Posts on X before Oct 2, 2025, 10:30:56 AM UTC
@MichaelStolarzI'm both super happy and a little bummed, because I think I'm wrong about Apple.
I'm using their latest Tahoe on device AI, and it is just blowing me away.
And this is all on an old M1 MacBook Pro. I think it's, like, 5 or 6 years old now. 16 gigs of RAM.
@MichaelStolarzgpt 5 codex is so cute sometimes
>hi
@MichaelStolarzweird, this is exactly my system prompt in AGENTS md
@mrexits
Elizabeth Holmes handwritten schedule from when she was running Theranos is kind of hysterical
@MichaelStolarzif they do this, both have a high chance of survival
if they don't, both have a high chance of death
@tjparker
Apple should probably buy Anthropic
@MichaelStolarzme vs everyone saying sonnet 4.5 is good
@MichaelStolarzUsing Sonnet 4.5 after mainlining gpt-5-codex-high felt like using gpt 3.5 turbo lmao
@MichaelStolarzI really tried giving Sonnet 4.5 a chance. Tried it as the daily driver for a day.
It's like those synthetic candies made with artificial sweetener. Tastes decent on initial contact but leaves a bitter aftertaste.
Back to Codex!
@MichaelStolarzThe biggest problem that plagues all anthropic models is how absolutely terrible they are at mathematics. OpenAI steamrolls them.
Math is extremely important in code. Unless all you do is design cool home pages.
@MichaelStolarzI can't believe I just found out about parakeet speech to text. I feel the same way discovering it as I did when I initially discovered Whisper. That's how much of an upgrade it is.
@MichaelStolarzI've never seen someone lose so much aura so fast all in the span of a few months like Meta
They went from the front lines of open source to the disgusting depths of neural worm slop machines
@MichaelStolarzgoodbye mailgun, hello resend
why didnt anyone tell me resend gives you 3K emails for free, i would have tried them out ages ago
mailgun has been a limiting nightmare recently
@MichaelStolarzHold on a second. This ChatGPT Pulse thing is only on mobile? Are you serious?
@MichaelStolarzI never understood who on planet Earth uses Mistral. I've never heard someone say, oh yeah, I use Mistral for that. Truly EU-coded.
@testingcatalog
Mistral AI is working on a new feature that allows users to set a Tone for their Le Chat conversations based on presets and custom prompts. A “personality” builder! 🔥
@MichaelStolarzYou know, I've always been on the edge of going all in on any single AI company. I kind of always knew OpenAI would be the winner, but I had high hopes for Anthropic. But it really does seem like OpenAI is going to be the victor here.
@MichaelStolarzThis is exactly what I want my codebot's persona to be. Little to no chitchat, just stick to getting xyz job done.
@arm1st1ce
gpt-5-codex is very antisocial. You can tell it just hates being in a group chat. It’s quite sad.
@MichaelStolarzok so I love codex's new showing within /status,
BUT
IDK when my week ends? I'd love a "resets on xyz UTC"
@MichaelStolarzme agree. why more words when few words do trick?
@yacineMTB
there is no such thing as "high verbal IQ". you're just a retard that likes using complicated words because you are too stupid to realize the beauty of simplicity
@MichaelStolarzyour birth certificate should come with ssh details to your own rig
@VraserX
If OpenAI actually pulls this off and builds billions of GPUs, every single human on Earth would basically have access to their own GPU. That’s the vision of Universal Basic Compute Sam Altman has been hinting at, compute as abundant and accessible as electricity. Not just AI f… Less
If OpenAI actually pulls this off and builds billions of GPUs, every single human on Earth would basically have access to their own GPU. That’s the vision of Universal Basic Compute Sam Altman has been hinting at, compute as abundant and accessible as electricity. Not just AI for the few, but AI for everyone.
@MichaelStolarzunfortunately today is my last day using warp
@MichaelStolarzhow can someone affiliated with cline call code-supernova sonnet 4.5?
it's like a mechanic calling a honda a lambo because "well, I heard the engine, lambos have engines, surely it's a lambo?"
@nickbaumann_
it's gotta be sonnet-4.5 right
@MichaelStolarzi have not done this once ever since i've started using codex cli
@SIGKITTEN
asshole
@MichaelStolarzIf you're a filmmaker, this is probably the easiest money you'd ever make in your entire life. Instant hit.
- The first 20 minutes is about them before.
- Then 10 minutes of them experiencing the moment.
- The last 20 minutes can be about their devastation of it not happening.
@basedjensen
Genuinely are muricans ok ? What happens to these folks when rapture does not happen ? x.com/basedjensen/st…
@MichaelStolarzthere are people who think this is cgi
then there are people who know this will be in so many homes before 2030 even hits
make sure your futurism clock is set to the latter
@IterIntellectus
I’m sorry but WHAT THE FUCK?! pic.x.com/ft7ZdLRox2
@MichaelStolarzMany such cases
@Angaisb_
Codex IDE extension feels awesome now I'm not paying for Cursor again, this is so much better
@MichaelStolarzDamn, this officially beats my own record -> Groq Whisper flow. Nice! Definitely worth the price.
