in trade negotiations with the US. OODALOOP ARTICLE
Germany and UK team up to hunt Russian submarines with Boeing submarine detectors Germany's buying eight P-8 Poseidon planes that drop sonar buoys to track Russian subs across the north Atlantic from British and German airbases. DW ARTICLE
Switzerland is spending millions upgrading its network of secret nuclear bunkers The country has enough bunker space for its entire population and they're hidden everywhere, including a seven-story parking garage in Zurich's old town. WASHINGTON POST ARTICLE
AI
Amazon cuts 14,000 corporate jobs while pouring billions into AI infrastructure Amazon's cutting 4% of corporate staff to fund a $10 billion AI data center push, basically trading people for cloud computing capacity.
The people saying AI isn’t affecting jobs are going to have an increasingly difficult time making arguments. It’s so obvious from so many angles at this point. AMAZON JOB CUTS ANNOUNCEMENT

Researchers define AGI as systems that can do any economically valuable work Really cool paper here. They propose AGI means automating any remote job a human can do, which gives us a testable definition instead of the usual philosophical handwaving. The methodology they use for defining cognitive work is quite interesting. Dawn Song is one of the authors, and she’s brilliant. ARXIV PAPER | HN DISCUSSION
Apple starts shipping Houston-built AI servers with custom silicon for Private Cloud Compute Tim Cook announced Apple's American-made servers are now shipping from Houston to power Private Cloud Compute, which runs custom Apple silicon and promises zero data logging. COOL NOW FIX SIRI TOM'S HARDWARE ARTICLE | APPLE SECURITY BLOG
Anthropic expands Google Cloud TPU deal to hit 1GW of compute by 2026 Anthropic's getting up to a million TPUs from Google Cloud, letting them skip the risky infrastructure buildout while betting on cloud scalability. TOM'S HARDWARE ARTICLE | GOOGLE CLOUD PRESS RELEASE
AI researchers are burning out because the window to stay relevant is actually closing In an essay by Nathan Lambert, he argues that training frontier LLMs now requires athlete-level dedication, and starting from scratch gets harder every day as established labs compound advantages in tools, data, and team culture. INTERCONNECTS ESSAY | WSJ AI WORKERS ARTICLE | ATOM PROJECT
A chart that’s been making the rounds. It’s from Bloomberg, so the numbers are likely accurate at least. Don’t accept or dismiss it. Just think about it.

Context engineering is sleeping on the humble hyperlink Michael Bleigh argues LLMs should dynamically load context through links like humans do—it's token-efficient, works anywhere, and consolidates dozens of tools into one. MBLEIGH ARTICLE | ANTHROPIC CONTEXT ENGINEERING | HATEOAS WIKIPEDIA | GENKIT | MCP RESOURCES SPEC | FIREBASE MCP SERVER
Trump administration wants AI data centers connected to the grid faster The Energy Department proposed rules to speed up grid connections for data centers, especially ones that bring their own fossil fuel or nuclear power. THE HILL ARTICLE
Apple starts shipping Houston-built AI servers for Private Cloud Compute TOM'S HARDWARE ARTICLE
OpenAI acquires the Apple Shortcuts team to build deeper OS integration OpenAI bought SAI, the team behind Apple Shortcuts, who've been building Sky—a context-aware AI layer for macOS they're now folding into ChatGPT. They’re clearly pushing for a way to bypass the phone and go direct from user to AI Device. OPENAI ANNOUNCEMENT | ARS TECHNICA ARTICLE
Claude now has persistent memory across conversations CLAUDE MEMORY ANNOUNCEMENT
Gemini CLI now runs vim and interactive commands without breaking context Google added pseudo-terminal support so you can run vim, htop, or git rebase inside Gemini CLI and it all stays in the AI's context. It’s super compelling, but the ecosystem still isn’t anywhere near as solid as Anthropic’s. GEMINI CLI ANNOUNCEMENT | GEMINI CLI DOCUMENTATION | GITHUB REPOSITORY
TECHNOLOGY
YouTube quietly becomes the dominant force in TV watching and creation YouTube now captures 13.1 percent of TV viewing (topping Netflix's 8.7 percent), pays out $100 billion to creators, and is positioning itself to absorb sports, late night, and scripted TV through its creator ecosystem. HOLLYWOOD REPORTER ARTICLE
Eugene Yan's advice on being a principal engineer who actually ships Eugene Yan says principal engineers should focus on execution over architecture astronautics—build stuff people use, not frameworks nobody needs. EUGENE YAN ARTICLE | HN DISCUSSION
Focus on building things people actually use instead of abstract frameworks
Principal engineers should prioritize execution and shipping over pure architecture
Technical leadership means solving real problems not creating complexity
The best principal engineers write code and deliver working systems
Avoid becoming an architecture astronaut who designs but never implements
All the senior people in these roles that I’ve seen have been exceptional coders and highly collaborative at the same time. Constantly talking to people and finding shared problems, and then solving them and rolling out the solution.
