Our paper on Subliminal Learning was just published in Nature! Last July we released our preprint. It showed that LLMs can transmit traits (e.g. liking owls) through data that is unrelated to that trait (numbers that appear meaningless). What’s new?🧵

The paper Agents of Chaos, co-authored by scholars from MIT, Harvard, and Stanford, studies autonomous AI agents in real environments. It finds they obey unauthorized users, execute harmful commands, and misreport outcomes—exposing a core alignment gap where autonomy creates security and governance risks. As cities integrate AI into infrastructure (e.g., traffic or energy grids), the risk of "uncontrolled resource consumption" or "partial system takeover" documented in the paper could translate into physical chaos. Both cities and firms must address the "delegated authority" problem. If an agent acts on behalf of a citizen or employee but destroys the "owner’s" assets to protect a third party’s secret, the traditional lines of liability and responsibility vanish.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

There's no doubt that peptides are making young people gay at an alarming rate. Our job is to capitalize long term on this trend by studying the second-order effects and investing accordingly. Here' how I'm playing this trend: Obvious buy is Grindr. And Match Group owns Hinge and Tinder which are both deep into the gay market. Bumble less gay exposure but still benefits. Gays also love and spend alot of money on pets so you want to be long Chewy. Trupanion for insurance, Zoetis, and Idexx for animal health, Freshpet for premium food. Pet spending is probably the biggest second-order effect by total dollars Gays travel more and index heavily toward urban destinations, cruises, and resort experiences. Royal Caribbean, Norwegian, Marriott, Hilton, Booking, and Airbnb. Cruise lines in particular have a long history of chartered gay cruises. Gays are heavy into personal care and grooming and spend rises materially. Estée Lauder, Ulta, and e.l.f. are all obvious beneficiaries of this trend. Gays are into fitness and adore Lululemon. And there's the publicly traded gyms like Life Time, Planet Fitness. Gym membership rates among gay men run well above the general population. Premium fashion and accessible luxury is a major gay market. Benefits Tapestry (Coach/Kate Spade), Capri (Michael Kors/Versace), Ralph Lauren, and the European luxury houses (LVMH, Kering) on foreign exchanges. You can probably short Mattel, Hasbro, Carter's, Children's Place, Kimberly-Clark (Huggies), and suburban-focused homebuilders like Lennar and D.R. Horton would face significant headwinds as gays hate the suburbs and flock to cities. P&G's Pampers business will be in trouble. Finally an asymmetric bet that fertility benefit companies like Progyny will see exponential growth as gays still want kids but through IVF and surrogacy. Hamilton Thorne makes IVF lab equipment.

