bravura

So what's at loggerheads here is:

* The LLM has strong and deep rooted belief in its knowledge (that a seahorse emoji exist).

* It attempts to express that concept using language (including emojis) but the language is so poor and inaccurate at expressing the concept that as it speaks it keeps attempting to repair.

* It is trained to speak until it has achieved some threshold at correctly expressing itself so it just keeps babbling until the max token threshold triggers.

Couldn't see it, what was it?

As an American living in Europe, I have seen Europe do "no cameras by default" quite successfully.

Right but you have a good security posture and hygiene. MCP as a use case (not a protocol) is encouraging risky usage by less security minded people.

Question: Is the computational cost of verifying the proof significantly less than the computational cost of creating the proof?

Can you share an actual example demonstrating this potential pathology?

Like many things in ML, this might be a problem in theory but empirically it isn’t important, or is very low on the stack rank of issues with our models.

An old bit in Harper's Index quoted that the average number of time you have to taste something you have an aversion to and develop a taste for it is 12. Like most statistics in Harper's Index, the veracity of that number is questionable and amusing. However, the idea that you might just need to try more than you think is what's thought provoking about citing this probably fallacious statistic. (Harper's Index was basically a nicely arranged list of thought-provoking cited statistics you weren't supposed to take at face value, arranged in a way to create a narrative arc.)

Another piece I read on this topic was aimed at foodies. The thrust was basically: "You're trying to broaden your horizons by trying different obscure dishes from other cuisines. Why not revisit the foods you hated as a child? I used to hate eggplant and thought it was slimy and now I love eggplant."

There's something compelling, in matters of taste, about addressing a huge blind spot you have been avoiding tackling instead of hyper-specializing in your chosen niche.

There's a difference between being present and enjoying something in the moment versus equating the background process with the moment itself.

Have you reversed engineered the Ableton project format?

I got frustrated the last time I tried, and currently LLMs tell me "Live updates break things constantly". Are my fears oversold? At the time I was doing it, I was trying to control all the parameters of Ableton native synths which was a headache.

Now what I want to do is mainly create different kits of loops + one hits using ML, at the bare minimum.

I have a background in audio ML so I imagine there would be other sorts of thing I might want to do later. But currently it's just one "template" with different filler sounds.

Sure, I can describe some of the basics I'm trying to do.

My goal is to autogenerate "projects" with loops and one hits, so I have like 100 candidate "songs" or "kits" to work with just using hardware and avoiding a DAW. And then to quickly jam on each song/kit to create an arrangement in hardware that I can save. Ending up with 20 scratch arrangements that I can load into a DAW and then produce / etc more carefully.

I want to load a "kit" into the MIDI controller and trigger sounds by hitting buttons. That part seems clear to me with JACK. That I can basically drop straight into hardware and start working with my audio material.

Now where I feel stuck is I'd like my knobs to do basic filtering, I'd like the loops to loop when I trigger them, and I'd like to be able to save my jammed arrangement to load into a DAW later.

Saving it seems like I could just save MIDI.

For filters, I am not sure which library I would use, and here is where it seems like I should be controlling a real DAW (so I get the same filter during my jam as in postproduction).

For triggering a loop that goes and loops, I have to write that by hand?

Apologies if what it sounds like what I want is: "I can use software to create 100 candidate Ableton projects with the clips pre-mapped and do a looped arrangement very quickly."

This is very illuminating, thank you. I feel like a veil has been lifted from me.

What sort of scripting are you doing?

Do you mind sending me an email (in profile) if I have a few questions?

I'm aware that the format is undocumented XML, thank you for sharing anyway.

I've found it painful enough to reverse engineer that it sucks the joy out of any Ableton programming.

I honestly believe this is one of the most end-user unfriendly decisions in the digital music world.

> If you're already using Ableton, this should be easy as long as your computer/audio interface has a midi out port, you basically just select what midi channel a Ableton midi track should use, and start sending stuff, no coding or fiddling required :)

That requires clicking around in Ableton though, right? My goal is to generate like 100 different projects and jam on the effortlessly.

How easy is it to integrate reaper with hardware?

I find it really frustrating that the Ableton format is closed, so you can't write software to generate projects for Ableton Push.

Want I want is the fastest path to making beats using physical interaction with MIDI hardware based upon loops and one-shots I have, without endless coding or fiddling. Open to purchasing new hardware.