Apple releases 400K image editing examples with instructions and quality scores Apple's Pico-Banana-400K dataset has 400K text-image-edit triplets covering 35 edit types—Gemini writes the instructions, Nano-Banana does the edits, then self-grades them. PICO-BANANA-400K GITHUB | SFT MANIFEST | PREFERENCE MANIFEST | MULTI-TURN MANIFEST | OPEN IMAGES DOWNLOAD
Good code reviews shouldn't have more than five or six comments SEAN'S CODE REVIEW ARTICLE
Windows 10's end-of-support deadline accidentally becomes Apple's best sales driver Nearly 40% of all PCs still ran Windows 10 before the October cutoff, forcing mass upgrades that pushed Mac shipments up 14.9% year-over-year. And it’s happening to Linux too. Starting to worry a bit about Microsoft, actually. Both on the traditional tech front, but also on the AI story front. Where’s the actual moat? Office. Outlook. Corporate tech stack basically. But it somehow feels vulnerable right now. Just a feeling. COUNTERPOINT RESEARCH REPORT | MACRUMORS ARTICLE
Code like a surgeon means making precise changes without understanding everything Geoffrey Litt says you can effectively modify code by targeting specific areas like a surgeon operates without knowing every organ. But my favorite interpretation of this is how you can use AI to delegate certain things, but that you should never delegate the surgery itself. And of course what that surgery is depends on the task. GEOFFREY LITT ARTICLE | HN DISCUSSION
Tech's second layoff wave is flooding startups with top talent they can finally afford. PAVEL SHYNKARENKO
Derek Sivers open sources his entire database-driven web stack Derek Sivers released his full PostgreSQL-powered website architecture where database functions return complete HTML responses, eliminating the typical app layer entirely. GITHUB REPO | DEREK'S SITE | CONTACT DEREK
HUMANS
Over a million people weekly confide suicidal thoughts to ChatGPT OpenAI says 0.15% of its 800M weekly users show suicide planning indicators, plus hundreds of thousands more show signs of psychosis or dangerous emotional attachment. I wish I had a few weeks to dig into the implications of this number on the state of American mental health. On the one hand the number seems high, but .15% seems low. That’s the trick with stats; you need to do work to get truth out of them. OPENAI ANNOUNCEMENT | TECHCRUNCH ARTICLE | GPT-5 SENSITIVE CONVERSATIONS ADDENDUM
Corporate America is crushing earnings despite tariffs and economic concerns Over 85% of S&P 500 companies beat expectations in Q3—the best rate in four years—while jobs keep disappearing. MORNING BREW STORY | BLOOMBERG EARNINGS ANALYSIS | AXIOS JOBS MARKET REPORT
Americans can't afford their cars anymore and Wall Street is worried TELEGRAPH ARTICLE
A4L architecture aims to personalize adult online education using AI agents Researchers at the National AI Institute for Adult Learning describe a data architecture that collects learning analytics and feeds them back to teachers, learners, and AI agents for scalable personalization. A4L ARXIV PAPER
People walk 15% faster in cities and hang out 14% less than in 1980 MIT TECHNOLOGY REVIEW ARTICLE
Intelligence is just compression plus the ability to act on it A fascinating interactive essay argues intelligence isn't about thinking or consciousness—it's pattern compression that enables prediction and action in the world. WHAT IS INTELLIGENCE ESSAY | HN DISCUSSION
Intelligence is prediction all the way down, from molecules to societies to AI Blaise Agüera y Arcas argues that prediction isn't just what brains do—it's what life itself does, and modern AI systems might actually have real intelligence, consciousness, and free will. WHAT IS INTELLIGENCE BOOK
Humanity trades physical survival for psychological battles as we conquer each layer of need In an essay by Dave Shapiro, he argues that solving hunger gave us obesity, solving communication gave us digital addiction, and solving work will give us sloth—the struggle just evolves upward. DAVE SHAP'S ESSAY
Be careful with Obsidian because perfectionism kills actual thinking Phong warns that fiddling with your note-taking system becomes procrastination disguised as productivity, and the best ideas come from just writing stuff down. PHONG'S ARTICLE | HN DISCUSSION
DISCOVERY
Evan Hahn shares his most-used personal scripts EVAN HAHN'S SCRIPTS
Boring lives might actually be the happiest ones WHY A BORING LIFE MIGHT BE HAPPIEST
Hackers share their favorite cult sci-fi books HACKER NEWS DISCUSSION
Dostoyevsky wrote The Brothers Karamazov while staring into the abyss of his son's death THE LIGHT OF THE BROTHERS KARAMAZOV
China controls 55% of the world's high-IQ working-age population SOFIE CHAN ARTICLE
Burp extension maps Next.js server action hashes to actual function names NEXTJS SERVER ACTION ANALYZER
I'm drowning in AI features I never asked for and I hate it MAKEUSEOF ARTICLE
RECOMMENDATION OF THE WEEK
Brute force maintain contact idea.
Make a list of all your friends and associates from your whole life that you think it’s worth something to keep up with. Probably dozens.
Put it on the calendar, like once a month to twice a year, to text or call them.
Follow the schedule.
Virtually every study that’s looked at it has shown that people live longer and are happier based on the relationships they maintain.
Put it on the calendar if you have to.
APHORISM OF THE WEEK
The essence of the independent mind lies not in what it thinks, but in how it thinks.
GET THE MEMBER EDITION
You’re currently receiving the STANDARD edition.
Members get numerous benefits, including:
25-50% off all UL Paid Content, including the upcoming Human 3.0 / AUGMENTED ONLINE portal!
Access to the extraordinary UL Member Community that includes vibrant conversations with ~1,500 of the smartest and kindest people you’ll find on the internet
Member-only Content, such as EDC guides on tech stacks, personal productivity routines, my recommendations on Critical skills to Build Going Forward, Trend Identification and Analysis, and more…
Access to the Member Archive of previous Member-only content, the Book Club archive, etc.
Access to The UL Book Club that’s been going monthly since 2017! One of the highlights of my and many attendees’ month!
Access to the Monthly Member Meet-up where we talk about our routines, productivity workflows, what’s on our minds, etc.
Access to In-Person Events like our dinners in Vegas, San Francisco, etc.
And much more coming…
This is the moment to connect with others who are smart, kind, and asking the same questions we are. Where is this all going? And how do to prepare?
Join the conversation.

