It is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure right here.
Firstly of April, the outstanding AI researcher Andrej Karpathy tweeted out an concept that was, for a sure form of productiveness nerd, an infohazard. He described the thought this fashion: “One thing I’m discovering very helpful lately: utilizing LLMs to construct private information bases for numerous matters of analysis curiosity.” He proceeded to explain his course of: including supply paperwork to a neighborhood folder and utilizing an LLM to extract and arrange their contents right into a Markdown wiki that will get up to date as he provides new materials.
Karpathy’s AI-related pronouncements are carefully adopted on-line — he’s, in any case, the one who coined the time period “vibe coding.” And so, seemingly inside hours, the online had stuffed up with GitHub repos, YouTube movies and Substack posts about learn how to arrange an LLM wiki for your self. I name the thought an infohazard as a result of, just by changing into conscious of it, I had ensured that I might commit the following a number of weeks to constructing it, with out having any concept whether or not it might profit me in any respect.
Because it so occurs, it has. Of every thing I attempted this yr to get higher on the deskbound components of my job, the LLM wiki has simply been essentially the most helpful. Like a classic sports activities automotive, it requires a good diploma of upkeep, and there are virtually actually simpler methods to create a private information base. However in case you do any form of work that has made you crave the assistance of a very good analysis assistant, an LLM wiki is perhaps value your time.
The wiki is the centerpiece for my annual productiveness publish, the place I run down any modifications to the way in which I work that is perhaps fascinating to others troubled by an inexplicable enthusiasm for software program. (Listed below are the posts from 2023, 2024, and 2025.)
So with that, right here’s what I’m nonetheless doing from final yr, what I finished doing, and what’s new.
What I am nonetheless doing
Given how typically I change apps, the most effective check for whether or not one thing truly makes me extra productive is longevity. Do I set up the app on a brand new machine? Do I renew the subscription when it’s time? Can I level to the locations the place it truly saves me time?
Three apps I’ve really helpful in earlier years nonetheless clear that bar.
Raycast, a launcher app that replaces Highlight, stays my most well-liked approach to navigate a pc. Once I press ⌘-space, the Raycast window immediately materializes and lets me carry out actions throughout a lot of the apps that I take advantage of. I lookup phrases; I do math; I reposition home windows; I entry my clipboard historical past; I open web sites. And due to a paid improve, I do tons of straightforward AI searches within the Raycast window. (I take advantage of GPT-5.5 Prompt right here for the excessive quality-to-speed ratio.) Once I’m executed, there’s no getting misplaced in a jumble of open tabs or home windows — Raycast merely fades again into the background. At this level, I actually can’t think about my Mac with out it. And in a pleasant growth since final yr, it’s now out there for Home windows as effectively.
Capacities, which payments itself as “a studio on your thoughts,” is the place I hold my each day journal. Every morning, I write a bit about no matter’s on my thoughts. Then, I add notable information hyperlinks from Techmeme to the underside of my journal and tag them. The result’s that after I’m writing a narrative or getting ready for a podcast, I can click on a tag like “Labor” and immediately see all of the tales I’ve saved on that topic since I began constructing this technique a pair years in the past. This has been enormously useful in planning our present podcast miniseries on AI and productiveness, since we focus on jobs information on every episode. I used to spend so much of time digging via databases or operating fruitless Google searches in an effort to jog my reminiscence about one thing I had learn; Capacities ensures that it’s all only a click on away. You could possibly simply do that with any variety of apps, however after three years I nonetheless discover myself appreciating the simplicity and calm of Capacities.
Lastly, final yr I discussed testing an app known as Recall that (anticipating the LLM wiki!) helps you save and arrange content material from the online. I discovered myself utilizing it much less for that objective over the previous yr, aside from one remaining killer use case: its Chrome extension offers near-instant textual content summaries of YouTube movies. Over the previous yr, I’ve saved myself many hours by dumping podcasts I really feel like I ought to hearken to for work into Recall and easily skimming the summaries.
