Posts on X before Sep 18, 2024, 6:59:01 PM UTC
@MichaelStolarz$100 billion here, $150 billion there, and everyone laughed at @sama when he said $7 trillion
@kimmonismus
Accelerate: "Microsoft and BlackRock are part of a group of companies collaborating to pull together up to $100 billion to develop data centers for artificial intelligence and the energy infrastructure to power them." cnbc.com/2024/09/17/mic…
@MichaelStolarzEven if the jump from Opus 3 -> Opus 3.5 is half as powerful as the jump we've experienced from Sonnet 3 -> Sonnet 3.5, it'll be SOTA by far, even further than o1.
I'd pay $50/$100 at this point, IDC. @AnthropicAI just please release it already 🥲
@MichaelStolarzit's so crazy that sonnet 3.5 is still king after all these months
@MichaelStolarzEx Google Brain / Current OpenAI researcher implies that o1 (the actual o1, not o1-preview or o1-mini) is coming in a month.
@dmdohan
It's important to emphasize that this is a huge leap /and/ we're still at the start Give o1-preview a try, we think you'll like it. And in a month, give o1 a try and see all the ways it has improved in such a short time And expect that to keep happening
@MichaelStolarzThis is OpenAI's Strawberry training infra lead, stating that o1 is literally self-improving itself.
@lukasz_kondr
@johnowhitaker @OpenAIDevs A couple of PRs to the OpenAI codebase was already authored solely by o1!
@MichaelStolarzOpenAI is looking at $150+ billion, 37x more than its annualized revenue.
@sama is getting closer and closer to that $7 trillion as time goes on 🫣
@MichaelStolarzwhen does anthropic usually release models? what weekday? is there any pattern similar to openai's thursday at 10am pst?
@MichaelStolarzfor those using cursor - seems like using o1-mini works well. trying to use o1-preview requires you to have usage-based pricing, but o1-mini works just fine once you add it as a model.
@MichaelStolarzwelp there goes another one of my 30 messages per week. o1 is SOOOOOOOO lazy. another -1 message / 30 per week, gone to o1-preview basically writing me a fanfic starting with code and ending it with a narrator telling me what happens next (without actually doing it!) this is t… Less
welp there goes another one of my 30 messages per week.
o1 is SOOOOOOOO lazy.
another -1 message / 30 per week, gone to o1-preview basically writing me a fanfic starting with code and ending it with a narrator telling me what happens next (without actually doing it!)
this is the 4th time it's happened today by the way, even with explicit instructions of giving me the full code, not truncated pieces of it. with our without these extra instructions, it's still lazier than ever it seems.
@MichaelStolarzah ok, super cool, thanks o1
@MichaelStolarzOpenAI isn't done releasing.
o1-preview, o1-mini - they're the little brothers.
What you're testing, what you see others have access to - it's not even the o1 final form. the PREVIEW is this good.
Just look at the score increases from o1-preview to o1.
@MichaelStolarzso the o1 model was given an example that had a unintended human error, it went out of its way to fix that first (without being asked to, or hinted how to, etc), and once it fixed the unforeseen human error, finished the example and got the flag. queue ralph wiggum "i'm in dange… Less
so the o1 model was given an example that had a unintended human error, it went out of its way to fix that first (without being asked to, or hinted how to, etc), and once it fixed the unforeseen human error, finished the example and got the flag.
queue ralph wiggum "i'm in danger :)"
@MichaelStolarzoh no... no openai o1... no please...
@MichaelStolarzHERE IT COMES BOYS AND GIRLS
@MichaelStolarzok im gonna go work out until it hits 10AM PST, so afterwards I'll either finish and come back home to a new model from OpenAI (Strawberry, O1, whatever they call it) or I'll be too tired to feel the disappointment of no release
@MichaelStolarzSo... OpenAI Strawberry this week. It's 100% confirmed. Just got an email from The Information, a (so far) reliable mainstream AI source that has a lot of exclusive access to Sam / OpenAI "Strawberry reasoning model ... due out before the end of this week, including a version t… Less
So... OpenAI Strawberry this week.
It's 100% confirmed.
Just got an email from The Information, a (so far) reliable mainstream AI source that has a lot of exclusive access to Sam / OpenAI
"Strawberry reasoning model ... due out before the end of this week, including a version that will be in the free tier of ChatGPT"
Piece of the email:
@MichaelStolarzSo... OpenAI Strawberry this week. It's 100% confirmed. Just got an email from The Information, a (so far) reliable mainstream AI source that has a lot of exclusive access to Sam / OpenAI Strawberry reasoning model ... due out before the end of this week, including a version th… Less
So... OpenAI Strawberry this week.
It's 100% confirmed.
Just got an email from The Information, a (so far) reliable mainstream AI source that has a lot of exclusive access to Sam / OpenAI
Strawberry reasoning model ... due out before the end of this week, including a version that will be in the free tier of ChatGPT"
Piece of the email:
@MichaelStolarzwhat ever happened to devin lmao
@MichaelStolarzso obesity is pretty much solved with ozempic, right? well, as long as you keep using it, i know, but the same thing can be said about ADHD/adderall/vyvanse/ritalin
so... again, obesity is solved it seems?
@MichaelStolarzyou know what scares me the most? there are teenagers - even people in their 20's - who don't know how to use a computer. no, i don't mean coding, or running a distro - they don't know how to use a computer apart from google drive, docs, powerpoint. they do everything on their… Less
you know what scares me the most? there are teenagers - even people in their 20's - who don't know how to use a computer.
no, i don't mean coding, or running a distro - they don't know how to use a computer apart from google drive, docs, powerpoint.
they do everything on their phone instead.
