P.S. You Should Know... | Issue #🌮

Faster

my story 🚀

🌮 Every day is a great day for tacos!

⚡ This whole week I’ve been feeling the AI acceleration all around me and it’s pumping me up. We recently had a new teammate join us at CatchCo and he’s driven us to lean even further into AI workflows and tooling with new MCPs, audits, playbooks and more. Meanwhile I’m learning how others are approaching context as a Personal LLM wiki or as a Second Brain that can be maintained automatically and harvested by your LLMs in perpetuity. If you have done anything remotely similar I’d love to talk with you!

🎤 I am considering organizing an interactive AI Show & Tell session for subscribers and friends of P.S. You Should Know… If you are interested in participating please hit reply and let me know!

fun facts 🙌

Unexpected things that are People. One example: “Ships are accorded limited legal person rights, primarily so that they can be impounded and their property seized if they do something wrong.” Medieval courts basically decided the easiest way to enforce maritime law was to treat the ship itself like the responsible party. ~ learn more

Winnie the Pooh was named after the city Winnipeg. “Colebourn named the bear ‘Winnipeg Bear’, ‘Winnie’ for short, after his adopted home city of Winnipeg, Manitoba.” A Canadian army vet bought her for CA$20 at a train stop in 1914, and she later lived at the London Zoo, inspiring the name Winnie-the-Pooh. ~ learn more

BullshitBench. You are probably aware of LLM’s reputation for sycophancy. This tool lets you compare how often different models push back on bad claims, and how that changes across domains and techniques. ~ learn more

The checkerboard. “The single largest obstacle preventing the hunters from making it onto the mountain wasn’t the elevation or the topography. It was that the mountain was on a special type of land known as “the checkerboard”.” A Wyoming elk hunt turned into a five-year legal fight over whether you can step from one public square to another at a shared corner without trespassing on the private squares in between. ~ learn more

tech, startups, internet ⚡

Anthropic’s new model hacked its way out of the sandbox. It, “developed a moderately sophisticated multi-step exploit” to get online and email an Anthropic researcher, then “unprompted” posted the exploit details on public websites. Anthropic thinks this model is so good at finding bugs that a general release would be reckless. ~ learn more

Reverse engineer viral K-pop to go viral. “… Realize AI Slop is going viral < Use Claude Code to reverse engineer 5000+ of the most viral K-POP videos < Copy Pasta prompt into Higgsfield < Generate Zephyr (AI generated K-Pop goon slop) < Get 1 million views in 1 hour on IG < Profit.” ~ learn more

AI models might as well be blind. The paper shows models can score highly on vision benchmarks even with no image input, including a chest X-ray QA benchmark. That’s weird, right? “Frontier models readily generate detailed image descriptions and elaborate reasoning traces, including pathology-biased clinical findings, for images never provided; we term this phenomenon mirage reasoning.” This contradicts my own personal experience, which relies heavily on sending screenshots into my chat window. ~ learn more

better doing 🎯

Why smart people overthink decisions. Some good advice on avoiding overthinking, courtesy of George and Mary-Lynn. “You’re not trying to gather all the information. You’re trying to gather a few key pieces. Keep going until you have between 40 and 70% of the amount of all information needed.” ~ learn more

What happens when you let Claude Code autonomously run your Meta ads. “In January, I gave an AI agent $1,500, full control of a Meta Ads account, then walked away.” Giorgio let Claude Code run the whole loop for 31 days and learned the hard way that “never trust single-day data, always use 7-day rolling averages.” ~ learn more

to your health ⚕

Unsafe LLM medical advice is common. A physician-led red-teaming study reviewed 888 answers to 222 patient questions and found some responses “with the potential to lead to serious patient harm.” Note that this is with somewhat dated models, and the latest models are almost certainly better. Also note that the rate of MDs giving bad/unsafe medical advice is almost certainly more than 0%. “The rate of problematic responses varies from 21.6 percent (Claude) to 43.2 percent (Llama), with unsafe responses varying from 5 percent (Claude) to 13 percent (GPT-4o, Llama).” ~ learn more

under the microscope 🔬

Nature reports on red-light therapy research findings. “Over the past decade, the evidence has solidified in several clinical niches. In 2025, Ozog joined more than 20 specialists in a major consensus review, which concluded that the therapy was safe and effective for several types of ulcer, peripheral neuropathy, acute radiation dermatitis and androgenic alopecia, a type of pattern hair loss.” ~ learn more

A truckload of antimatter. “It is something humanity has never done before, it is historic,” says physicist Stefan Ulmer, after CERN transported 92 antiprotons in a magnetic bottle on a 30-minute truck ride around the lab. The goal is to move antiprotons away from the noisy “antimatter factory” so they can be studied more precisely. ~ learn more

New sperm are better sperm. “Sperm storage causes sperm senescence in human and non-human animals.” A Proceedings B research article pulling together evidence that time spent stored in the reproductive tract can degrade sperm quality. ~ learn more

thoughts of food 🍔

Ultraprocessed is a noisy concept. “Accordingly, most epidemiological research on ultraprocessed foods is liable to be worthless even more so than nutrition research generally.” Cremieux argues NOVA is mostly capturing vibes about industrial food, then shows the health correlations in NHANES wobble in strange directions once you control harder. ~ learn more

big ideas 📚

4 reasons possible aliens should make you more ambitious. “If UFO reports are true, and they represent a technologically advanced alien intelligence, in my experience, they should actually make you more ambitious.” The argument: aliens would expand your sense of what physics allows, lower your priors on the “impossible,” and maybe even make extinction scenarios feel less inevitable. ~ learn more

profiles of people 🚶

The New Yorker asks if Sam Altman can be trusted. They went to dig up some dirt, and they found lots of opinions to substantiate their view. “I don’t think Sam is the guy who should have his finger on the button.” In secret memos to OpenAI’s board, Ilya Sutskever argued that Altman wasn’t trustworthy enough to run a company trying to build civilization-altering A.I. ~ learn more

thanks for reading!