Hey folks! Hope you had an excellent Fourth of July (that statement holds worldwide) and are having an incredible summer.
I've got some slick new tools for you:
- Industry Performance Screener - this adds the Fama-French 12 Industry Classifications to the SIC and NAICS tools I've been working on recently.
- Security Correlation Calculator - compare daily return correlations between stocks, mutual funds, and ETFs.
Standard caveats and disclaimers apply; please let me know if you notice any issues.
These are some of the more complex (but high-demand) tools I've developed, especially the industry screener. I've got the data pre-computed nightly for that one, so hopefully it isn't too slow if you're looking at something like Technology or Finance, haha.
The correlation tool allows up to ten tickers, which I hope is a good compromise. I appreciate all the suggestions on that one; you all really came through after the last Weekender. I soft-launched that one a few weeks back, and it's already getting hundreds of uses a day – thanks for trying it out!
We do a little enhancing
Sticking to the security tools, I have some updates for you:
- The S&P 500, Dow Jones, NASDAQ, and Treasury Return Calculators (and their derivatives) should now update automatically five times a week. Dividend data lags a few months, though (that part – and earnings/sales data – is still manual).
- The Stock, Mutual Fund, and ETF return calculators now let you bookmark individual tickers and have new dividend treatment options (including "ignore dividends"/price returns). Open the Advanced Options to play with those. The URL and Title should superficially change when you run a calculation on a ticker. You can bookmark (or share, I appreciate it!) that URL to quickly return to a company's, uhh, return page.
- I recovered the Gold, Platinum, and Silver return calculators. You might laugh when you see how it works (check out the new methodology), but it's been effective. I'm still looking for a source for Palladium.
Send me your ideas!
AI may or may not replace us/end the planet/force us into clinical trials/paper clip maximize, but as of today, it can certainly help me ship ideas to DQYDJ quicker.
Lean into the current milieu and contact me if you have ideas for enhancements or new tools. As you've seen, I'm especially focused on end-of-day security pricing data or quick finance tools – I'm not trying to compete with CNBC here!
We're a few months out from the annual updates to tools and posts on US income, and I don't expect we'll see a net worth microdata drop from the SCF until next year. So I'm all ears (and eyes and fingers and whatever).
PK's house
We're getting there, folks. Most of _my_ recent projects have been of the checklist type - fix that light, move that chandelier, repair that trim, hang that door, and the like. We do have some work to do on one side of the house, but there's an end in sight!
Even the hilariously drawn-out deck project is finishing up (I know, Don't Quit Your Day Job.).
After reinforcing and replacing one-quarter of the joists and half of the deck boards, we worked on some visual improvements. The stain, new vinyl rail, and deck lighting are now a reality. The last step for us will be drink rails, but they can wait for a few weeks (the rails are tall enough without them!).
Those gutters are new (and the roof, too). We hired that out; I'm scared of heights!
My wife and daughters say "hi!". Sports season just switched up; it's volleyball and gymnastics for us now.
AI is software, software eats everything, etc...
Behind the scenes, AI and I have been busy tackling a lot of those "I should really..." projects together.
My current setup is Claude Code inside Cursor. Claude Code helped me tackle two longer-horizon projects I've wanted for a long time: expanding the test suite on the security tools (including the ones I mentioned above), and implementing a cron job monitor for DQYDJ.
Previously, I would sweat every new security tool deployment. I was hitting my own cognitive limits for how everything fit together. There's no way I would have shipped the correlation and industry performance tools this quickly without AI assistance. And, if you've been following the discourse around "vibe coding", I have been using AI tools much the way Andrej describes.
I'm no longer Googling CSS rules or combing through the Bootstrap Docs when working on DQYDJ tools... I'm complaining in the Cursor chat and dropping a screenshot (sometimes with a red arrow on it if I'm feeling particularly frisky). I'm not building spreadsheets with every pro and con of every style of testing, I'm switching Claude Code to Plan Mode and asking it the best way to test new features and tools, then picking the one that sounds best to me.
Limits? Sure! There are many!
I mostly agree with Dwarkesh Patel's and my former colleague Miguel Grinberg's negative takes on LLMs and Agents, especially around continuous learning. Current AI tools aren't – contra Twitter/X's takes – a brilliant, well-read intern; agents and LLMs lose context quickly and don't continuously learn your preferences and style. This can be extremely annoying, to put it mildly. Interns would pick up on those harder-to-score traits, such as taste, that seem to be lacking in current LLMs and agents.
Claude Code tries to get around these limits by "summarizing" (compacting) your previous conversation into text, maintaining "memory"" (which you can use directly, and I have been manipulating after extremely annoying behavior: "Stop putting emojis in debugging statements"), and using a Claude.md file with broad instructions.
But these do feel like hacks.
There's a Chekhov's Gun element in work experience, where seemingly insignificant details end up mattering in future projects. And the summaries these agents are using to simulate gaining experience are designed only to remember the most significant (at the time) details. I often find myself in a good groove with Claude, nervously watching the "X% until compaction" ticker under the tool, only to lose said groove right after the compaction.
But for me, and for some of the use cases that matter to me both on this site and at Twilio, AI is already useful. Rewind to 2020, and there's no way I would have guessed how quickly AI would become helpful in working on my interests (and, ahem, working on my actual work at my day job). If you'd asked me then, I'd have predicted that I still wouldn't have cron job monitoring in place by 2030.
Shorter PK: where we sit today, in July 2025, AI is already useful. Turn off the AI scaling tomorrow, and I'll still get a lot of use out of it. I bet you can, as well.
