a fun thing to do for papers with good source code available is to git clone the paper repo, drop the paper pdf in, and then tell claude to, for example, write an interactive opengl visualization of Figure 4 (adding the plot from Figure 2 on top) with sliders for playing with the parameters
Magnetic Fields: Find a central metaphor that's so good that everything aligns to it. Design meetings are no longer necessary, it designs itself. The metaphor should be crisp and fun. Smalltalk is object-oriented, but it should have been message oriented. Snobbery: Turn up your nose at good ideas. You must work on great ideas, not good ones. Appreciate mundanity: after all, a pencil is high technology One goal: the computer disappears into the environment The computer shouldn't act like it knows everything. The whole notion of 'programming language' is wrong.
Combobulating...
vibecoded signal generator, now without the sine peaks cut off
it's not just programming any more. Claude Opus 4.6 is a pocket mathematician. an unreliable one, but experts can get real results from these things, it seems.
That's a lot of proofs:
> In two weeks, Gauss then autoformalized the 24-dimensional case using only the original paper as input, performing autonomous literature searches when needed. This brought the total sphere packing formalization from 70k to ~200k lines.
> In just 5 days, Gauss automatically proved all remaining results needed to verify the result in 8 dimensions. The Sphere Packing team estimated that the 8-dimensional case alone would have taken six more months of work with existing tools.
www.math.inc/sphere-packing
but then you have the problem of making some backend system that categorizes all the possible topics that people are posting about. this is at minimum LLM-complete, and doing this at bluesky scale seems to require a lot of infrastructure!
Your feed was then often things about the topics you followed, but somethings adjacent topics that the recommendation algorithm thinks you may also like. in this model, I suppose the idea would be to extract topics from a user's bluesky interactions and blog posts.
Screenshot from Maven's page for the "RLHF" tag, showing other users following that topic and what other topics they follow
yeah I think "how do you register interest in topics in this system" is not fully worked out in my mind. On maven (I discovered it's still up, app.heymaven.com) every post has topics extracted automatically, and then each user had to follow individual topics, and you could see what everyone followed
Kenneth Stanley talks about this in www.youtube.com/watch?v=73sv..., the phrase "what is transformative to you cannot be decided through consensus" has been rattling around in my brain ever since (esp. as I browse For You, unfortunately). I wish atproto was more mature when he'd tried making this
added a voltage vs time plot and Gaussian signals
a custom visualization, mostly vibecoded. still needs work, but source here: tangled.org/oscillatory...., and i deployed a version here: dev.oscillatory.net/tline-viz/
impedance discontinuities
still bouncing around
Because of the scaling work, I became completely absorbed with how the exponential increase in complexity of integrated circuits would change the way that we think about computing. The viewpoint of the computer industry at the time was an out- growth of the industrial revolution; it was based on what was then called βthe economy of scale.β The thinking went this way: A 1000-horsepower engine costs only four times as much as a 100- horsepower engine. Therefore, the cost per horsepower becomes less as the engine is made larger. It is more cost eο¬ective to make a few large power plants than to make many small ones. Eο¬- ciency considerations favor the concentration of technology in a few large installations. The same must be true of computing. One company, IBM, was particularly successful following this strat- egy. The βComputing Centerβ was the order of the dayβa central concentration of huge machines, with some bureaucrat βin chargeβ and plenty of people around to protect the machines from anyone who might want to use them. This model went well with the bu- reaucratic mindset of the timeβa mindset that has not totally died out even today. But as I looked at the physics of the emerging technology, it didnβt work that way at all. The time required to move data is set by the velocity of light and related electromagnetic considerations, so it is far more eο¬ective to put whatever computing is required where the data are located. **Eο¬ciency considerations thus favor the distribution of technology, rather than the concentration of tech- nology.** The economics of information technology are the reverse of those of mechanical technology.
the dream of the 90s
Some thoughts on AI and math, inspired by βFirst Proofβ: www.daniellitt.com/blog/2026/2/...
Phasor Agents: Oscillatory Graphs with Three-Factor Plasticity and Sleep-Staged Learning
https://arxiv.org/abs/2601.04362
"Be the Signal" I ~ v x C_L where: I = current out of our foot v = speed with which we move down the line, charging up regions C_L = capacitance per length of the line
Be the Signal
just set up a tangled.org knot for fun. you can't yet migrate an existing repo to a new knot (tangled.org/tangled.org/...), but it's easy enough for me to re-create my repos since they currently don't have state outside of git
impedance mismatch
so the kanye seeddance video didn't give me me too much motion sickness, but this one does. it's the rapid camera pull back that does it. it's been years, seems like the labs are never going to fix this problem. looking forward to a glorious future of AI-generated ads that literally make me nauseous
new end of winter metric just dropped: the day that Veronika starts removing her winter coat
β«βͺ the best part of waking up is rainbows in your cup βͺβ«
yeah, the flip side of my post above is "i am not a computer scientist, so I will not be learning git, and AI will handle all of my version control tasks". there's two LLM modes: running straight through complexity walls that I don't care about, and delving into the gaping maw of complexity i do
i think this is right. LLMs are great for overcoming incidental complexity, we just need to be careful to not to use it to avoid understanding essential complexity. it's way more important to understand eagain.net/articles/git... than it is to memorize git syntax (the CLI is not great!)
the main barrier to some of the transhumanist visions I'm seeing is that "reality has a surprising amount of detail" (johnsalvatier.org/blog/2017/re...), which must be delicately managed when you make physical things. the poor chatbots don't have high resolution interfaces to the world yet.
feedback loop diagram of an AI agent with "4. observe" circled in red. original diagram from https://www.argmin.net/p/secrets-of-intelligence-services
but what I'm seeing is reliable agentic coding requires sufficiently rich observations for tight feedback loops, which is much easier for pure software plays than for physical systems (there's a good reason why anthropic's recent team-of-claudes experiment wrote a programming language compiler).
it took me awhile to come around to this because i've been programming since i was 12, have only ever worked as a software developer, have an M.S. in C.S., etc., so I'm fairly invested.
Itβs one thing when newcomers fall into this dopamine loop and produce something. When Peter first got me hooked on Claude, I did not sleep. I spent two months excessively prompting the thing and wasting tokens. I ended up building and building and creating a ton of tools I did not end up using much. βYou can just do thingsβ was what was on my mind all the time but it took quite a bit longer to realize that just because you can, you might not want to. It became so easy to build something and in comparison it became much harder to actually use it or polish it. Quite a few of the tools I built I felt really great about, just to realize that I did not actually use them or they did not end up working as I thought they would.
it's been clarifying seeing writing like this because my reaction has been exactly the opposite: it seems like software is no longer the main bottleneck, so i've de-identified as a software developer and pivoted to learning more about hardware, which seems to me where the remaining bottlenecks are.
