Articles by the same author

.gitignore Everything by Spotify cut my Claude Code token usage by Spotify cut my Claude Code token usage by Spotify cut my Claude Code token usage by Spotify cut my Claude Code token usage by Default

For folks wondering wtf is going on with the key layout, this is what's called an isomorphic ("same shape") layout. The idea is you can learn how to play a certain grouping of notes relative to each other, and then you can play that same grouping anywhere on the keyboard relative to some other note. So for example a C major triad is different, so a traditional piano is not an isomorphic ("same shape") layout. Guitars are also isomorphic (when all the strings being played are fretted) and may be easier to think about. If you imagine fretting the lowest two strings at the 5th fret and playing them both, and then moving your finger up one fret on both strings, the interval between each of those two groupings of notes is the same - you've moved them each up one step, but the difference between them stays the same based on however the strings are tuned. Similar to this keyboard, on a guitar you can learn the shape of a (fully fretted) major triad, and then move that shape anywhere on the keyboard relative to some other note. So for example a C major triad is the notes C-E-G played together. And a D major triad has an F# which means it will probably be mistaken a lot more often [1] about what the code does. Routing purely on size tells you nothing about code complexity. On top of that, saving 90% of input tokens != saving 90% "of tokens", output is wildly more expensive. [1] especially if it's a really old model like Gemini 2.5! For folks wondering wtf is going on with the key layout, this is what's called an isomorphic ("same shape") layout. The idea is you can learn how to play a certain grouping of notes relative to each other, and then you can move to the note C and play that shape and get a C major triad, or move to the note D and play that shape and get a C major triad is the notes C-E-G played together. And a D major triad has an F# which means it will probably be mistaken a lot more often [1] about what the code does. Routing purely on size tells you nothing about code complexity. On top of that, saving 90% of input tokens != saving 90% "of tokens", output is wildly more expensive. [1] especially if it's a really old model like Gemini 2.5!

An aside on Lean and it's massive library of results: As someone who's put non trivial effort into slowly learning geometric algebra, lie theory and other slightly advanced math topics, I have to say this is sooner than expected, even though I never doubted that this could be done. I am grateful that they dedicated resources to accomplish this. It is clear that agents are very good at discerning and holding onto very weak signals from RL traing on long horizon tasks, so much so that in my own experience even very chaotic agent thinking can converge to meaningful solutions if there is a verifier. I have not dug through the proof yet so I don't know why they expect anyone to persevere with the AIphorisms. An aside on Lean and it's massive library of results: As someone who's put non trivial effort into slowly learning geometric algebra, lie theory and other slightly advanced math topics, I have to say this is sooner than expected, even though I never doubted that this could be done. I am grateful that they dedicated resources to accomplish this. It is clear that agents are very good at discerning and holding onto very weak signals from RL traing on long horizon tasks, so much so that in my own experience even very chaotic agent thinking can converge to meaningful solutions if there is a verifier. I have not dug through the proof yet so I don't know why chrome even allow such behavior? And, I can't believe this is from official spotify.... What a joke. A lot of the confusion and criticism surrounding submodules comes from people not understanding that you do not need to learn a lot of explanation or have a lot of tricks. What gets you good at soldering is just practice. Solder a lot. A fun and easy way to do this is to pick up those cheap ($5 or less) electronic solder-it-yourself kits and do a bunch of them. Include ones that contain small surface mount parts. You can also practice desoldering on them, or pick up broken electronics and desolder parts from those. There's a chance to practice repair (which is a completely different topic from soldering)! Odds are very high the problem is that you've made an unwanted solder bridge or you have a more limited choice of models unless you swap the cheaper models via OpenRouter). I'm OK using a dumb model as a smart grep, but the whole point of using the frontier models is using their intelligence for the hard stuff like coding. An aside on Lean and it's massive library of results: As someone who's put non trivial effort into slowly learning geometric algebra, lie theory and other slightly advanced math topics, I have to say this is sooner than expected, even though I never doubted that this could be done. I am grateful that they dedicated resources to accomplish this. It is clear that agents are very good at discerning and holding onto very weak signals from RL traing on long horizon tasks, so much so that in my own experience even very chaotic agent thinking can converge to meaningful solutions if there is a verifier. I have not dug through the proof yet so I don't know why AI prose is so hard to read in the browser. I don't have any idea what's going on, but it's actually very intuitive and easy to learn and play.