links (95) (2025-12-22)
Prompt Caching: 10x Cheaper LLM Tokens
That's it, those K and V matrices above, they are the 1s and 0s that the providers save in their giant datacenters to offer us 10x cheaper tokens, and much faster responses.
Providers hold on to these matrices for each prompt for 5-10 minutes after the request is made, and if you send a new request that starts with the same prompt, they reuse the cached K and V rather than recalculating them. What's really cool is that you can partially match a cache entry and still use the bit that matched, not the whole thing.
Noclip - Video Game Map Viewer
A digital museum of video game levels
CLICK AND DRAG to look around and use WASD to move the camera
In this large, prospective study of US adults aged 40 years or older, eating one meal per day was significantly associated with an increased risk of all-cause and CVD mortality compared with eating three meals per day. Skipping breakfast was associated with increased risk of CVD mortality, whereas skipping lunch or dinner was associated with higher risk of all-cause mortality.
MerkleMap collects Certificate Transparency data by continuously monitoring and live tailing the CT logs.
Rattlin' Bog feat. the Tree of Wisdom (video)
I wanna learn to dance like that :(
Screenshots from Developers: 2002 vs 2015
Richard Stallman - 2002: I don’t know how to make a screenshot, because I normally use my computer in text-mode.
The Food Lab's Chocolate Chip Cookies
Corn syrup is so darn powerful, in fact, that even a small amount of it will completely alter the texture of your cookie. In the cookies above, the batch on the left was made with 5 ounces each of granulated and brown sugar. The batch on the right was made with 5 ounces of brown sugar, 4 ounces of granulated sugar, and 1 ounce of corn syrup — a substitution of only 10%.
Well, it all comes down to three perfectly synergistic events:
- OpenAI executed two unprecedented RAM deals that took everyone by surprise.
- The secrecy and size of the deals triggered full-scale panic buying from everyone else.
- The market had almost zero safety stock left due to tariffs, worry about decreasing RAM prices over the summer, and stalled equipment transfers.
links!
The real prediction is that the diagnostic task in radiology will be performed by AI, eliminating the need for human intervention. That is within Hinton’s expertise. The prescription that we should stop training radiologists has a logical flow from that prediction, but many other factors are outside of Hinton’s expertise. I have seen all too often that scientists often take the things they know and extrapolate those into calls for action that they don’t fully understand.
"xchg rax,rax" is a collection of assembly gems and riddles I found over many years of reversing and writing assembly code. The book contains 0x40 short assembly snippets, each built to teach you one concept about assembly, math or life in general.
Anthony Bourdain's Lost Li.st Archive
Anthony Bourdain published about 30 lists on the defunct li.st web site around 2015. This page presents a partial archive of those, recovered from the Internet Archive.
40% of people answering this survey have been using the terminal for 21+ years
95% of people answering the survey have been using the terminal for at least 4 years
I see a lot of engineers run into a weird thing - commonly a 403 or 400 status code from some other service - and say “oh, I’m blocked, I need this other service’s owners to investigate”. You can and should investigate yourself.
Understanding The Player Brain, Pt. 1: Loss Avoidance
They designed in a fatigue mechanic. After you played two hours, you got smaller rewards for your efforts.
Early players hated this. HATED it.
So they fixed the system. They didn’t actually change anything about the numbers of the system. Just the labels. Instead of saying when you played too long you were tired, they said that if you logged out for a while you became “rested.”
Yes, I use Dropbox and GitHub to hold all the data that I care about, but the beauty of these systems is that they work with local copies of that data, so with a couple of computers here and there, I always have a recent version of everything, in case either syncing service should go offline
Some projects I worked on... A slide rule for salesmen to estimate prices on site, instead of making clients wait until the salesman could talk to engineering.
Think Deep Research for GitHub.
We Should All Be Using Dependency Cooldowns
In the very small sample set above, 8/10 attacks had windows of opportunity of less than a week. Setting a cooldown of 7 days would have prevented the vast majority of these attacks from reaching end users (and causing knock-on attacks, which several of these were). Increasing the cooldown to 14 days would have prevented all but 1 of these attacks.
The f32 is an ultra-compact ESP32 development board designed to mount directly behind a USB-C receptacle.
Modern Hardware Numbers for System Design (2025)
Single PostgreSQL or MySQL instances now handle dozens of terabytes while maintaining millisecond-level response times. They'll process tens of thousands of transactions per second on a single primary. You can push single-instance databases much further than conventional wisdom suggests. While the largest tech companies still need sharding, most applications can run on a single well-tuned database.
What's in a Passenger Name Record (PNR)?
PNRs are airline records, but few airlines host their own databases. Most airlines store their PNRs in a virtual “partition” in the database of a Computerized Reservation System (CRS).
The source code was contributed anonymously and represents a snapshot of the Infocom development system at time of shutdown - there is no remaining way to compare it against any official version as of this writing, and so it should be considered canonical, but not necessarily the exact source code arrangement for production.
ACT-1: A Robot Foundation Model Trained on Zero Robot Data
By ensuring the glove and the robot hand share the exact same geometry and sensor layout, we eliminate the translation gap entirely. The promise is simple: If a human can do it in the glove, the robot can also do it.
Good Engineering Management Is a Fad
The conclusion here is clear: the industry will want different things from you as it evolves, and it will tell you that each of those shifts is because of some complex moral change, but it’s pretty much always about business realities changing. If you take any current morality tale as true, then you’re setting yourself up to be severely out of position when the industry shifts again in a few years, because “good leadership” is just a fad.
LibrePods unlocks Apple's exclusive AirPods features on non-Apple devices. Get access to noise control modes, adaptive transparency, ear detection, hearing aid, customized transparency mode, battery status, and more - all the premium features you paid for but Apple locked to their ecosystem.
Only three kinds of AI products actually work
Chatbots, Completion, and Agents