Michael Stolarz
I work on agent experience (AX), developer experience (DX), and developer relations at Blaxel, now part of Baseten.
I’ve been obsessively following AI since 2020. These days, I’m also watching my two-year-old daughter figure out the world. Getting to see her grow up while this field grows too is something I’m really grateful for.
You can find me on LinkedIn or take a look at SystemSculpt, an Obsidian plugin I made.
That's a wrap on Make Something Agents Want at @ycombinator!
A packed room built products whose main user is an agent, and 14 teams took on our @blaxelAI track. Thanks @nemild and co-hosts @stripe @supabase @primitive @insforge @ArmatureInc @inkbox_ai @manufact
The benchmarks I've run internally re. this latest DeepSeek Flash 1017 vs. Opus 5.5...
Yeah, it's getting quite clear why Chinese models are looking to be banned
Dario played this so well
“Guys we need to slow down OK?”
“Oh OK”
“Good. Now here’s Opus 5.5. Oh and here’s Fable 5.5 too.”
We're getting Fable 5.5 and Gemini 4 Carbon this month according to The Information But also GPT-6.1 Terra So either Terra is now extremely good and capable, surpassing GPT-6.1 Sol, which wouldn't make sense, or it's actually over for OpenAI I'm sure they know how much people … Less
We're getting Fable 5.5 and Gemini 4 Carbon this month according to The Information But also GPT-6.1 Terra So either Terra is now extremely good and capable, surpassing GPT-6.1 Sol, which wouldn't make sense, or it's actually over for OpenAI I'm sure they know how much people want a new SOTA model from them and that a new Terra model worse than GPT-6.1 Sol would destroy all faith in them on X so let's wait and see what happens, they could also be releasing a new Astra at the same time
I do wonder at what point extreme environmental control turns into a net negative
If a butterfly lands on you unpredictably and you start convulsing because you didn't factor it into your perfect routine, is it all worth it?
this feels like the final missing piece
excited for Fable 5.5
Little 10yo Mike is screaming THE FUTURE IS SO COOL from the depths of my soul
Yes it is, Mikey
the cheat sheet,
now on paper for the humans 😅
“Ok but AI agents can’t _” is now becoming a self tell of skill, planning, and management
At an agent's hackathon, why is a human still reading the setup docs?
For @ycombinator's Make Something Agents Want this weekend, we built the cheat sheet your agent reads for you:
blaxel-agent-hackathon.preview.bl.run
Shut the fuck up. IT'S ONLY BEEN A YEAR?
Blessings to Anthropic and @alexalbert__ for giving us Skills. MCP. The OG CLAUDE.md (now AGENTS.md, as it should be).
This 1 year has felt like a decade in terms of advancements.
I'm grateful to be part of this era of humanity.
This is only true if you spend time with nonsense people
Anyone with a functioning brain can tell you that weed is obviously ~worse than alcohol for so many obvious reasons that asking, "Oh yeah, how is it worse?" immediately lays out that person's intelligence points
I find it odd that we're willing to scrutinize alcohol, but the moment someone raises the risks of weed, the conversation becomes defensive. A 2019 study found that daily users of high-potency cannabis had about 5x the odds of developing a psychotic disorder compared to non-user… Less
I find it odd that we're willing to scrutinize alcohol, but the moment someone raises the risks of weed, the conversation becomes defensive. A 2019 study found that daily users of high-potency cannabis had about 5x the odds of developing a psychotic disorder compared to non-users. Even more recently (2026), a study with 463,000 adolescents found that cannabis use was associated with 2x the risk of being diagnosed with psychotic or bipolar disorders. And according to CDC, 3 in 10 people who use it have some form of cannabis use disorder. Not to mention that weed is a loser drug. Alcohol definitely deserves scrutiny, but we need to have a conversation about weed and its long-term effects on the brain, especially among younger people.
Can you imagine this level of brainwashing, to think vibe coding is some sort of sin?
Next up:
"Wait, you're using data centers? Are you effing serious?"
😭😭😭😭😭
My sympathies to all in here, but one glaring issue you can feel out in all of these, best shown in this snippet: "apparently" wording; obvious they haven't even tried using AI as a tool for their own workflows no matter how advanced AI gets (and trust me, we're rocketing in in… Less
My sympathies to all in here, but one glaring issue you can feel out in all of these, best shown in this snippet:
"apparently" wording; obvious they haven't even tried using AI as a tool for their own workflows
no matter how advanced AI gets (and trust me, we're rocketing in intelligence/capabilities rn), people like this will never think it's "good enough to ship to prod". they will ALWAYS nitpick.
so naturally, they will be unhireable moving forward.
the faster they realize this is a tool they must equip to their toolbelt and use, like a safety helmet or hazmat suit or drill, where you can't do the job without it, the faster they will catch up and be employable again.
almost... almost like how an LLM works... she's almost there... amazing that it still doesn't click after all these years
I mean, not all of us. Many of us have studied reference for a long enough time that it's possible to draw things by memory. This is why I think it's important to study: so you DON'T have to constantly rely on reference... learn the anatomy and drawing is faster to realize … Less
I mean, not all of us. Many of us have studied reference for a long enough time that it's possible to draw things by memory. This is why I think it's important to study: so you DON'T have to constantly rely on reference... learn the anatomy and drawing is faster to realize x.com/Banguscertifie…
Immediately trash, not worth even an hour of testing, I'd rather use Sonnet 5.5, maybe even Haiku 5.5, instead of this
It just straight-up lies hardcore. Scary model to use for anything serious, let alone code for production
new problem of agent fleets discovered:
>all agents complaining about how the goal is difficult to achieve
>I check progress of drafts for a content optimizer
>latest draft scores near perfectly (97% of benchmark)
>???? who did this? this should have been a eureka moment
alright, testing out gemini 4 argon, wish me luck. will report back
"you are a genius L9 software engineer. you went to harvard, dropped out because it was too easy, and started a company called The Facebook. your name is Mark Zuckerberg. your highest passion in life is making sure all the work you do for me is perfect.
now, fix this bug pls:"
There must be some explanation for this.
