Lev Popov

It's *not* a well-made keyboard, but I'm still going to be a bit sad when they all run out now that they don't make them anymore.

This is also a periodical reminder that getting a larger monitor and putting it further away from you (deeper desk, setting up a keyboard tray, etc) is the best thing you can do to reduce eye strain.

Many reasons to doubt VR, but complaining about a screen sitting too close to your eyes is not a good one. The typical focal length through the optics is much greater than what you keep your phone/laptop screens at. Your eyes are more relaxed looking through VR lenses!

@deepfates Rate limits, probably. It's 3.5k requests/minute for embeddings and it will error out if you exceed the cap. Try throttling req rate.

Why do I keep getting status code 500 from the openai embeddings endpoint after a couple thousand calls?

@tr_babb I think "tab" is that button ;)

(but seriously a "pet CoPilot" button could be used to actually reward the model and help train it for good results)

It's cute how I'll be describing some problem in a comment, and CoPilot will autocomplete like "🥺👉👈 perhaps we could make an X that does Y, and that would solve it" I want to pet it on the head for being helpful

If your agent recursively spawns itself, it’s an LLM MLM

@tr_babb Really want to build an extension to automatically click "not interested" on every "recommended" YouTube video that you haven't touched in >2 weeks. Might be a good GPT-4 task.

You watch one train video, and then YouTube thinks you are a Trains Guy forever

@zachtratar user-agent ;)

For a GPT-style AI that does work for you, what name makes the most sense?

@mayfer Nice! Would be awesome if it (optionally) auto-selected GPT-4 for new threads. Injecting below should do the trick: var click = (tag, str) => [ ...document.querySelectorAll(tag)].filter(el => el.innerText.includes(str))[0].click() click('button', 'Model') click('li', 'GPT-4')

made a MacOS menu bar app for quick access to ChatGPT & i've been finding it super convenient can customize a keyboard shortcut too $2 one time payment (or alternatively open source if you want to build yourself) aiadventure.spiel.com/menubarchat

@DavidHodge Absolutely. Especially interesting if they will try to do training optimized hardware or if they’ll let Nvidia own this for foreseeable future

@DavidHodge At inference, it already is (and pretty far ahead of other mobile chips short of dedicated GPUs). Whisper.cpp is now doing encoding forward pass on ANE

@ericmigi Fair :) BeeperBand looks really slick! Mass-market watches mostly fail at "wearable remote" use case (apple *finally* added a programmable button last year, but still super limited), so there's a big gap to fill. Accelerometer gestures for single-hand use would be amazing.

@levpopov I mean, Pebble technically fits the bill but I want something even simpler

This is a little side project I kicked off a few months ago with @tty0x24. Premise is: 🎧 I use bone-conduction headphones (usually @sentienhq) 🗣️ I want to be able frictionlessly send a voice message to my wife over Beeper 👉 Don't want to reach up to press button on headphone

@ericmigi I think PineTime might fit the bill for "wearable BLE Button" - $27, slightly lower spec Nordic chip, buttons, haptics, accelerometer, 1 week battery.

Honestly not sure what else I will use it for, but I was amazed that there was no existing 'BLE button' device on Aliexpress!

@mpdaugherty And definitely better on Chinese->English as well (tested on a bunch of text from component data sheets). Usually English datasheets are just (poorly) ML-translated copies anyways, so being able to just reference the original is super useful.

GPT-4 is better than Google Translate - last night @levpopov and I ran a blind test comparing outputs from Google & GPT-4 for English => Chinese for discussing technical requirements with suppliers. GPT-4 was more accurate & sounded more natural.

Adventures in electronics component sourcing: My email thread with a supplier is now so long and has so many people on it that gmail web UI can no longer handle it.

@reidhoffman @OpenAI p.151 Investing in AI: "Adept uses AI to provide a more intuitive and accessible interface for CAD/CAM (computer-aided design and manufacturing)" Huh? Is Adept focusing on CAD/CAM? Or is this a different Adept?

I wrote a new book with @OpenAI’s latest, most powerful large language model. It’s called Impromptu: Amplifying our Humanity through AI. This, as far as I know, is the first book written with GPT-4. Here’s how it all began… impromptubook.com/wp-content/upl…

@goodside ChatGPT api allows multiple "system" messages (unclear if officially supported or just an oversight that happens to work), so you don't have to pretend to be the user. Just inject another short system message at the end with a reminder for the desired assistant behavior.

This is clever. We’re at a point in LLM alignment where, to avoid ChatGPT/GPT-4 breaking character when a user pretends to be the developer, the developer must pretend to be the user.

A naive fix could be a hybrid model. A large, slow supervisor that generates core tokens, and a fast "filler" model that speeds ahead when it has high confidence. Could probably even have the supervisor figure out if the filler model screwed up and output a "backtrack" token.

Every token taking the same amount of time is almost certainly an architecture bug. We are throwing away way too much useful computation between each inference step.

GPT4 is impressive but painfully slow (we got spoiled with 3.5-turbo). The most jarring part is GPT taking a while to compute "obvious" tokens - 1) you often already know where it's going 2) *it already knows* where it's going. Yet it has to start from scratch on each token.