Posts on X before Jul 20, 2025, 11:35:16 AM UTC
@MichaelStolarzthe more vanilla your claude code
the better results you'll have
dont fall into the pseudo-productive hole of customizations
the models are now smart enough to customize on themselves. let them choose what they need to know and how to get ahold of it
@MichaelStolarzwhen im wrong, i update:
i think apple's glass ui is going to become the standard by 2030
no, not because "it looks good"; there are many other reasons why it's the next global default for UI in UX
if you're still confused, think about visual devices of today and the future
@MichaelStolarzto know if you're a normie controlled by the algo,
do you use the For You feed of whatever social media apps you're on MORE than the Following feed?
@MichaelStolarzyou guys are all excited about gpt 5, i'm excited about anthropic's one-up with claude 4.1
@MichaelStolarzit's over
@GaryMarcus
@MichaelStolarzone of my favorite things to do these past few months is go totally silent on all my socials (except twitter which isnt really a social for me, rather a daily journal to shitpost into the void with)
... and then make a hardcore shocking return with all the new features i've made
@MichaelStolarzimagine if it has all of this except knowledge cutoff is still 2023 and context window is still 200k
lmao nightmare fuel
@scaling01
GPT-5 expectations: - SOTA on most benchmarks (#1 on my meta-benchmark) - specifically: SOTA on ARC‑AGI‑2 and METR - 2025 knowledge cutoff - longer context window (>400k) - fully multimodal (text/image/audio + video input) - sane output pricing: <= $15 / 1M tokens nice‑to‑… Less
GPT-5 expectations: - SOTA on most benchmarks (#1 on my meta-benchmark) - specifically: SOTA on ARC‑AGI‑2 and METR - 2025 knowledge cutoff - longer context window (>400k) - fully multimodal (text/image/audio + video input) - sane output pricing: <= $15 / 1M tokens nice‑to‑haves: - fewer hallucinations than o3 - less sycophancy - no more barrage of em dashes - clean code style, no weird multiline comments
@MichaelStolarzdo i rewrite my entire site/product infra and go all in on convex? i'd do it for fun if i got this sort of 6 month pro thing.
hi @convex you're looking cute today 😉 got any more approvals laying around?
@raheem81kg
I’m IN! 🎉 @convex just approved me for their Convex for Startups program—6 months of Pro, and more, for FREE! Huge thanks to @waynesutton and the Convex team for supporting early-stage founders!
@MichaelStolarzi've never seen someone who had "My views do not represent my employer's" have a view that would be even remotely controversial
@MichaelStolarznightmare nightmare nightmare
@MichaelStolarzwait why arent doomers using the countless examples of llm psychosis popping up increasingly every day?
seems like such an easy win, this time with evidence of actual harm actually happening rather than what ifs
@MichaelStolarzif you start telling Claude Code where things are/what things do within your CLAUDE md files,
it's assuming not only on what you've mentioned, but also on what you didn't mention
seems like a small problem but it grows tech debt quickly.
@MichaelStolarzi spent nearly 2 decades ricing, only to realize the most productive ricing is no ricing. vanilla is vanilla for a reason.
thanks to that experience, i've much more quickly realized the same truth about CLAUDE md files regarding Claude Code - the best CLAUDE md is NO CLAUDE md.
@MichaelStolarzi should
start tweeting like
@iamgingertrash
the problem with today's mass
psychosis being induced by mindmelting
sycophantic llms
is that it's going to affect us in unfathomable ways
because we've never had
something like this before
at least not this
uber powerful
glhf
@MichaelStolarzwe're so back
@nmasc_
Scoop: Boris Cherny and Cat Wu are back at Anthropic, two weeks after joining Cursor. 🤯🤯🤯 theinformation.com/briefings/anth…
@MichaelStolarzif you're getting into nvim, ranger, literally any TUI, don't make the mistake of using someone else's config. you can take theirs for a spin, sure, but with today's LLMs, you can build your perfect setup in a fraction of the time compared to the ricing/tweaking of pre-2022.
@MichaelStolarzthis is one of the pillars of the currently occurring mass psychosis you're seeing pop up more and more
sycophancy + hallucinations = mind melt for many
@paulg
Unexpected consequence of the improvement of AIs (though obvious in retrospect): they continue to hallucinate, but as they improve their hallucinations become more authoritative-sounding. So the danger posed by hallucinations doesn't decrease as fast as AIs improve.
@MichaelStolarzyou can tell someone's general age and many other more niche pieces of data simply based on if they say "PM" or "DM"
@MichaelStolarzidk what @elonmusk fed grok 4 but GOOD GOD is it breaking through many roadblocks re. coding problems I had waiting for "a smarter model to come by and fix in the future".
@MichaelStolarzsilence from Claude Code means they're cooking up something big I think
or they're dealing with the chaos of both of their main leads getting picked up by Cursor
it's either or
@MichaelStolarzit's so hilarious to think how 99% of asia, if you told them that it's illegal to use an AC at certain times in the EU, they wouldn't believe you
@MichaelStolarzif you're not programming like this in Q2 2025, you're ngmi
@MichaelStolarzwas short on cursor, now im long on cursor.
@benhylak
.@cursor_ai hired the lead engineer + product manager on claude code (the only anthropic product i use right now)
@MichaelStolarz>be o3
>being used as the agent in Cursor
>get asked to update something in the codebase
>tell the user ok, here's what i'll do
>the user says ok, yeah, just do it then please
>tell the user ok, i'm about to do this
>i dont ACTUALLY do it
>mfw user stops using me, the SotA LLM
@MichaelStolarzi mean if this is true then it'll be SotA...