I was very excited for a strong off-the-shelf code vector embedding search tool.

I wanted to encourage you to explore that direction, since it's a) very powerful, b) annoying to hand-roll, and thus c) sorely needed as open source.

Does ccusage (or claude code with subscription) actually tell you what the limits are or how close you are too them?

1) Internet commentary is generally pretty low quality, but perfume nerds seem categorically to all be the most interesting person you would ever invite to a party: https://basenotes.com/fragrances/no-5-by-chanel.10210628

every single comment on that website is amazing.

2) I hate perfume. I met an avant garde perfumist called Christopher Brosius (label: "I hate perfume") and waited 20 years to buy his samples. They are AMAZING. So approachable. Everyone who has let me dab with his fragrances has been blown away. "In the library" smells like old books. "Wild hunt" has rotten leaves as an ingredient. "Walking on air" smells like fresh cut grass. I hate perfume but I am obsessed with his smells: https://www.cbihateperfume.com/

Pairwise rank constraints involve fewer assumptions that per-item scoring about the underlying nature of the data, thus they are more robust.

Claude Code with an API token, even when told: "Iterate until github is green without stopping" will sometimes just stop and wait to be told to continue.

Thanks for building this! The tool shows a lot of promise. Coming from Claude Code, the core functionality feels solid - just needs some permission refinements to match enterprise use cases. This is based upon quickly skimming the current code.

High ROI feature requests:

• Pattern-based permissions - Bash(git:) to allow git but not rm, Write(logs/.txt) for path scoping

• CLI permission flags - --allowedTools "Read,Bash(npm test)" --deniedTools "Write" for session overrides

• Allow/deny precedence rules - explicit deny should override general allow (security principle)

• Config file hierarchy - system → user → project precedence for policy enforcement

Medium ROI improvements:

• Command argument filtering - whitelist git commit but not git --exec-path=/bin/sh

• Multiple config formats - support both simple arrays and structured permission objects

• Runtime permission diagnostics - gemini permissions list to debug what's actually enabled

• Environment variable injection - top-level env config for OTEL endpoints, API keys, etc.

The permission engine is really the key piece - once you can express "allow X but not Y within X", it unlocks most advanced use cases. Keep up the great work!

And how do the results compare to hosted LLMs like Claude 3.7?

Sorry, I side with GP. Just because you don't want to use Llama/GPT because of cost, the middle-ground of DistilBERT etc (which can run on a single CPU) is a much more sensible cost/benefit tradeoff than VADER's decade old lexicon-based approach.

I can't really think of many NLP things that are one-decade old and don't have a better / faster / cheaper alternative.

I want LLM accessible bookmarks. That's it.

It doesn't work yet.

I use singlefile to archive pages I'm viewing Linkding.

Then I have a BeautifulScript4 script to strip the assets.

Then I use Jina's ReaderLM v2 to render the HTML to proper Markdown: https://huggingface.co/jinaai/ReaderLM-v2

Except, of course, for longer table oriented text documents like HN that doesn't work.

I want a plaintext archive of web pages in a github repo or similar. Not a fancy UI/UX

I usually like your takes, but where I disagree today is when you say: "Truffles have a functional role in a dish." but fonts don't.

Either both do have functional roles or both are luxuries like Hermès.

Sorry, does that mean I can use anthropic and gemini with codex? And switch during the session?

It's a peculiar and fascinating observation you make.

With static images, we always look for eyes.

With video, we always look for dancing.

Question: Does this use single file or something similar? I've basically been dying for an easy to use local bookmarking service that uses singlefile?

Second question: How easy is it to crawl the local store and apply LLMs or RAG?

This is an increasingly important area, with LLM generated code, and am curious about people's experiences with codesandbox vs e2b vs daytona

Recently, I signed up for two premium cards. The main benefit I wanted: Collision and damage coverage on INTERNATIONAL car rentals.

Honest feedback: I logged into your site and found it useless. I plugged in my two cards. Your site gave me such a cursory and incomplete picture of them, I would never have the information I needed.

p.s. they are literally the credit cards you have on your homepage, so I expected at least these to be complete.

Quantity has a quality all of its own.

When doing interior decoration, I am definitely interested in finding objects that fit very specific prompts.