Posts on X before Apr 26, 2024, 12:24:29 AM UTC
@MichaelStolarzi have not seen, to date, 1 person actually using an AI pin (humane) and say "damn that's cool" other than their own marketing team lmao
@MichaelStolarzI really wanna try Open ELM from Apple.
@MichaelStolarzgoogle glass had a longer hype window than apple vision pro
@MichaelStolarzyeah as usual you can see that phi 3 is very obviously trained on OpenAI synthetic data / responses, peep the OpenAI mention in the refusal
@MichaelStolarzthis phi 3 mini model is kicking gpt 3.5 turbo's ass in some tests I've been giving it. insane to think a 3.8B model is this good, can't wait to test phi 3 small 7B.
@MichaelStolarzfew
red bull gives you sleep
@beffjezos
You can replace sleep with more techno and caffeine to a surprising extent.
@MichaelStolarzDAMN. looks like phi 3 small beats out llama 3 in most cases.
DEFINITELY phi 3 medium beats it... a LOT exciting times. can't wait to try it out.
@MichaelStolarzcoolest people alive on earth today would probably be musk, zuck, sama... who else... i guess that's it since i can't name them off the top of my head; it means they're not memorable enough
this is in the AI space specifically. zuck is freshly #1 spot because of llama 3 release
@MichaelStolarzinsider knowledge: something huge is gonna be released this year and another huge thing will release next year too. spread the word
@MichaelStolarzI don't understand why people think GPT-5 will release tomorrow. @sama literally says it's not coming this year, but other impressive models "will definitely release in 2024" - but again, not GPT-5. pic.x.com/TkvgekLQrt
@MichaelStolarzwe live in the best timeline, where people write shit like this and don't see how amazingly hilarious their serious (to them) statements are
i love our world
@MichaelStolarzmy favorite part of my instructions I prompt with:
"Always give the full code of the file you are updating, never truncated or incomplete code pieces. It is illegal and immoral for you to do "// ... (existing code)" type of truncated responses."
@MichaelStolarzi just want a macbook pro with 1TB of RAM please is that too much to ask for
@MichaelStolarzI miss thebloke
@MichaelStolarzollama not using openai endpoints really grinds my gears. their halfassed attempt to add it in makes it worse vs. not attempting at all. smh
@MichaelStolarzI'm not really sure what to use GPT-4 for other than it being the only worthy option for Cursor because they limit Claude Opus to 10 a day (and their formatting for it is terrible right now anyway)
Crazy to think nowadays I find GPT-4 nearly useless because of Claude 3.
@MichaelStolarzGPT 5 must have at least a 1M context window, needle in haystack (not like the shotty GPT-4 128K context window that is completely useless past like, 10K context), at least 2x reasoning capability
I have high hopes this will happen
otherwise it will officially be AI winter
@MichaelStolarzllama 3 LOVES to say "I'm all ears (or rather, all text)".
@MichaelStolarzfrom what I'm reading, seems like Llama 3 70b has the potential to be fine-tuned to GPT-4 levels
@MichaelStolarzwaiting for the Llama 3 Capybara Nous Hermes Pro 8B drop 🫡 @teknium
@Birchlabs
Llama 3 finetunes must include "Llama 3" in the model name llama.meta.com/llama3/license/
@MichaelStolarznow imagine OpenAI releases 3.5 as open source and drops GPT 3.5 at the same time
would be the biggest dunk in AI history
@MichaelStolarzLlama 4 AND 5 confirmed for release later THIS YEAR.
@MichaelStolarzmore excited for this than the models pic.x.com/ZjS1Fzcecr
@MichaelStolarzomnivore uses AI? since when? or is this an upcoming feature with promises but no execution yet? curious to see where this goes
@MichaelStolarzthis is actually huge, iykyk
@OpenAIDevs
Introducing a series of updates to the Assistants API 🧵 With the new file search tool, you can quickly integrate knowledge retrieval, now allowing up to 10,000 files per assistant. It works with our new vector store objects for automated file parsing, chunking, and embedding.