What I finished doing
The 2 issues I finished doing over the previous yr are associated to one another — and to the LLM wiki. For years now, I’ve been searching for an answer to issues of reminiscence. I’ve been a tech reporter for nearly 16 years, have written a e-newsletter for nearly 9, and have written this article for six. For a lot of that point, I’ve been publishing tales and saving analysis supplies in numerous locations. And when a information story comes alongside that pulls on a few of that historical past, I wish to discover that context as shortly as potential.
Final yr I talked about how Notion had shipped a characteristic I’d wished for years: an agent that would search throughout the hundreds of hyperlinks I had saved into it through the years. It labored effectively sufficient, however I couldn’t flip it right into a behavior. The search characteristic within the database itself is a straightforward keyword-based search; agentic search takes place on one of many app’s many different surfaces; and the agent typically didn’t cite its sources with out extra prompting. It labored, but it surely was effortful.
Furthermore, it solely regarded backward. I additionally had a necessity for a system that will assist me arrange tales round new ideas, and at scale. In Capacities, I took to creating pages I known as “blips” (impressed by Andy Matuschak) the place I might collect unfastened string. For instance, in the summertime of 2024 I created a blip known as “AI might create large job loss,” and added related hyperlinks to tales about AI and jobs as I encountered them throughout my each day journaling.
For a number of months I felt extraordinarily intelligent, as a result of I created a dynamic object within the template for my each day journal in Capacities that will present me a number of blips at random. This is able to frequently remind me of blips that I had forgotten to replace, and labored higher than any system I had beforehand devised for monitoring long-term tales.
Finally, although, even this technique requested a bit an excessive amount of of me. Because the variety of blips proliferated, my skill to persistently observe tales throughout all of them waned. This, in flip, made me extra reluctant to create new blips.
And that’s why, after I noticed Karpathy’s tweet, I sat just a little straighter in my chair. What if AI might write and replace all these blips for me?
The LLM wiki
Actually, crucial change in my productiveness over the previous yr is that … I make software program now? Like seemingly everybody else on the finish of final yr, I started messing round with Claude Code, and have since made a small handful of apps that I actually do use on a regular basis. Generally I take advantage of Glaze, one other app from the maker of Raycast; it excels at design and polish.
For the LLM wiki, although, I simply requested Claude Fable 5 to jot down me a immediate that will get me a “Karpathy-style LLM wiki.” I pasted the outcome into the terminal (I take advantage of Ghostty), and earlier than too lengthy I had created a brand new folder of Markdown recordsdata in Obsidian.
One factor that makes the wiki significantly invaluable for me is that I first seeded it with my very own writing: all the Platformer archive, from which Claude expertly extracted all the varied individuals, firms, and ideas that I’ve coated right here since 2020 and wrote them up in Markdown recordsdata that dwell on my laptop.
These recordsdata can get fairly lengthy; my web page for Meta runs to greater than 12,000 phrases and comprises greater than 1,300 hyperlinks to different pages within the wiki. Everyday, that isn’t of a lot sensible use. However I’ve misplaced a couple of afternoon searching the archive in the identical state of blissed-out curiosity that I browse Wikipedia, remembering outdated tales and reasoning about how they match into present occasions.
However I wished greater than a wiki of my very own work. And so now every morning after I journal, I save a number of tales into the wiki through Obsidian’s internet clipper, which converts them into Markdown. Then a script on my laptop reads the tales and figures out the place they match into the wiki.
The result’s that I now have greater than 1,440 wiki pages masking most of what has ever me at Platformer: from the content material moderation focus of the primary few years to my rising curiosity in baby security and AI progress. The wiki creates detailed timelines that hyperlink to unique sources, and I can ask it questions utilizing a easy Obsidian plugin named Claudian.
