Senior Director, Developers & AI • Founder of Droid AI • Forbes Tech Council • LLMs • Data • AI/ML • Ex AWS, MSFT Research • Opinions are mine
my terminal
Joined December 2012
- Tweets11.2K
- Following5.3K
- Followers23.9K
- Likes31.2K
Pinned Tweet
New generation of ladies who are gossiping about distributed systems, FP, pragmatic architecture, chaos engineering and fast data analytics. – at Portland, OR
See you soon, San Jose + SF!
Excited for @WeAreDevs, @OpenAI Dev Day, and seeing my favorite people 💙
I made an always-on agentic on-call engineer 💚
It’s an AI agent that keeps working after you cut off its external network access.
It runs @NVIDIAAI Nemotron 3.5 Lightning locally on DGX Spark, investigates a broken service, changes the config, and verifies the repair.
Combination of open-weight models + local hardware is especially useful when you need control over both the model and the environment it runs in.
Open weights give you control over the model itself:
- which exact model version is running
- can you preserve that version
- can you customize it
- when you choose to upgrade or replace it
Local hardware gives you control over where the work happens:
- where your data is processed
- does it ever have to leave your environment
- does the system depend on an external model API
- can the application can keep working when external connectivity disappears
You don’t necessarily need both.
You might want open weights but still run them on rented GPUs.
You might need local inference for sensitive data without ever fine-tuning the model.
But when you need control over both the model and where it runs, the combination becomes very useful.
Watch the full video: youtu.be/_3MKLsli12w
So excited to be back in California next week! 💖
Keynoting We Are Developers San Jose (We Are Devs Berlin was amazing, by the way!) and attending OpenAI DevDay!
See you in San Jose and San Francisco 🌉
If you're there, definitely say hi 👋
I gave GPT-6 Astra a photo of one of my bookshelves
it turned it into a cozy interactive 3D library
with real covers
books I can pick up and rearrange
and vocabulary + takeaways that appear as I browse
who needs a tutorial when you have Astra
I gave Astra a link to the Reachy Mini assembly guide on @huggingface, and asked for a 3D app that shows where each piece goes
- it pulled 51 steps, diagrams, and video timestamps
- found 41 public 3D meshes + the assembly description in @pollenrobotics repo
- reconstructed the parts and fitted positions
- built step-by-step placement animations, like rotate, zoom, before/after, replay
- shipped an interactive web app
share your Astra experiments!
can you imagine two models understanding the same problem without exchanging a single word
congrats to the team @mostik_ai
#goals
meet @mostik_ai!
what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can.
everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning?
we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched.
how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance.
we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst
WIRED has the first external account of the company and the work: wired.com/story/russian-star…
full writeup, the setup, and all the numbers: mostik.ai/read-more
things are getting wild
I can now put a real AI supercomputer in my purse
when a new AI model drops, I don’t check if it’ll fit
I just download it and run it on 128GB of unified memory
and it’s quiet
we need to update our definition of "personal computer"
new video you've been waiting for about my @nvidia DGX Spark is up 💚
youtu.be/kAElYHzbKXg
if your girl:
- looks small but is insanely powerful
- makes other engineers jealous
- knows how to keep things private
that’s not your girl, that’s DGX Spark
a lot of engineers still hit their internal brakes when a problem feels difficult or ambiguous
brain goes straight into every technical risk and what-if, trying to solve the implementation before the problem is even clear
doesn’t matter as much anymore. just ask, is this even worth solving? if yes, the rest is figureoutable 🩵
fable is the only model that pushes back like crazy
I was working on something yesterday, dozens of iterations, and got things like - "I'm not going to evaluate another sentence" and "you said last one - hold yourself to it"
chatgpt would never
one quality I like in successful founders is that they are all to a big extent delusional
they treat their idea as inevitable
like not a single thing can wobble them off their path
what would make you change the programming language of your codebase in 2026?
does a programming language choice matter if agents are writing for you anyway?
Had a blast being a part of keynote talks at @ycombinator this weekend!
So many founders are going all in on building software factories without the slop and optimizing personal AGI workflows. I talked about how tech startups can make their products more likely to be chosen in agent-speed developer GTM⚡️
I’ve long been inspired by people like @paulg, and could feel the real no bs builder energy in every conversation.
Grateful for the discussions, and thanks to @ycombinator, @vaibcode, and @dexhorthy for the invite and the great event!
Excited!
See you at @ycombinator HQ! ❤️🔥
The last AI that works unconference brought together 70+ founders, systems engineers, and AI tinkerers to share the latest learnings and push the frontier together at @ycombinator's dogpatch HQ.
This weekend me and @vaibcode are running it back (signup below!) with a whole bunch of great engineers and plenty of room for audience-driven content - 8 hours of talks, breakout sessions, and some legendary Keynote Speakers
- @kwindla - @trydaily
- @swyx - @aiDotEngineer
- @lenadroid - @Akamai
- @bdougieYO - @papercompute
- @JohnKutay - @Rippling
- Leif from @SierraPlatform
- @codeshaunted - @boundaryML
- @0xblacklight - @humanlayer_dev
