iandanforth

"Every game needs a goal. In a Turing machine, goals are imposed from the outside. For true induction, the process itself should create its own goals. And there was a key constraint: Putnam realized that the dynamics he had in mind would only work mathematically if the system had just one goal governing all its behavior.

That’s when it hit him: The goal is to repeat. Repetition isn’t a goal that has to be programmed in from the outside; it’s baked into the very nature of things—to exist from one moment to the next is to repeat your existence. “This goal function,” Putnam wrote, “appears pre-encoded in the nature of being itself.”

So, here’s the game. The system starts out in a random mix of “on” and “off” states. Its goal is to repeat that state—to stay the same. But in each turn, a perturbation from the environment moves through the system, flipping states, and the system has to emit the right sequence of moves (by forming the right self-reinforcing loops) to alter the environment in such a way that it will perturb the system back to its original state."

I'm a big fan of this line of thinking. I've been arguing for years that RL should be based in homeostasis and this seems right along those lines. I wish I could have talked with him!

Man I saw the Fetch with its busted up gripper plastics and shocked to think they hadn't mothballed it and that it still worked!

Am I crazy or are the cats pictured under the graph swapped? The cat under the "leftward" bar is curled the right and the cat under the "rightward" bar is curled to the left. Did they just decide that egocentrism in directions doesn't apply to cats?

This is demonstrably false. Natural language acquisition is almost entirely listening and talking. The fastest and most consistently effective way to learn a language is immersion. The reason traditional material doesn't attempt immersive techniques is because it is much much easier to print a static book than it is to produce interactive and adaptive content.

The promise and potential of LLM based language learning apps is that you can cross that gap to full immersion in a way that has never been possible before.

Please be more ambitious.

Those outside the US at least ...

I learned from a recent post (https://sean.heelan.io/2025/05/22/how-i-used-o3-to-find-cve-...) that finding security issues can take 100+ calls to an LLM to get good signal. So I wonder about agent implementers who are trying to get good signal out of single calls, even if they are specialized ones.

Alright, having tried this with Japanese I can say it's frustrating. As a near complete beginner the tutor kept speaking in Japanese even when I said "sorry I don't understand" repeatedly and then when I asked it to start in English and then gradually transition to Japanese it lasted all of one sentence in English before switching back. I can totally see how this would be useful conversation practice if you've progressed that far, but I'd love to have something for even earlier beginners. Also since many of the models you use are natively multi modal this could readily integrate visual media for discussion and grounding.

Also, for the transcription it would be great to get pure romanji to start with!

Which spaced repetition algorithm are you using? I recently learned that there is a much improved one that has been adopted by Anki. (https://domenic.me/fsrs/) Have you adopted that as well?

That's not a given, Anthropic recently added Claude CLI access to their $20/m "Pro" plan removing the need for a separate API key.

I love how fragmented Google's Gemini offerings are. I'm a Pro subscriber, but I now learn I should be a "Gemini Code Assist Standard or Enterprise" user to get additional usage. I didn't even know that existed! As a run of the mill Google user I get a generous usage tier but paying them specifically for "Gemini" doesn't get me anything when it comes to "Gemini CLI". Delightful!

Interesting! I'd love to know more about how you built this. Also as a small note the twitter / linkedin buttons at the bottom of the page don't seem to do anything but point to the page itself.

Can I suggest, don't do this? The sustainable unit of code modification is the ticket, not the commit. When you're ready to merge to main, squash all commits into one that takes the ticket title as its commit message and appends the ticket description as its description. This aligns code changes with planned, scoped, and documented units of work. Anything more granular than that quickly becomes noise. By following the above pattern your main commit history becomes a clean, consistent log of tickets being completed, each linking directly back to your ticket management system.

Failure or not I have to say thank you to this guy. This left Jeff Hawkins with a substantial personal fortune which he went on to use to found Numenta. I thoroughly enjoyed my time there and none of that would have been possible without that acquisition.

Several discussions here should be making references to literate programming but as far as I can see, haven't. For the curious and/or Knuth fans:

https://en.wikipedia.org/wiki/Literate_programming

I'm probably just ignorant but shouldn't the graphic of the tiled transpose have the green vector column-oriented in the final matrix?

All AI tools I've tried have failed to generate animations for slimes in my experiments. (Including Veo3) I'm hesitant to even try this, any chance it would work?

I spent several hours and a few dollars with Claude Code last night on the PAYG plan so this is a welcome change if it gives me more for my Claude subscription.

I don't see any official announcement of this on their Twitter feed so its likely an in-progress change.

The most interesting and significant bit of this article for me was that the author ran this search for vulnerabilities 100 times for each of the models. That's significantly more computation than I've historically been willing to expend on most of the problems that I try with large language models, but maybe I should let the models go brrrrr!

AI studio, Gemini, Labs ... There are far too many entry points for Google's AI work.

Do you know of any examples which use MCP and oauth cleanly?

"It hasn’t always been a cakewalk, but we’ve been able to establish a stable position in the ecosystem despite sharing it with all of these different kinds of intelligences."

Or, more accurately, we have become an unstoppable and ongoing ecological disaster, running roughshod over any and every other species, intelligent or not, that we encounter.

I wrote "Why Anki Doesn't Work for Me" (https://medium.com/@iandanforth/why-anki-doesnt-work-for-me-...) six years ago, which means it was before the new algorithm was implemented. While Anki likely still suffers from all the other issues I mentioned, this directly addresses my point (3). I'm going to have to give it another try and see if the other points are still too much friction, or if the frustration caused by the algorithm was a majority of my pain.

I applaud this effort, however the "Does it work?" section answers the wrong question. Anyone can write a trivial doc compressor and show a graph saying "The compressed version is smaller!"

For this to "work" you need to have a metric that shows that AIs perform as well, or nearly as well, as with the uncompressed documentation on a wide range of tasks.

This paper is concerning. While divorced from the standard ML literature there is a lot of work on biologically plausible spiking, timing dependant artificial neutral networks. The nomenclature here doesn't seem to acknowledge that body of work. Instead it appears as a step toward that bulk of research coming from the ML/LLM field without a clear appreciation of the ground well traveled there.*

In addition some of the terminology is likely to cause confusion. By calling a synaptic integration step "thinking" the authors are going to confuse a lot of people. Instead of the process of forming an idea, evaluating that idea, potentially modifying it and repeating (what a layman would call thinking) they are trying to ascribe "thinking" to single unit processes! That's a pretty radical departure from both ML and ANN literature. Pattern recognition/signal discrimination is well known at the level of synaptic integration and firing, but "thinking?" No, that wording is not helpful.

*I have not reviewed all the citations and am reacting to the plain language of the text as someone familiar with both lines of research.

They are, at best, doing both. But more honestly they are attacking scientific institutions because they are perceived as liberal.

Mind sharing which school?

On the off chance you're interested in school lunches I highly recommend watching videos of Japanese school lunches on YouTube. There's a bunch out there now and if you were raised in the American system it will probably blow your mind. The idea that lunches can be freshly made, on site, out of healthy ingredients and children are active participants in serving and cleaning up is just crazy. When I encountered it for the first time I felt like a big part of my childhood had been sold to the lowest bidder.

Anyone who puts "Ulysses" at the top of a best books list is suffering from expertitis. Ulysses has a massive user experience problem. It's hard. Dense, convoluted, absurd. If your friend asks you for a good book, you don't recommend it. The only time you do is when your college English major, or advanced highschooler, who is bored with the tropes of even very good novels wants to stretch themselves. Then you hand them this book.

I don't like this answer so here's mine.

"Read. Not too much fiction. Mostly books."