How does this make me extra productive? I’ve discovered it extremely helpful in fast-evolving, advanced instances just like the OpenAI / Hugging Face agentic breach, the place we realized just a little extra in regards to the story each few days for a matter of weeks. Every day as I ready to jot down, or podcast, or go on another person’s podcast, I might pull up the web page and refresh my reminiscence — whereas additionally opening up the unique sources to ensure nothing I used to be about to say was hallucinated. Given the true complexity of that story — how precisely did the agent escape? From the place? To the place? — this saved me tons of time.
It additionally provides me helpful story concepts — and it does it by automating my outdated blips system. Every day because it reads, the wiki generates new pages for ideas within the information — like “tokenmaxxing,” or “youth social media bans,” or “AI copyright.” It additionally updates a “house” web page each morning that highlights tales within the information. This week, recognizing the “AI and Congress” idea in my wiki led me to open the web page and see plenty of latest tales on the topic; I later pitched it as a podcast phase.
All of that’s nice. What’s much less nice is that the wiki wants roughly fixed upkeep. Pages develop too lengthy and must be compacted. An error within the code implies that one course of or one other stops operating and needs to be fastened. And Claude’s hyper-compressed, borderline-unreadable home model (see this tweet and this one) led me to make use of GPT-5.6 Sol to rewrite a lot of the system to extra carefully approximate AP model. (It did an excellent job.)
It has now been simply over a month since I created the wiki, and I discover myself wanting ahead to checking in on it within the morning to see what it has constructed. On one hand, it appears too particular to me and too clunky for me to confidently suggest. On the opposite, “self-organizing information base” strikes me because the teleological finish of no matter course of started the day I first put in Evernote on my Mac in 2008.
That’s one motive why I sought out City CEO Jean-Denis Greze for an interview final week; one mind-set about that product (or what it would evolve into) is a form of Karpathy LLM wiki for work. As ever, the worth of utilizing these instruments on the bleeding edge is that I arrive in all places a lot too early. The payoff is that, by some means, I’m having fun with myself rather a lot.
A MESSAGE FROM OUR SPONSOR
The shared card period is over
Mercury Spend is expense administration constructed into Mercury. Give your workforce and brokers playing cards, set spending limits per individual or workforce, and the insurance policies implement themselves. Receipts arrive routinely from Gmail or through textual content, and playing cards with overdue duties auto-lock till resolved. No end-of-month scramble.
Discover Mercury Spend
*Mercury is a fintech firm, not an FDIC-insured financial institution. Banking companies supplied via Alternative Monetary Group and Column N.A., Members FDIC. The IO Card is issued by Patriot Financial institution, N.A., Member FDIC, pursuant to a license from Mastercard Worldwide Included.
Following
OpenAI’s security reset
What occurred:
OpenAI launched “ChatGPT for Teenagers,” a brand new mode with safeguards designed to guard teen customers. The corporate introduced work on the characteristic final September, after they’d been sued by plenty of households of teenagers who had died by suicide following conversations with ChatGPT.
OpenAI additionally introduced new measures to guard grownup people (together with themselves) from its personal merchandise. The corporate says its new fashions might need reached a “vital” degree of cybersecurity capabilities — which amongst different issues will increase the chance their brokers will escape containment and hack into different firms, as occurred lately.
The corporate stated in a weblog publish that it paused coaching on its newest era of fashions for the previous two weeks because it improved safety methods. OpenAI may even postpone a deliberate massive coaching run till it has executed extra assessments of the methods’ security.
Why we’re following: The corporate’s heightened give attention to safeguards within the wake of high-profile security incidents is welcome, even whether it is arguably overdue.
ChatGPT for Teenagers will apply safeguards “designed to cut back publicity to content material which may be dangerous or developmentally inappropriate” for customers beneath 18. It’ll construct on OpenAI’s earlier parental management options that enable caregivers to set quiet hours and get alerts about regarding exercise.
“ChatGPT mustn’t use romantic language, encourage emotional dependence, or suggest that it has emotions or consciousness,” the corporate stated.