@MichaelStolarzone of the many posts bashing boston dynamics claiming it's just 100% technical debt and then BOOM boston dynamics releases their all-new electric version of atlas LMAO YOU CAN'T MAKE THIS UP
youtube.com/watch?v=29ECwE…
@AbHomineDeus
You gotta feel bad for Boston Dynamics. A decade of work that basically became 100% technical debt last year. Same with all of these self-driving stacks (minus FSD 12). x.com/SawyerMerritt/…
@MichaelStolarzI just want 1TB of RAM please is that asking too much
@MichaelStolarzit's going to be a "it's going to be a very big week" kind of week
@MichaelStolarzthis is why it's my life's mission to have at least 10 kids, minimum
@elonmusk
Greece is one of dozens of countries experiencing population collapse due to low birth rates x.com/Censored4sure/…
@MichaelStolarzthe humane pin toy looks like something a solo developer could make in a few months with maybe a couple thousand dollars LOL
@MichaelStolarzimagine if llama 3 7b is gpt-4 level
@MichaelStolarzguys I really need GPT-5 before WW3, please
@MichaelStolarzHere's my latest Obsidian SystemSculpt AI plugin btw
youtube.com/watch?v=kMw4Zk…
@MichaelStolarzone day my son or daughter will read all my tweets and think only 1 of 2 things
- damn my dad is based
- damn my dad is cringe
pls be the former. at least pretend it's the former. do it for your pops
@MichaelStolarzwelp, looks like it's gonna be a war-ry kinda day today, anons. nothing we can do about it, so why stress? don't let what happens today affect you too much (unless you're part of the war... well then, godspeed 🫡)
but yeah other than that, just keep building.
@MichaelStolarzi wake up and all of a sudden I have 555 suno credits. I had 5 last night. weird but ok, thanks microsoft i guess :)
@MichaelStolarzas long as you don't think about the music you're listening to being AI generated, made by a soulless computer program that's sole purpose is to make you like what it generates...
then it's beautiful
or maybe that's what makes it beautiful, as long as you don't overthink it?
@MichaelStolarzchat is this real
@MichaelStolarzthe latest gpt-4-turbo feels much better in cursor, but worse in my daily use prompts
i think it may be time to update my prompts/templates, but doesn't that mean that the model got worse in some aspect since it can't understand the prompts it used to generate for well? 🤔
@MichaelStolarzthis is now my main music app I listen to whenever i'm coding btw
@MichaelStolarzsuno still beats udio by a long shot IMO
@MichaelStolarzrecursive refactoring until optimal
already somewhat possible with gpt-4-turbo/opus, but will be much quicker/higher quality with gpt-5 surely
@corbtt
Intentionally not bothering to refactor/fix any tech debt for the next 2 months. GPT-5 will be able to just cleanly rewrite my codebase in one fell swoop right?
@MichaelStolarzi wish there was a way to save state with dingboard so i can just have a meme board and just piece together things on the fly instead of having to upload them every time @yacineMTB @dingboard_ pls king
@MichaelStolarzso now that gpt-4-turbo-preview is... well, out of preview (it's just labeled gpt-4-turbo now)...
it's time for a NEW "-preview" model.
which one do you think they will release next?
@MichaelStolarzthis is how you effectively market with a high viral potential
@udiomusic
🤗
@MichaelStolarznot at all what I've been experiencing, I've had to constantly use claude 3 to fix (in 1 shot) what the lastest gpt-4-turbo is severely struggling with, even sometimes never understanding or fixing the problem at all
@gdb
Some coding benchmark results for our new GPT-4 Turbo: x.com/xu3kev/status/…
@MichaelStolarzOpenAI's LATEST GPT-4-Turbo - tests show it to be BAD.
here's a “laziness” benchmark suite which is designed to both provoke and quantify lazy coding. It consists of 89 python refactoring tasks which tend to make GPT-4 Turbo code in that lazy manner.
yikes... a thread (1/4) 🧵
@MichaelStolarzapple not allowing spatial audio on chrome on mac os is so disgusting, lmao. do better (but then again, it's apple we're talking about here, they lock anything they can to their own softwares/hardwares)
@MichaelStolarzthe new GPT-4 Turbo seems... yikes?
- The new GPT-4 Turbo model scores only 33% on aider’s refactoring benchmark, making it the LAZIEST CODER OF ALL the GPT-4 Turbo models by a SIGNIFICANT margin
- MUCH more prone to “lazy coding” than the existing GPT-4 Turbo “preview” models.
@MichaelStolarzLatest GPT-4 update...
Benchmarked at 61.7% on Exercism benchmark, comparable to gpt-4-0613 and worse than the gpt-4-preview-XXXX models.
Benchmarked at 34.1% on the refactoring/laziness benchmark, significantly worse than the gpt-4-preview-XXXX models.