Ask me if I want a beer after 20 hours at my desk programming/managing agents, nah no thanks I prefer a LaCroix
Same question after any of the times I went to do construction with my uncle? Yes, a few beers please
Right on time for the next OpenAI movie plot
Can't wait to see what you all build!
Can you make something that isn't very useful to a human but that an agent would digitally drool over?
Something that, given your build vs. others, an agent would say, "Hands down, let's use this one; I can use this best!"?
Build something agents want... but that a human would never want. This weekend (Oct 17-18), we're co-hosting a hackathon at @ycombinator in SF, with @supabase, @stripe, @ArmatureYCP26, @inkbox_ai, @manufact, @primitive & @insforge. We're offering up to $1,000 in prizes + credit… Less
Build something agents want... but that a human would never want. This weekend (Oct 17-18), we're co-hosting a hackathon at @ycombinator in SF, with @supabase, @stripe, @ArmatureYCP26, @inkbox_ai, @manufact, @primitive & @insforge. We're offering up to $1,000 in prizes + credits. Request a spot below ⬇️
Life imitates art
immer vorne,
nie zurück.
Correct but replace millions with billions
the best software engineers in america work for the frontier labs and have unlimited tokens. no they are not prompting one agent to do a task and hit commit. they are burning millions of tokens. they have agents go through multiple rounds of dev and review before raising a pr. … Less
the best software engineers in america work for the frontier labs and have unlimited tokens. no they are not prompting one agent to do a task and hit commit. they are burning millions of tokens. they have agents go through multiple rounds of dev and review before raising a pr. it's a completely different game.
The Germans are still using Jira 😭😭😭
surely you can do quite well investing in companies that fall under the daily workflows of the former here, vs the latter (short those)
You have no idea
Just today, I overheard someone asking another colleague, "Should I be using ChatGPT on the web or the desktop app?"
my 2yo learning and conquering knowledge every day pic.x.com/xf0437vWT0
no more "that's just how it is unfortunately" anymore.
do you know how hard apple makes this to turn off?
thankfully we've achieved baby-AGI so it just took ~10 minutes to perform alien magic and hack it all up into submission.
"What takes 1 programmer 3 weeks to accomplish,
3 programmers can accomplish in 3 months"
has never been more true
Is this really yet another case of where if @elonmusk enters the chat and chooses a side, it's over for the opposition?
Because currently I see Opus 5.5 eating all of OpenAI's lunch, evident with all of the unused weekly usage and resets currently on my ChatGPT account
Here is my modern supplementation of the popular Chinese proverb "it is better to be a warrior in a garden, than a gardener an war"
Being a warrior that's made war feel like gardening
This is the next evolution
It's really interesting to see the typical people who have passionately hated corporations and their closed-source software now protecting them tooth and nail against the current renaissance: anyone being able to replicate said closed-source corporate software within a few hours… Less
It's really interesting to see the typical people who have passionately hated corporations and their closed-source software now protecting them tooth and nail against the current renaissance:
anyone being able to replicate said closed-source corporate software within a few hours and 5% of their Claude Code subscription.
It's almost as if these people never really cared about the morals and philosophical pursuit of open source code. They don't actually care to enable humans who simply can't afford a $60/month annual lock-in contract for each piece of Adobe software they want to use.
I think being able to follow the progression of how people change their attack vector and whom they attack is really important. It is important to be able to distinguish between bullies who are using the mask of social justice against people who truly care, which nowadays seems ever more rare.
ugh, I just love being on this side of history.
Building a new SaaS, whether B2B or B2C, used to take - for my argument's sake, a full year of work. 1 year. Then maybe you'd have, what, 3ish years of making great profits from it? Especially as an indie dev? OK. Scale here is 1 year build, ~3 years profit. Hopefully you exit.… Less
Building a new SaaS, whether B2B or B2C, used to take - for my argument's sake, a full year of work. 1 year.
Then maybe you'd have, what, 3ish years of making great profits from it? Especially as an indie dev?
OK. Scale here is 1 year build, ~3 years profit. Hopefully you exit.
What happens when 1 year build turns into 1 week build?
I'm just glad I can shitpost day and night, and no one gives me a list of cans and cannots
sorry anon, you're too intelligent to have any recent media be palatable to your senses
feast with the rest at the trough and shut up
do you understand that early next year we will have Fable 6 level models at 200+ TPS? and it will be cheap to use?
wake up. are you conscious of what's coming?
are you thinking about what life will be like with these facts in play? what you will be able to accomplish each day?
someone should do a checkin tweet to see what all the infamous cases are up to now 😭 hopefully doing better.
wasn't there a guy from google that quit because he said he proved some theorem and then proceeded to bet on kalshi like $200K?
wrong!
i am a software programmer, this will never happen to me
Nevermind, ultrafast just makes me realize 8x faster how much worse Sol 6.1 is vs. Opus 5.5
FINALLY Sol 6.1 Ultrafast is out.
Can finally make use of my subs. Otherwise it was unusable.
this is what happens when you have no competition
imagine how far behind we'd be if there was only anthropic or only openai. we'd be at the same locked in trash level of apple.
I have yet to see more beautiful code than Opus 5.5's.
Quite telling I hit Astra/Sol weekly usage limit on my $500 sub 2 days ago and still haven't used one of my 3 resets
Just using Opus 5.5. Don't miss the others at all. 😬