d4rkp4ttern

No time to try other options but I’m betting on Rex from Superlogical -

https://www.superlogical.com/rex

On MacOS, dictation was solved a while back with Parakeet V3, and I haven't seen anything that has a better speed/accuracy tradeoff when talking to AIs, where slight inaccuracies are ok since they can "read between the lines".

I got annoyed by various UI and other things in tools like VoiceInk, Handy etc and Claude-Coded my own CLI dictation tool called VoxType that uses parakeet-mlx:

https://pchalasani.github.io/claude-code-tools/tools/voxtype...

Install this skill for your coding agent and let it help you set it up, there are many config options:

    npx skills add pchalasani/claude-code-tools -s voxtype-install -g

Or if you must do it yourself,

    uv tool install voxtype

This is just an example of where we are today with software today: people building their own tools, and not bothering to market it (or perhaps I'm contracting myself here ;).

The frontier live voice models (especially GPT Voice/Live or whatever they call it this week) do not have the typical AI-tells. So I wonder about a few things:

Why is this?

Why couldn’t whatever went into post-training these voice models be applied to the text-gen models to get rid of AI smells?

Could we set up a pipeline that leverages the voice models to clean up text that reeks of AI ?

Every single AW I’ve owned has this weird problem: when I go for a run it often doesn’t show my HR even after 10 mins, sometimes even longer. Super irritating.

Since one of the big improvements here is supposedly the writing style, on that topic I'm mystified about something:

Why is it that the voice models in Claude and ChatGPT have a perfectly normal style with barely any "AI smell", while the writing models are so obviously recognizable as AI?

The answer is likely that models underlying the voice modes are (post) trained differently. If so, then why can't the writing model be similarly trained? Presumably they haven't found a way to train them to be both "smart" (i.e. solve tasks etc) and pleasant to talk to?

Hmm first I thought it’s just $400 but with shipping 30 and duties 80 it’s more like $510

What I find equally interesting is finding out how well an AI model knows an open source software library you’ve built.

E.g I’m quite pleased to see most frontier-level models are very competent setting up agent systems using the Langroid framework, in development since Apr 2023:

https://github.com/langroid/langroid

An interesting fact I learned while reading The Dream Machine[1], is that Newell, Simon and Shaw’s Logic Theorist (1956) was considered to be the “first AI program”.

Quoting from Wikipedia: https://en.wikipedia.org/wiki/Logic_Theorist

“Logic Theorist is a computer program completed in 1956 by Allen Newell, Herbert A. Simon, and Cliff Shaw.[1] It was the first program deliberately engineered to perform automated reasoning, and has been described as "the first artificial intelligence program".[1][a] Logic Theorist proved 38 of the first 52 theorems in chapter two of Whitehead and Bertrand Russell's Principia Mathematica, and found a new and shorter proof for Theorem 2.85.[3] [1] https://press.stripe.com/the-dream-machine”

Amusing and amazing to see this in the context of today’s Erdos-slaying LLMs.

A lot of people, including me, don’t want to bother with GPUs, they’d rather run it on their M1-M5 MacBook. For example the 35B-A3B is very usable even on a M1 64GB MacBook.

Tangential to this post but Peter Bloem’s Transformers from Scratch (2019) was one of the best early tutorials on transformers and self-attention. There have been many more tutorials since then but this was the one that solidified a mental picture for me.

https://peterbloem.nl/blog/transformers

Agree with the spirit of this, but it assumes a human is still looking at the code, and human inspection of code is vanishing faster than people are willing to admit (especially when coding in a language said human doesn’t know).

Which electron apps?

I’m wondering if this would replace my trackpad placed between the two splits of my UHK

I have a few variations of pruning (trimming all long tool outputs, agent-directed trimming, etc) in my aichat collection of session tools:

https://pchalasani.github.io/claude-code-tools/tools/aichat/...

An interesting fact I learned while reading The Dream Machine[1], is that Principia was the basis of Newell, Simon and Shaw’s Logic Theorist (1956), considered to be the “first AI program”. Amusing and amazing to see this in the context of today’s Erdos-slaying LLMs.

Quoting from Wikipedia:

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

Logic Theorist is a computer program completed in 1956 by Allen Newell, Herbert A. Simon, and Cliff Shaw.[1] It was the first program deliberately engineered to perform automated reasoning, and has been described as "the first artificial intelligence program".[1][a] Logic Theorist proved 38 of the first 52 theorems in chapter two of Whitehead and Bertrand Russell's Principia Mathematica, and found a new and shorter proof for Theorem 2.85.[3]

[1] https://press.stripe.com/the-dream-machine

All these "intelligence" benchmarks miss something extremely important when using an LLM in a code-agent harness: How it communicates with you about what it did.