The characteristic takes an identical strategy to under-18 accounts created by social media firms in response to baby security issues. (Ask social media firms how that is going!) And as with teen accounts, the safeguards will be evaded, as OpenAI stated themselves in a weblog publish a number of months in the past: “Guardrails assist, however they’re not foolproof and will be bypassed if somebody is deliberately attempting to get round them.”
Have been I a guardian, I’m undecided I’d belief my teen round a mannequin with “vital” cyber capabilities. So let’s hope OpenAI retains taking the time it wants to make sure it’s protected!
What individuals are saying: On X, Ex-OpenAI board member Helen Toner permitted of OpenAI’s determination to pause coaching. “Actually good to see this,” she wrote. OpenAI’s coverage, she wrote, is “by far one of the best ways to consider “pacing the frontier”— not as some fastened period of time (e.g. ‘six month pause’ or ‘go 10% slower’), however merely ensuring that sufficient time is taken to fulfill an affordable security/assurance bar.”
—Ella Markianos
Meta’s social media dependancy trial opens (once more)
What occurred: 4 states — California, Colorado, Kentucky, New Jersey — went up towards Meta at the moment in a landmark trial in California over allegations that Meta designed its platforms to be addictive for younger customers and misled the general public in regards to the dangers.
In opening arguments, Megan O’Neill, deputy lawyer basic at California’s justice division, stated Meta’s enterprise mannequin might be summed up as follows: “Hook the customers; maintain them for so long as they will; harvest their knowledge; conceal the reality from the general public when making public statements.”
Meta knew in regards to the addictive results its platforms had on youngsters’s psychological well being, O’Neill stated, regardless of the corporate telling Congress and oldsters that Instagram and Fb had been protected. “All of the whereas it hid the reality,” she stated.
The primary witness of the case, whistleblower and a former chief of Fb’s Integrity and Care workforce Arturo Béjar, testified that Meta’s metric for measuring hurt was “very slim” and that the corporate under-reports the prevalence of dangerous content material.
Meta lawyer Paul Schmidt argued that the corporate has not focused youngsters and stated he’ll current knowledge that exhibits greater than 90 p.c of Fb customers are adults (he conceded Instagram skewed youthful, however stated the information was related). Plus, he pointed to previous statements from Mark Zuckerberg and Instagram head Adam Mosseri stating the platforms weren’t designed to be addictive, and that analysis hasn’t but supported the idea of social media dependancy.
Why we’re following: Meta has been going through an onslaught of lawsuits over baby security. And it’s shedding.
Only a few days in the past, a New Mexico court docket ordered the corporate to pay $567 million on high of the preliminary $375 million in civil penalties to fund the state’s abatement plan after the corporate was deemed to have contributed to a teen psychological well being disaster. In that trial, the decide known as Meta a “public nuisance” akin to air air pollution.
Individually in March, a Los Angeles jury discovered Meta and Google chargeable for a younger lady’s despair and anxiousness after she compulsively used social media as a baby.
This landmark trial might additionally imply potential modifications to Meta’s platforms, together with a change in advice algorithms, parental verification, and the tip of “autoplay” for video content material.
What individuals are saying: Kentucky lawyer basic Russell Coleman was the most recent to check this trial to the states’ success towards tobacco firms within the Nineties: “AGs are within the excellent place to get this executed. We did it with the Tobacco Settlement within the Nineties…We’ll do it once more with Meta.”
Meta, for its half, has persistently denied the allegations. Schmidt painted the corporate as one which has truly handled its points: “There will be no dispute that Meta has acknowledged individuals wrestle, or can wrestle, with their use of social media, and has provide you with instruments to try to handle that.”
Meta’s motto was to “transfer quick and break issues,” California lawyer basic Rob Bonta stated. “Sadly, the factor that was damaged right here was the psychological well being of our youngsters.”
—Lindsey Choo
These good posts
For extra good posts day-after-day, comply with Casey’s Instagram tales.
(Hyperlink)
(Hyperlink)
(Hyperlink)
Speak to us
Ship us suggestions, feedback, questions, and LLM wiki strategies: casey@platformer.information. Learn our ethics coverage right here.