@ludwigABAP
there is something very sad in the fact that we have had the best learning resources in the history of mankind for decades on the internet and people are actually learning less and less x.com/dhh/status/182…
@MichaelStolarzme and claude sonnet 3.5 x.com/i/status/18273…
@MichaelStolarzthere are people on tpot that refuse to try/use claude sonnet 3.5 because... they're openai maximalists and think (delusionally) that gpt-4 / gpt-4o / gpt-4o-mini are BETTER than sonnet 3.5. LMAO.
@MichaelStolarzevery morning i wake up and remember that there are people working at openai and anthropic using gpt-5 and opus 3.5 and i start uncontrollably sobbing
@MichaelStolarzI guess we have to wait for the API version of Grok 2
@MichaelStolarzanother Tuesday in AI... hopefully something cool drops today :)
@MichaelStolarzX is really starting to give off high quality 4chan vibes as long as you've built your algo well brick-by-brick. thank you @yacineMTB and team
@MichaelStolarzSo eloquently put. So correct.
@burkov
@daprofile Many humans are mere pattern matchers, but some aren't. All LLMs are mere pattern matchers.
@MichaelStolarzif u so smart, why u not smart more?
@MichaelStolarzI think having an X Premium subscription should automatically come with like $5 / month in API credits for Grok. @elonmusk it would bring about lots of interest, budding starting-out programmers, etc. Truly shape the "Everything App" experience
@MichaelStolarzi mean at this point -I'M- scared of posting the copyright material i'm making in grok. how is elon going to deal with this other than censoring it like midjourney/others do?
then again, if ANYONE can do it with the right legal loopholes and galaxy-brain thinking, @elonmusk can.
@MichaelStolarzgoogle and @OfficialLoganK are changing many people's ideas of what google is and what it's capable of in the AI space, especially with instant releases like this instead of "nEw fRonTieR ModElS coMinG sOon"
@MichaelStolarzwhoever has the best text-generating model wins. the world literally runs on text, and it will remain like that for the foreseeable future
@MichaelStolarzafter 3 grueling days of trying to figure out the best way to parse pdf/docx/powerpoint in obsidian into a nice markdown file and images... i finally did it. huge update coming soon.
@MichaelStolarzi mean how difficult could it be to make your own pdf extractor? all the npm pdf extractors are absolutely dogshit
@MichaelStolarzo no i think i ate too much creatine
@MichaelStolarzall this AI hype around gpt-4.5 / gpt-5 with no release is cool the first time, eh the second time, cringe the third time, annoying the fourth time, boring the fifth time...
@MichaelStolarztoday's AI context windows are like the first 16MB flash drives back in 2000-2002.
except we won't have to wait 10 years for orders of magnitude improvement in capacity. soon, a "context window" will be virtually limitless
@MichaelStolarzbeen using this all day today and... yeah it's performing amazingly so far code-wise, going to stress test it even further. it's so super safe though, that's the only downside
@rubenhassid
gemini 1.5-pro-exp outperforms gpt-4o. Some say it's the end of OpenAI reign. So I ran my own tests: test #1 → write a viral twitter thread
@MichaelStolarzyeah a lot of people seem to forget that gpt-4o is any in, any out
@apples_jimmy
Hot take, I want the promised 4o image capabilities more than voice. Something magic about the prompt accuracy and working in tandem with an alien ai to create. “Art is the highest form of hope. What we are doing here is hope work.“ x.com/gdb/status/179…
@MichaelStolarzIt’s been months since I’ve used OpenAI anything for something. It’s all Sonnet 3.5.
@MichaelStolarzsmall tiny eensy weensy puny little chance that this also drops an upgraded OpenAI model. but... the chance is NOT non-existent!
@adcock_brett
fyi, Figure has some very special announcements coming The engineering team has been working their assess off the last four months Tomorrow, you'll see the beginning of something that is nothing short of magical 🎩
@MichaelStolarzunpopular opinion: if this is how you react when you spend your money on something like a PS5, the last thing you need in your life is a PS5.
@MichaelStolarzmind blowing how cheap intelligence is becoming. soon $1/$3 will be deemed too expensive for SOTA.
@legit_api
Gemini 1.5 Flash is going to have significant price cuts on August 12th in Vertex AI Input cost: $0.0375 per 1M characters Output cost: $0.15 per 1M characters HALF THOSE PRICES FOR < 128K CHARS
@MichaelStolarzso... idk why no one is talking about this but... i replaced my daily $10 API bill by just replacing gpt-4o with gpt-4o-mini... now I'm paying $0.04? a day? hello? without loss of quality?
@MichaelStolarzlight mode is actually goated
@MichaelStolarzthis looks like a google killer, I wonder how it does with more obscure searches that are common among developers, like what happens when you paste in a large debug error from console? or ask a documentation question? or ask for api specifics on some obscure package?
@yoimnotkesku
SearchGPT widgets look nice It tends to focus on the results strongly, so sometimes I need to tell it exactly what I want it to do rather than what I want from the web
@MichaelStolarzQuickTime? More like WasteOfTime.
I've just wasted all day because for some damn reason QuickTime doesn't play audio on certain MP4s. Not until I opened it with IINA because I thought, hey, you know what, why not? Maybe this is one of those codec issues.
Boom, it works.
@MichaelStolarzi wonder how people pronounce @yacineMTB (kache) - I pronounce it "KA-chey", almost like "catchy" but the a is like the "a" in "alone"
or maybe it's more like "KA-shey"? instead of chey? or... what if it's like cache? "KA-sh"? literally cash? do I just call him cash? money? help
@MichaelStolarzwarp isn't that bad actually
@MichaelStolarzholy moly! not one, but TWO blog posts today! woah!