This is devastating and should be 1) free to read and 2) broadly published

Even once he gets confirmed, he's just one BoG/FOMC vote. He doesn't set rates himself! It's all highly contingent! Plus there is ~zero correlation between what any Trump Admin nominee has said to Congress and what they do on the job. All a massive waste of time to focus on.

But there's dirty filthy nature outside. Ewww!

this is just a warmed over david duke speech from 1990

You are in CO. Just walk into the mountains.

I will laugh though if it turns out that Mamdani is sitting on some massively appreciated, no dividend paying compounder stock (like Berkshire) and is a #neversell bro.

$NPK Annual Report design never ceases to crack me up.

A white garment with some kind of food stain. Humiliating.

A pair of dirty hands rest on a car engine. Rough, tough, cool.

Here is an easy solution:

Food stains are humiliating, as they suggest you're a little baby who can't feed themselves.

Oil stains, such as those you'd get from working on your car, suggest you're tough, independent, and skilled.

Thus, simply cover your food stains with used motor oil.

Very much not trying to call out this specific person so PLEASE do not pile on, but this is a good example of how this very common false belief that people can get status if they try hard enough often leads to a kind of presumption that if someone HASN'T gotten status, they are somehow at fault.

Birthright Americans really have no appreciation for how long and filled with pitfalls the process of obtaining U.S. citizenship is *even* if you have followed all the rules and processes to the letter.

Lady missed one deadline after fleeing a pogrom and essentially the doors all slammed shut.

Real engineering judgment:
If a regex solves it, use a regex.
If a SQL query solves it, use a SQL query.
If a deterministic function solves it, write the function.
Reaching for an LLM when you don't need one is how you get slow, expensive, unreliable software. Use the right tool.

Lol.

Took me a minute to figure out what this was (it's nothing mysterious - long story), but I was creeped out for a moment when I saw it in the TUI.

Anyway, I think the main event with $CODI has now occurred. Mostly out of the common. I.e., show's over, you are on your own at this point - not that any of you were crazy/stupid enough to listen to me in the first place (I hope).

...the other extreme of this: sometimes I need to put a whole lot of crap into context and the harness shouldn't increase the friction of doing that. Trust that the user is aware of all the dangers of saturating context (otherwise you end up with Anthropic bulls**t where they truncate silently).

Added later: philosophically I would say that a big part of why I am a big fan of pi is because it adheres to giving me control of what goes into context in a no-fuss manner. I think Mario came at this from the angle of "it wont put things in context behind your back". I am essentially asking for...

"Now I have the full picture" my a**! I just saw that you only read the truncated file!

Caveat on the Sonnet vs Opus distinction: based a bit on vibes and anecdotal experience. I haven't tried running any rigorous tests of this yet.

Opus 4.6. In fact, I find that Sonnet is often "better" about things like this: Opus tries to be "too smart" in situations where it's absolutely the one thing that will get you killed. Sonnet OTOH ends up being just "dumb enough" to do the drudge work of reading every single character w/ offsets.

Anyway, just remember that the Gell-Mann amnesia principle/idea also applies to LLMs.

...solves the problem. However, my hope was/is that simply having a customizable setting would be a simple and more elegant solution.

For the most part, it makes me very happy that pi surfaces truncation *loudly*. Claude Code and the like have situations where they silently truncate (abominable).

...minified files that being able to tweak the limits would help with.

And yes, I admit this is a quality of life request rather than some big issue. As you point out, just overriding the core tools with custom ones (or preprocessing files)...

...chunked to be within the truncation limits and then just force feed them to the model. Allowing truncation limits to be easily modified would smooth this (whole document force fed to the model; ofc, the context window has to be large enough to allow this). There are also some issues with...

...to ground truth provided by source documents. So I have a few times caught the model hallucinating because it read 995 lines (say) of 1000 and inferred the rest (even though it was prohibited from doing so). It's partly a model issue (vs harness). Current workaround is to have the source docs..

Yes, I know I could replace the commands (but seems like too much ado for just modifying truncation limits). To be clear, the idea is NOT to get rid of truncation, but to be able to modify the limits easily.

The reason/use case isn't coding. It's knowledge based work that needs faithfulness...

Mechanical sympathy Use the tool as meant to be used.

New post: On mechanical sympathy vickiboykis.com/2026/04/13/m...

...on truncated content. Manual chunking is a workflow solution to this, but often I am left wishing that I could just tweak the limits depending on the project.

To be clear though, mostly a "nice to have", not a real "irritation". The "loud" surfacing of truncation is very much appreciated.

Where this plays a role is for (my) use cases where fully ingesting certain files is non-negotiable. Although pi surfaces truncation to the model very clearly (and suggests reading with offsets to fully ingest), models are often lazy and try to even ignore specific prohibitions on inferring based...

Not that big a deal, but wish it were different: as I understand it (possibly mistaken?), truncation limits in core read/bash tools are hardcoded. So need to change truncate.js directly to modify (fragile across package updates). Settings.json having an entry that overrides would be nice.

Ain’t big evil insurance fat cats trying to fleece you, it’s your compatriots.

www.newyorker.com/magazine/202...

To this point, here are the metrics from a large buy here pay here financing op (think very low credit or completely broken credit, high mileage cars). The way it works is that they are essentially “forced” to increase terms to keep monthly payments affordable.