hoc

For an adhd mind that serial interface is a bless and hell at the same time. Great for the detailed planning discussion, bad for checking the results.

I am currently working on canvas-based interfaces for that reason and i would think the only way to really create value here is with a deep independent analysis and visualization of the changes afterwards to reach some ease of mind. Live would be cool (if you like that)

When it comes to planning itself, I recently tried the token-saving planning plus phases execution agents approach and had to find out that agents actually don't necessarily communicate better by prose-reduced specs than we do.

I had to go back to the planning agent to implement or fix things with our full planning context in mind. So if you want really high control for a "tight" implementation, I'd say just sharing plans is not enough. The probability of things getting filled in by the executing agents rises and you either find yourself holding those agents' hands or fixing things afterwards.

Actually phases are still to large and you would actually want the planning agent to hold that hand all the time, meaning small context is not the way to go, as you might need the full checking context much more often than current phases sizes suggested by the planning "doc" would use it.

Distributed building still is the way to go though, steady control by the overall context or one specifically thinned out for the particular job is. But don't go prose-based plans anymore. These are dead indeed.

"I didn't know one could get lead out of a zoom meeting"

I'll follow it.

For me, in the early 80s, this was all about state. Reliably keeping a state, well defined, numbers or strings, bytes preferably, and being able to act upon them.

Only later I learned that this, ehat I was missing in the analog world back then, was what made up the core of a turing machine.

So this beauty is still the same for me. Just the sheer amount of state that is available and provided by others make the concept much less powerful than back in the day.

Now AI brings that back a bit, by finding the right items that you can keep and iterate on. But we tend to let AI also do rhe iteration and that introduces that non-deterministic character that the computer had overcome.

So, no wonder, that I, and quite a few others, at the moment, still mostly use AI for finding and typing code to describe the structures for the machine, but keep trying to define the iterations ourselves, guaranteeing clear insight and access to the state we are trying to work with.

Everything else is more like working with an assistant back then (or today), extending your actionable potential instean of your mind. And depending on how you see the world or what the tasl at hand might be, you might prefer one over the other, control over action, insight and perspective over tinkering with the matter to push it somehow in the right direction or implementing a known process.

But that state thing, still priceless, timeless. The right augmentation to our fuzzy brains and better than paper, since, who thought, we can express the algorithms in the same way.

So, I guess I will always see this beauty, even in a simple flip-flop, coin, switch and any array thereof, and anything that can be controlled by that binary configuration.

Nice. I loved his Short Circuit tune back then and looked at its code with my cartridge's monitor to extract it into a standlone player shell of mine (which might actually have worked out). Great to see the sources for the addresses and their meaning that one had to make guesses about 40 years ago...

Also in the linked player under Short_circuit.sid , btw. Thanks!

Great project and write-up. I wonder whether most of those "hints" are really needed, though, as you are already using Claude CODE. Aren't things like "simple" and "clean" assumed to be part of its system prompt already (idnividual documentation style etc can't be, of course). While they were useful when using a general LLM for coding, I would think that they are now part of the overall setup of any coding agent. These days I run more into problems with language and api version drifts, even when specified beforehand.

Any numbers on practical pricing per country for these scans?

I'd rather identify with the "These docs are old and won’t be updated." part of your linked page.

Well, it seems it uses the word "research" quite a lot for what it's doing (first paragraph and last one with "deep research") but does choose some artificially constructed one as the result.

Personally I had similar results when searching for known terms for certain concepts that I didn't know the name for. And usually I had to guide the process to find the actual expression used in the domain (it usually made up a lot of fancy and well-fitted names itself). And sometimes it helped to gonback and change the query. Harder if you have to wait that long, though :)

So.. I guess I would trust its process (mentioning "research" a lot) more than its chosen result :) Not sure if it you wanted it to be a synthesizer or rather an assistant. I'd go with what is closer to your intention or what you think it resulted in in the end.

An interesting observation might be that guiding that (I don't dare to say it, but anyway) research process might still be an important part and the network's self-evaluation might not be as good as one would need it to be at this point in time. I'd guess it's about adding ones personal judgement early to not end up at the wrong spot after a long processing time. Then again you are using a certain selection of simple and more complex models, so I can't say if there might be a way to have that kind of harsher judgement that one would apply oneself emulated in the proces and what side effects might come form that choice of models you made.

In the end I was just surprised by the number of picky replies your post got, so I just thought that you discussing the perfect description with the LLM might make be a fun solution. Personally I am a big fan of interactively talking to LLMs at the moment though, so I might be the wrong guy to know your intended audience and use case enough. Just couldn't see how using the term "research" would be the problem.

By having to defend your (thesis)/work like this, the whole piece is getting lifted into academic heights in a way, so you could as well keep calling its result and process research :)

What description would itself come up with, BTW?

When you anwer with "I agree, LLMs have biases.", I immediately suspect that to be an LLM calming me after correcting it, though. So, the world has definitely changed and we might need to allow for correcting the broadness of words and meanings.

After all you did not write thesis, scientific research or similar and I remember it being called researching when people went looking up sources (which took them longer than an agent or LLM these days). Compressing that into a report might make it a review, but anyway. Great that you assembled a useful work tool here for some who need exactly that.

Pretty neat, actually. Thanks for looking uo that link.

That also is similar in a sense to a typical human bahavior of "rounding" a "logical" argument, and then building the next ones on top of that, rounding at each or at least many steps in succession and bacically ending up at arbitrary (or intended) conclusions.

This is hard to correct with a global training, as you would need to correct each step, even the most basic ones, instead. As it's hard to convince someone that their result is not correct, when you actually would have to show the errors in the steps that led there.

For LLMs it feels even more tricky when thinking about complex paths being encoded somehow dynamically in simple steps than if there was some clearer/deeper path that could be activated and corrected. Correcting one complex "truth" seems much more straightforward (sic) than effectively targeting those basic assumptions enough so that they won't build up to something strange again.

I wonder what effective ways exist to correct these reasoning models. Like activating the full context and then retraining the faulty steps, or even "overcorrecting" the most basic ones?

That missing Playdate phone accessory.

It somehow feels that one has to come up with a tailored strategy for each CVS import that contains decimal numbers, dates and/or other data with dots in them. Most of the time I end up regexing my way out of the failing (and different) attempts the different build-in options in Excel deliver. LibreOffice is better, Apple's Numbers partially worse in that regard. Funny to see these things in that state in 2025.

Discussing number data with current LLMs also does not seem right at the moment. Not sure if simple column-specific conversion tasks might work.

The promptly emerging Excel-Schmerz reminds me that the author avoided the use of floating point numbers in the provided example for a reason.

In german Excel it's might be ok due to its original focus on numbers, but replacing decimal "." with the german "," also forces you to use a semicolon everywhere where you would use a comma (function, vector/array notations). To me the most annoying issue with localized programming languages. And consecutive ";;;" just looks awful. And what would happen to command/line endings...

BTW, Apple once tried to translate content-specific parts of their AppleScript language (like Dialogs etc) but in the process also hit some enumerations which often sit at the border between programming language and content. Big desaster in certain edge cases.

Anyway. Of course I like that the effort was made. There are famous Asterix comic books translated into all kinds of languages and dialects. I'd really like to see great localized coding languages, but I guess they will have to avoid line ending semicolons and only use integers :)

Gov sites issueing warnings about citizen data security seems a bit funny these days.

Elon: "Comedy!"