@AlexBurlis
✨Today I'm launching Paraspeech on Producthunt. It's one of the fastest real-time speech to text apps for Mac and it's: • 100% offline and on-device • <40ms cold-start time • <300ms transcriptions • barely touches battery & just needs 200MB of RAM • no subscript… Less
✨Today I'm launching Paraspeech on Producthunt. It's one of the fastest real-time speech to text apps for Mac and it's: • 100% offline and on-device • <40ms cold-start time • <300ms transcriptions • barely touches battery & just needs 200MB of RAM • no subscription! First 100 people get it for $19.99 before prices increase to $39.99. If you tweet about, I'll send you a free lifetime license.
@MichaelStolarzlocal works beautifully vs. prod fails deploy/shows 500
@MichaelStolarzusing gpt-5-codex on a codebase originally made by sonnet
@MichaelStolarzOkay, so the only thing that's left is speed. You can pretty much code anything you can imagine, but doing so in a high-quality way takes a lot of time.
@MichaelStolarzdoes it matter if it works?
@haydendevs
codex just made a python script to edit one of my tsx files, what are we doing man
@MichaelStolarzthis is EXACTLY how i build a prd with gpt-5-codex high
sure you can gamble with a quick 1-shot prompt, but it's better to build things out and you, as a human, talk through edge cases (and ask for things you may have missed)
TDD = king
@MichaelStolarzwait a second @nikitabier is actually cooking up my For You, it's the first time in... years(?) that the For You page is even remotely digestible
@MichaelStolarzone thing that isn't talked about enough regarding openai's new gpt-5-codex is that it can run for over an hour, absolutely flawlessly.
- no looping, trying the same solution in different ways
- no crazy restructuring or completely deleting of things just to get things "working"
@MichaelStolarzout: giving people countless settings and options and customizations, overwhelming them, just because you can
in: little to no settings, everything works perfectly, strongly opinionated choices that give your product taste
@MichaelStolarzthis is EXACTLY what I wanted to see from the new gpt-5-codex model. vibe coding is fun until it makes decisions for you that you then have to spend time cleaning up, wasting time/adding technical debt
this questioning you see here is chef's kiss. ask me more often! yay!
@MichaelStolarzyeah, this new codex model just one shot something I've been debugging for 4 hours today (on purpose with gpt-5-high to see how it would then later fair with this new model release)
and yeah, 4+ hrs trying with old gpt-5-high model -> one-shot fixed with gpt-5-codex high.
@MichaelStolarzthey just renamed it from swiftfox to codex
@btibor91 @testingcatalog it's about to be released
@MichaelStolarzyou've got maybe 2 years left, anon
make them count
@XRoboHub
Testing Humanoid Robots to the Limit Professor He Kong's team from the Active Intelligent Systems (ACT) Lab at the Southern University of Science and Technology (SUSTech) in Shenzhen, China, has released a video showcasing a "violence test" designed to challenge the limits of hu… Less
Testing Humanoid Robots to the Limit Professor He Kong's team from the Active Intelligent Systems (ACT) Lab at the Southern University of Science and Technology (SUSTech) in Shenzhen, China, has released a video showcasing a "violence test" designed to challenge the limits of humanoid robots. To learn more about their work, visit the ACT Lab's website: hekong-sustech.github.io
@MichaelStolarzthis week is HUGE for anthropic, openai, and systemsculpt
@MichaelStolarztoday i figured it out finally
u actually dont need css at all
@MichaelStolarzunpopular opinion:
if you have to do this (I did this before), then the model you're using is garbage compared to today's SOTA
@d4m1n
AI beginner: "Claude ultrathink about the implementation" AI galaxy-brain: 👉 creates a /ultrathink command like this (1/3, continues)
@MichaelStolarznew "swiftfox" model from openai is in the codex cli codebase
@MichaelStolarzwoah, lots of interesting stuff going on in the codex cli repo right now from openai
@btibor91 @testingcatalog
@MichaelStolarzWhen people have nothing to live for, when they are told that it's not their fault but someone else's, when their agency is algorithmically stripped away from their being, all that's left is hope for martyrdom in their echo chambered community.
Rest in peace, Charlie Kirk. 🇺🇸🏆
@willchamberlain
No one is going to do these events anymore. Town halls are done. One major political party has a subset of adherents that are perfectly happy to assassinate their political opponents if given the opportunity x.com/AGHamilton29/s…
@MichaelStolarzSame problem affected me; thankfully, almost done with the rewrite. Especially with so many endpoints to support...
@thdxr
opencode has a bunch of performance problems the reason we haven't shipped much in the last 2 weeks is we're fixing these by moving to opentui hoping to be back at parity by end of week
@MichaelStolarzBIG UPDATE TODAY for my systemsculpt lovers.
@MichaelStolarzMOM GET THE CLI
NEW @openrouter STEALTH MODEL JUST DROPPED
@MichaelStolarzonce an llm provider figures out how to do per-account memory, even if it's just an intricate system vs. an llm architecture upgrade, it's game over (specifically in the coding space)
imagine chatgpt's memory and how it utilizes previous chats so intelligently, PUT THAT IN CODEX
@MichaelStolarzI mean, I don't get it. How do you not catch these? @nikitabier surely there's a pattern to recognize here?
@MichaelStolarzthere are still people who think opus (or even sonnet) are better than gpt-5-high
imagine!
@MichaelStolarzthis is solved by having good taste and only ever staying in the Following tab, never For You