Opus-5 is practically unusable (for complex tasks) in this sense - its updates are voluminous, and dense with cryptic language (there are numerous reddit threads complaining about this, so it's not just me). I often have to ask it to re-state concisely in plain terms.

For a fairly gnarly task, after fighting with with Claude-Code + Opus-5, I ported my session to Codex + GPT-5.6-sol, and it was like a breath of fresh air.

Arguably a key aspect of intelligence is concise, clear communication, and current benchmarks miss that, at least as far as I'm aware. I would think some arena-type benchmarks where humans rate responses would measure this, though I'm not sure which those are.

I know about these. My question was about Tmux. Also the Claude resume search is not full text, it only matches session names.

I didn’t quite understand. I use Tmux extensively and have tons of sessions running different panes etc. But how does Tmux solve the session search problem, maybe I’m not aware of some Tmux feature?

Related: several months ago I made a TUI session search/manager tool as part of my suite of Claude/codex productivity tools:

https://pchalasani.github.io/claude-code-tools/tools/aichat/...

There’s Rust/Tantivy-based search and a whole host of utilities to view, resume, trim (to clean up context when close to limit), port to Claude/codex, move accounts, etc.

The search has a json mode for agents, and a corresponding skill so you can just ask Claude/codex to search for the “latest session where I worked on xyz”.

Probably referring to my post:

https://news.ycombinator.com/item?id=48938375 (100 up votes)

Direct link to the doc page:

https://pchalasani.github.io/claude-code-tools/plugins-detai...

This was in the context of how I use ChatGPT Live to brainstorm/discuss/learn while on a walk, but of course useful in other contexts too.

what guarantees are these? You mean data retention, use for training etc?

Interesting take. I should try that !

Regarding the urge to “check what’s out there”, one simple rule I follow is: spend around an hour in the morning catching up with HN, a few reddits, a couple newsletters, and Bluesky. I completely avoid Twitter/X and LinkedIn, trusting that anything significant enough bubbles up to those other places.

Most important — do NOT check any of those the rest of the day.

Love it. Like some others here, I've been making html presentations with code-agents and they come out really nice. Other alternatives for that suit different purposes are these:

- slidev markdown-based slides, I often have a code-agent make these. https://github.com/slidevjs/slidev

- typst slides are great looking too, especially with math etc. Again easy for agents to do

One minor nit: Based on the title, I thought this was literally a tool that had something to do with PowerPoint, but "PowerPoint" is used here as a synonym for "Presentation", which is ironic/unfortunate given how many dislike PowerPoint.

I use the Hex[1] app on MacOs for near instant transcription with Parakeet V3. This is how I speak to Code agents at least 80% of the time. The idea is - I tend to be lazy if I need to type lots of details, so speaking lets me get into details that I otherwise wouldn’t and this helps give more useful context to the agent. This often tends to be an unstructured brain dump so I sometimes ask the agent to repeat back what it understood, so I can make sure, and this also likely helps it stay on track.

[1] hex https://github.com/kitlangton/Hex

I used to use Handy but as of a few months ago it had stuttering and other issues so I switched to Hex. Even for relatively long minute long dictation, transcription is near instant and accuracy is more than sufficient, especially when talking to AI since it that can “read between the lines “.

Using voice with code agents is a huge unlock and I’m surprised to see some people I recommend it to, still resist it.

As for TTS I found it fun to make a voice plugin for Claude Code that uses PocketTTS, so it can give brief updates whenever it finishes a turn:

https://pchalasani.github.io/claude-code-tools/plugins-detai...

The other day I was trying to find out exactly how much context size 5.6 has in codex CLI (via subscription) and I didn’t find that documented in any of the obvious places. Sort of ridiculous that we have to find this out in X or Reddit, e.g:

https://www.reddit.com/r/codex/s/hVv29obfFD

In the Reddit post there was clarification that this was done with Sol Pro not Ultra - curious what is everyone’s mental model of the difference.

My understanding is that ChatGPT Pro is effectively a multi agent system, or somehow uses multiple LLMs in parallel and selects a best answer. And Ultra is more similar to Claude-Code UltraCode where the main agent can choose to create a dynamic JS workflow that deterministically orchestrates multiple agents to handle different parts of a task and have adversarial checkers etc.

Is that more or less the difference? Any substantiating sources would be great to see.

Nice, it works for discussing concepts baked into the LLM but fails when you want it to read contents of an article or do web search. ChatGPT Live doesn’t have this limitation. Their Voice mode did have this limitation, but the Live mode released a couple weeks ago works exactly as you’d want, with link-following and web search

Too much friction in notebookLM - several minutes to make a podcast, and they disappear after a while, and the interrupt feature is super-janky.

Related book by Blum, Hopcroft, Kannan:

Foundations of Data Science (2020):

https://home.ttic.edu/~avrim/book.pdf