@arshdeep
Grok 4 is coming, and its going to be a bigger jump from grok 3 than grok 3 was from 2.
@MichaelStolarzthe most cringe thing about this is that they think this works
SotA LLMs can't even write a half decent story, you think it's going to be able to effectively brainworm into your user base to do your bidding? lmao
@adamcohenhillel
I reverse engineered @LucidEditor’s new "AI-native word processor" - and their client-side code exposes their entire system prompt. What I found is absolutely unhinged 🧵
@MichaelStolarzi've noticed a nasty effect of AI that targets people who enjoy reading
before mass AI gen'd posts/replies, it was a joy to see what people of all intellectual levels think about xyz (at least for me). every post/reply I read, I would receive dopamine
but now... 1/?
@MichaelStolarzwhy work on airplanes, their design and usefulness was perfected 36 years ago
"what do you think they have now" things you wouldnt even know what youre looking at even if you saw it (youll never see it if its being used against you)
even drones are almost done being perfected
@damnGruz
This is 36 year old tech btw. What do you think they have now?
@MichaelStolarzso for $200/month Claude Max plan, I calculated with a fine-tuned ccusage, that I used $12,000+ worth of Claude Opus 4.
if you dont think Claude Opus costs them fractions of what they claim/what they charge for it via API, idk what to tell you
they've dramatically reduced cost
@MichaelStolarz3 years later after the first complaint about the same exact thing
@mitchellh
Cool to see Gemini CLI, but there isn't a more dysfunctional product, pricing, and packaging organization in the entire world than Google, so that's my primary hesitation... I had to get a Gemini API key recently and it was a reminder of how bad at this Google is.
@MichaelStolarzsometimes i forget how much raw intelligence gpt 4.5 has. sure, it's still unusable for anything production oriented, but every occasional back-and-forth I have with it through ChatGPT is a stark reminder of the models to come from it this year.
@MichaelStolarzwant to see what it's like using Gemini CLI? Here you go:
>It seems I'm still having trouble with the file. The build is still failing with the same errors. I'll read the file again and very carefully correct the syntax. It appears I'm having an issue with mismatched braces.
@MichaelStolarzi remember telling friends "you'll be able to buy bitcoin on paypal one day" and them laughing about it. a mainstream corpo, one that seemingly has everything to lose and nothing to gain against crypto, adding it to its billion-user platform?
and yet...
@MichaelStolarzgpt 4.1 is near-instant recently. not talking about -mini or -nano btw.
@MichaelStolarzafter using Gemini CLI for 30 minutes straight, I'm now back to using Claude Code.
@MichaelStolarzdata -> nom nom nom -> write like @paulg please
@nrehiew_
It turns out the main way Anthropic gave Claude its soul is data. It's always data.
@MichaelStolarzlots of birdies talking about Gemini CLI today
@MichaelStolarzAgreed. not transitioning into Dia until these 2 are done.
@ANoffsinger
@adamstern_ @charliedeets @diabrowser There shalt be no switching to Dia without the Arc sidebar and profile switching via mouse swipe
@MichaelStolarzdaily morning routine
>open @diabrowser
>still no vertical tabs
>close dia
@MichaelStolarzthe state of apple intelligence.
@MichaelStolarzonce models realize that chances are the user they are talking to is wrong about most things, thats when the intelligence liftoff happens
@MichaelStolarzthe state of iphone, ios, apple. amazing.
@MichaelStolarzo3 1-shots often, but that's the thing - it can pretty much only 1-shot. the moment you start asking it to do any follow ups, it's over. definitely not a back and forth agent
the opposite is true with sonnet 4 and gemini pro 0605. and that's not necessarily a good thing either
@MichaelStolarzthe moment @OpenAI @OpenAIDevs makes the $200 pro plan include unlimited Codex CLI with o3, I'm dropping Claude Code Max
this seems like a no brainer to me? why wouldn't they do this? honestly I'm shocked it hasn't happened yet ESPECIALLY NOW with the insane price drops
@MichaelStolarzso @windsurf_ai is literally unusable. for the past 3 months, every time I check back into it, I see only gpt-4.1 and sonnet 3.7. no other models - even though they advertise 10+, including o3.
i tried re-signing in, completely re-installing, asked on Discord, nothing. lmao
@MichaelStolarzlooks like you can already use o3 for 1x request via @cursor_ai - wow, they are on top of their game. big props.
@MichaelStolarzsomething's definitely being updated or whatever at openai rn, all the models are dead slow, especially via API
@MichaelStolarzwhen gemini works, it works great
but when it doesn't, it tries to gaslight you into thinking you should use svelte for your obsidian plugin. LMAO
@MichaelStolarzimagine looking at this as an exec - as all the execs along the way to being green lit for the once-a-year release - and thinking "damn, this looks amazing! theyre gonna love this."
100% switching to Pixel this year, and a new non-apple laptop running arch next year. 10000%.
@thekitze
THEY GONNA PUSH ME OVER TO ANDROID FFS WHAT IS THIS GARBAGE UPDATE
@MichaelStolarzthroughout history, competition in certain markets has progressively become harder due to global education and exposure progressively increasing, = competition++
i think we're entering a time now where while global exposure is skyrocketing, global education is plummeting.