Everyone talks about AI enabling the "one-person Hollywood studio." I spent the last two months putting that to the test. The result? A full, 2+ hour feature film. Introducing "The Four Classics." Using @suno, @soraofficialapp , @grok Imagine 1.0, and Genie 3 by @GoogleLabs , I adapted China’s greatest literary epics into a sprawling, cyberpunk hip-hop opera. Grab some popcorn. Welcome to the future of AI film production. 🍿🎥 Timestamps: Water Margin 0:00:09 “The Tiger (十八婉)”- Wu Song rejects the law and engages in a drunken battle defeating the Tiger 0:03:20 “Forced to the Mountain (逼上梁山)”- Lin Chong seeks revenge on Gao Qiu for destroying his life and familial honor 0:09:24 “Black Whirlwind (黑俊峰)”- Introduction of the brute Li Kui and his love for violence 0:11:42 “The Gathering / 108 (大聚义)”- The Liangshan heroes formally assemble to announce their reign at the Hall of Loyalty and Righteousness 0:14:08 “The Amnesty (招安)”- Song Jiang, leader of the heroes, debates the pros of accepting an offer of Amnesty from the Emperor 0:18:05 “The Poison Wine (神聚蓼儿洼)”- Song Jiang encourages Li Kui to join him in drinking Poisoned Wine gifted by the Emperor Romance of the Three Kingdoms 0:22:20 “The Oath (桃园)”- Liu Bei is a small time hustler with ambition to take over the nation. He makes a blood oath with Guan Yu and Zhang Fei to do so. 0:25:57 “The Mastermind (卧龙)”- Liu Bei recruits the strategist Zhuge Liang to join his cause and take over the kingdom 0:30:49 “The Burden (​仁义)”- Caocao’s troops attack causing Liu Bei and Shu to flee south. Liu Bei burns the city of Xinye and uses common people as human shields to make his escape. 0:33:44 “The License (挟天子)”- Caocao plans his move to crush the growing nation of Shu from his penthouse 0:36:45 “No Regrets (​宁负)”- Flashback to Caocao’s betrayal of his ally Old Lu. He chooses to massacre a banquet after hearing the sound of knives sharpening. An examination of ruthless ambition. 0:39:59 “The Flex (​八十万)”- Caocao’s army of 800k soldiers races south to destroy Liu Bei’s Shu. 0:43:26 “The Estate (​江东)”- Sun Quan, leader of Wu, forms an unlikely alliance with Liu Bei to quell the approach of Caocao’s army 0:47:27 “The Shift (东风)”- Zhuge Liang summons the East Wind to change the fate of Caocao’s army at the Red Cliffs 0:51:15 “The Chains (​连环)”- Pang Tong of Shu deceives Caocao into chaining his fleet of ships together setting the perfect trap 0:54:37 “Symphony No.7 (​​第七交响曲)”- General Zhou Yu of Wu sparks the trap set by Shu strategists lighting Caocao’s entire fleet ablaze. The Epic Battle of the Red Cliffs concludes Caocao’s push south. 0:59:25 “The Split (​​三分)”- Opening Credits. Sima Yi reviews the status of the Three Kingdoms after The Red Cliffs result in an uneasy truce. 1:02:23 “Still G.U.A.N.”- Following the war, Guan Yu runs operations in Jingzhou. Sun Quan sends a formal message requesting Guan Yu’s daughter marry his son strengthening the alliance. Guan Yu promptly refuses insulting the leader of Wu. 1:05:58 “The Statue (刮骨)”- After an assassination attempt, Guan Yu is struck with poison that makes its way into his bone. He requires immediate, invasive surgery. He survives through his grit, maintaining his arrogance, but weakened from the procedure. 1:09:42 “The Funeral (白衣)”- Guan Yu is ambushed by Wu’s forces after retreating to the mountains of Maicheng. The Wu assassins approach as merchants, disguised in white clothing, catching Guan Yu by surprise. 1:13:19 “The Lie (桃园)”- Liu Bei receives news of Guan Yu’s death and sets out on a war path against Wu. He loses his composure leaning into pills to cope with the loss of his blood brothers. 1:17:08 “The O.D. (七百里)”- Liu Bei’s madness leads him to make a fatal strategic error. He stretches his army thin in a 700 mile long “snake” formation. 1:20:32 “The Boomarang (​回旋镖)”- Liu Bei’s army burns to ashes in a brutal defeat reminiscent of his triumph at the Red Cliffs. 1:23:46 “The Flop (​扶不起)”- Zhuge Liang tries to pick up the ashes of Shu nation by educating Liu Bei’s son Adou. Adou is too busy enjoying his life and inherited wealth to bother with the burden of leadership. 1:27:48 “The Frequency (​空城)”- Cornered by the army of Wei led by Sima Yi, Zhugeliang employes the empty city strategy to outsmart his rival. 1:31:12 “The Flatline (​五丈原)”- Zhuge Liang tries to outlive his corporeal form to ensure the longevity of the nation of Shu and prevail in the war. 1:34:33 “The Liquidation (​归晋)”- Sima Yi consolidates control of the Three Kingdoms and unifies China under Wei. Journey to the West 1:38:20 “Eight Trigrams Furnace (炼丹炉)”- A family flees the grind of life in China in search of “Freedom” offered by the American Dream 1:41:36 “White Bone Spirit (白骨精)”- The family experiences the demons and tribulations of modern American life in their struggle to assimilate 1:45:33 “Thunderclap Temple (雷音寺)”- The family realizes the truth of the Blank Scriptures. Dream of the Red Chamber 1:49:39 “顽石GENESIS BLOCK (Final Cut)” - Rejecting its status as a useless system glitch, the sentient Stone signs a smart contract to download its consciousness into the mortal avatar of Jia Baoyu. Trading the safety of eternal numbness for the sensory overload of a physical body, Baoyu logs into the simulation specifically to chase the lust, trauma, and "pain" of human existence. 1:54:17 “大观园THE GARDEN (HYPE HOUSE)” - Inside the simulation, Baoyu lives as the pampered admin of the Grand View Garden, a walled-off, hyper-capitalist echo chamber where the elite ignore the rotting code of the outside world. Surrounded by the Twelve Beauties, he gorges on digital excess and mindless consumption—literally "eating the rouge"—pushing the server to the absolute brink of failure. 1:57:45 “葬花DELETING ASSETS (THE CRASH)”- The manic party violently ends with a massive system crash, forcing Baoyu to watch the algorithmic destruction of his decadent world. Amidst the digital decay, his tragic soulmate Lin Daiyu desperately tries to save the pure, beautiful files of their love, burying her fragile memories to protect them from the corrupted stream before she is permanently deleted. 2:01:09 “白茫茫WHITE NOISE (LOGOUT)”- The corrupted simulation undergoes a complete format, wiping out the Jia family's vast digital empire and leaving nothing behind but a stark, quiet white wasteland.