Pinned Tweet
Did you know this wiggly thing can be a computer?
We think of computation as symbolic: add, subtract, multiply....
…but…you can also compute directly with voltage, current, charge, waves, materials (octopus arms…look it up)...or wiggles.
Today, we’re (@unconvAI ) sharing Un-0: our first step toward scaling this idea to modern AI workloads.
Michael Carbin retweeted
Today, we’re excited to launch the Datology Curation Studio, customizable frontier data curation for every team training its own models.
Data quality is the ultimate compute multiplier, and Curation Studio makes it easy for everyone.
Michael Carbin retweeted
At COLM in SF on Tuesday and Thursday! Would love to meet with people, old friends and new friends to be! DM me if you want to meet!
Also, come by our poster on "switching linear attention" Tuesday 4:30-6:30 pm, Grand Ballroom, Poster Location: #136
Michael Carbin retweeted
Fisler, Littman, & I have an upcoming CACM op-ed: "Teaching Novice Computing and Programming in the Agentic AI Era". Based on our experimental course+research. We argue that the «flaky compiler» is a useful metaphor for course design. Details in preprint:
cs.brown.edu/people/sk/Publi…
Michael Carbin retweeted
Retrieve more, rerank, get better results. That's what most of us expect from search, but it turns out to be wrong! 🤯
I'm SUPER EXCITED to publish the 141st episode of the Weaviate Podcast with Mathew Jacob (@mat_jacob1002)! Mathew led the work behind "Drowning in Documents" during his time at Databricks and is now a Ph.D. student at the University of Washington working on ML systems!
This episode dives deep into "Drowning in Documents". This has been one of the most influential papers for us @weaviate_io as we are exploring scaling reranked retrieval. I think it is a must read for those working in Search and Information Retrieval. 🙌
We begin with an overview of the paper, and then dive into full scoring with cross encoders and phantom hits. We then cover Listwise Rerankers, what next generation cross encoders might look like, and ranking cascades.
On the topic of ranking cascades, I loved learning about Mathew’s work with Melissa Pan (@melissapan), Negar Arabzadeh (@NegarEmpr) and collaborators on “Natural Language Query to Configuration for Retrieval Agents”. Per-query “effort” prediction is certainly going to be a huge component on the future of these search systems! (Congratulations to OpenRouter 😆)
We then dove into Mathew’s work on TraceLab, a super exciting effort to understand coding agents such as Claude Code and Codex. ⌨️
This was a super fun conversation, and I really hope you find it useful!
YouTube: youtu.be/fWuavBcoTzk
Spotify: open.spotify.com/episode/2A8…
Michael Carbin retweeted
I'm really excited to share this work! Sparsity has been the holy grail to gain hardware efficiency. Today we present results that show, not only does it work, it actually gains performance and enables near infinite scalability!
Brains🧠are extremely sparse both in connectivity and in activity. By framing AI inference as a dynamical system, we start to see similar sparse constructs show up to actually make the system work.
unconv.ai/blog/less-is-more-…
I can't quit sparsity! At @unconvAI , we've been investigating the interaction between sparsity and physics-inspired dynamics for AI.
We're still on the chase, but, like others, we're onto some interesting findings that sparsity unlocks model scaling (up to our contemporary understanding of how to train dense dynamical models).
We removed 93.25% of the connections in our Un-0 image model, fully expecting to pay for it in quality. But it got better.
FID 7.15 on ImageNet 64x64, roughly 1.9 ahead of the dense baseline at matched size. Same family of model, a fraction of the couplings, a better score.
Here is why that is not as strange as it sounds. Un-0 is a coupled oscillator model, and in the dense version every oscillator talks to every other one. That sounds like a strength, but it means the whole system can fall into catastrophic synchronization: everything locks into step, gradients go flat, and learning stalls.
Sparser connectivity leaves room for coherent and incoherent activity to coexist. The dynamics stay alive, and the model keeps learning.
Connectivity turns out to be a control knob, not a dial you turn up until it stops.
Learn more here: unconv.ai/blog/less-is-more-…
Michael Carbin retweeted
Wake up babe, new benchmark just dropped!
📣 Excited to introduce OfficeQA Pro V2, the next generation of OfficeQA Pro!
It's built on a new corpus of 120,000 PDFs provided by the U.S. Treasury.
Frontier AI agents average 26% accuracy. More info below 👇
databricks.com/blog/introduc…
There is a precedent for this: RAFT!
a best paper that authors argue is no technically better than previous work: just understandable.
hard part: understandable requires human eval, which they actually did:
"Results from a user study demonstrate that Raft is easier for students to learn than Paxos"
impressive because us cs folks usually -- sadly -- averse to human studies.
usenix.org/conference/atc14/…
Michael Carbin retweeted
I think there will come a time when we need to sharpen our latents. I.e., to learn discontinuities jointly with smooth structure and to automatically identify discrete structures and then iteratively refine the boundaries.
With his academic hat on, one of our teammates at unconventional, Ang Li, has a long-time collaboration on the hardware architecture for systems with physical dynamics. Take a look
What if your AI chip didn't calculate the answer, but physically settled into it?
That's the premise behind dynamical system hardware, and there's never been a standard way to program it. In an upcoming International Symposium on Computer Architecture (@ISCAConfOrg) 2026 paper, Unconventional AI MTS Prof. Ang Li and co-authors introduce DS-ISA: a minimalist 9-instruction architecture that bridges digital processors and continuous-time analog physics.
It's a proof-of-concept, but it's the abstraction layer the field needs to build real compilers and software stacks for dynamics-based AI hardware.
Full deep dive: unconv.ai/blog/speaking-the-…
Michael Carbin retweeted
Really cool work! