Sabitlenmiş Tweet
on january 24th i asked gpt-5.2-pro to predict the outcome of the tariff case before the supreme court.
gpt absolutely nailed it.
English
mark erdmann
4.9K posts
mark erdmann
@markerdmann
co-founded @pulley in 2019. grew to 5k happy customers and $XXm ARR. now on pat break with tiny new human. 👶 exploring - ai eng, voice agents, edu games
Calgary Katılım Aralık 2008
2K Takip Edilen1.4K Takipçiler
MiMo-V2-Pro & Omni & TTS is out. Our first full-stack model family built truly for the Agent era.
I call this a quiet ambush — not because we planned it, but because the shift from Chat to Agent paradigm happened so fast, even we barely believed it. Somewhere in between was a process that was thrilling, painful, and fascinating all at once.
The 1T base model started training months ago. The original goal was long-context reasoning efficiency. Hybrid Attention carries real innovation, without overreaching — and it turns out to be exactly the right foundation for the Agent era. 1M context window. MTP inference for ultra-low latency and cost. These architectural decisions weren't trendy. They were a structural advantage we built before we needed it.
What changed everything was experiencing a complex agentic scaffold — what I'd call orchestrated Context — for the first time. I was shocked on day one. I tried to convince the team to use it. That didn't work. So I gave a hard mandate: anyone on MiMo Team with fewer than 100 conversations tomorrow can quit. It worked. Once the team's imagination was ignited by what agentic systems could do, that imagination converted directly into research velocity.
People ask why we move so fast. I saw it firsthand building DeepSeek R1. My honest summary:
— Backbone and Infra research has long cycles. You need strategic conviction a year before it pays off.
— Posttrain agility is a different muscle: product intuition driving evaluation, iteration cycles compressed, paradigm shifts caught early.
— And the constant: curiosity, sharp technical instinct, decisive execution, full commitment — and something that's easy to underestimate: a genuine love for the world you're building for.
We will open-source — when the models are stable enough to deserve it.
From Beijing, very late, not quite awake.
English
@GroundhogStrat @corbtt hm the caltrain does stop a lot. i could see that being a big hit with the protestors.
English
@markerdmann @corbtt It’s faster to take the Caltrain bullet train than Waymo. Also you should know that Waymos are not AGI, and the runs to train them fall well below the 10^26 FLOPs limit in most pause proposals!
English
turns out that google's core structural advantage in the AGI race is that its hq in mountain view is too far away from sf for the pause ai people to get to
David Krueger@DavidSKrueger
A week from today, we will be at Anthropic, OpenAI, and xAI, demanding that leaders agree to a conditional AI pause. These companies are recklessly endangering all of our lives. Their excuse is that they can't pause unilaterally. So they must commit to pausing if others do.
English
@lacker although right now the emerging bottleneck is code review, so maybe it makes more sense to optimize for "easy for humans to read"?
x.com/gabriel1/statu…
gabriel@gabriel1
only bottleneck is consuming code, so make sure to tell codex that you want just that: "write extremely easy to consume code, optimize for how easy the code is to read. make the code skimmable. avoid cleverness. use early returns."
English
Eventually the agents are going to write most of the code. What sort of programming language do they prefer?
English
@DavidSHolz @max_spero_ is it mean regression due to the l2/mse loss function?
or maybe denoising happening in sRGB instead of linear RGB?
English
@max_spero_ It's funny cause I totally know whats causing it but I feel like I can't say 😅
English
I die a little inside every time I see another piss yellow ChatGPT slop image generation.
Doubly so knowing that midjourney and a little taste can get you so much farther!
theseriousadult@gallabytes
midjourney is still the only image generator that comes in color
English
@aidan_mclau DM: you find a legendary artifact in the dungeon
me: does it write unit tests
DM: it grants +3 to wisdom
me: so no
English
the median high fantasy novel i read growing up does not contain an artifact more powerful than codex
English
@allgarbled i hadn't noticed this but now that you say it it's so obvious.
just showed up again today in a codex UI redesign: x.com/markerdmann/st…
mark erdmann@markerdmann
i gave claude code and codex the same task: redesign the loqi UI/UX. see if you can guess which is which.
English
I don’t know what you call them, but these little side tabs are like the emdash of vibe coded UIs
Juri Strumpflohner@juristr
What's your AI adoption level? (according to Steve Yegge)
English
i gave claude code and codex the same task: redesign the loqi UI/UX.
see if you can guess which is which.
mark erdmann@markerdmann
when to use claude code: design, UI/UX codex: everything else
English
this is what motivated me to hack on loqi, a voice agent optimized for maximum intelligence and long-form conversations. it's awesome for hands-free work on long walks. i use it for 1-2h per day.
if you'd like to try it, soleio, i can add you to the testflight.
x.com/markerdmann/st…
English
I used to think this was a little ridiculous…
…but now I have come to accept the extraordinary wisdom of the laptop harness.
Trung Phan@TrungTPhan
Stephen Wolfram uses a 'walking desk' laptop to maximize productivity. When it is sunny, he will walk around for hours and "type perfectly well" while having "enough peripheral vision" to not trip. You may not like it, but this is what peak performance looks like.
English
@ashleevance dogs won't help us. we fake-threw the ball for 10,000 years and now they have agi and a grudge.
English
I won't really be impressed with AI until a dog uses ChatGPT to cure a human's cancer
English
@CJHandmer do you still have effort set to high?
boris posted about this change recently
x.com/bcherny/status…
Boris Cherny@bcherny
👋 we recently changed the default effort to medium to better balance intelligence with speed, based on median usage we were seeing in the wild. You can change it anytime in the /model selector if you prefer low effort (faster) or high effort (more intelligence). The setting is sticky and will persist for your next session. If you’ve changed effort to high already, can you double check you’re on the latest version, and that your CLAUDE.md’s (in /memory) don’t have confusing/conflicting instructions? That’s the most common cause. [deleted a previous post where I accidentally said default is high, it’s early and I haven’t had coffee..]
English
Claude code is derping tonight. It took 3 weeks but it doesn't feel magical anymore. 3 weeks for the scope of my projects to rise and the precision of my communication to fall enough to get to the point of annoyance.
For context, I have about 1000 legacy projects that I'm re-excavating for new insights or better graphics.
English
@seconds_0 so much of the software builder's job is now qa monkey. 😂
it's gone from ~20% to ~80% of my day.
makes it even more important to follow @paulg's advice in the "how to do great work" essay and work on things you find fascinating.
English
The problem with most people's vibe-coded apps is they don't use them.
Your cycle time is so fast that you should basically have it open on one screen and be doing UI testing and reporting bugs, but nobody wants to be a QA monkey, so they just get bored and quit
English
when to use
claude code: design, UI/UX
codex: everything else
English
mark erdmann retweetledi
GPT 5.4 trounces Claude on mathematical proofs bullshit test. Claude keeps claiming it has proven mathematical statements that are incorrect, failing to spot the fault in the question
Opposite result to BullshitBench where Claude is king
Jasper Dekoninck@j_dekoninck
The results are not only poor, differences between models are also huge: Opus-4.6 completely fails, and Gemini-3.1 performs only half as good as GPT-5.4. The capability does not seem stressed equally among all frontier models, even though it is essential for math use cases.
English
@sailaunderscore top half: policy summit.
bottom half: do you have anything to declare
English
Diabolical work from the Australian Government's graphics team.
Kevin Rudd AC@AmboRudd
The rate of AI innovation is like nothing Marc Andreessen thought he’d see in his lifetime. Australia is making major investments in compute, renewable energy and a smart regulatory framework to support the next wave. Great discussion with you, Marc.
English
mark erdmann retweetledi
@Suhail inference rationing begins. each american will receive a booklet of 50 monthly prompt stamps. black market claude access priced in bitcoin
English
at @pulley we were remote-first and it attracted some absurdly talented working moms who would never have applied to yet another "fun office culture" startup with a ping pong table and a 90-minute commute. there is genuine alpha in being the best remote employer. most founders haven't figured this out yet.
English
Had no idea it’s National Working Moms Day, or that working moms even get a day. Regular reminder that the greatest needle mover for working mothers is working from home.
Incentivize employers to offer work from home as a benefit and most mothers of young children will take it.
English
the good news though is there are always new problems to sleep on, just higher up the stack. you used to go to bed stuck on a bug, now you go to bed stuck on agent orchestration conundrums or what product to build next. as @DavidDeutschOxf says, problems are inevitable, problems are soluble. it's fractal. we solve one layer and unlock exciting some new problems. 🚀
English
I think this collective feeling of "I don't enjoy coding anymore because it's so easy with AI" is good to talk about and realize, and I have it too
I miss going to bed with a coding challenge I have to get through and then wake up and in the shower I get the answer and I scream EUREKA!!!!!
But then you quickly just have to accept that the world has permanently changed now and it's just not going back because letting AI code for you is simply so much faster and effective and will only get better with every passing year
So the better mental approach for me to these things is to just aggressively embrace it and change myself instead, if the fun in solving the challenges is gone, where else can I find the fun?
I'm lucky a bit because for me the fun has always been building new things in general, not so much the coding part, although the coding challenges were fun for me too. But having ideas and just building new things was always the most fun. So I have to double down on that now, making more things and making better things and making them much faster than before.
Especially now that literally everyone in the world has access to the same coding skill as everyone else (which is AI), the focus will have to aggressively be on what remains as a differentiator for me as a creator, which is my ideas and the way I execute them, not coding them
So that's what I will try focus on from now on I think
@levelsio@levelsio
If the skill part of making things moves to the AI Then everyone now has access to the same skills So then it's either not about skills anymore and everyone is competing with everyone on equal footing and all of us ending in a perfect competition with close to zero profit So then nobody ends up winning anymor but the AI companies (since we pay them) Or skill is replaced by stuff like ideas, originality, taste, getting users, attention, distribution, audience, capital (who's rich) etc
English
@karrisaarinen the office exists so you have someone to turn to and say "holy shit did you see what the agents just built"
English
I love this timeline where the future is humans directing swarms of agents building whole products perfectly remotely, but humans must show up to offices for essential “serendipity” and “tap on the shoulder” purposes
English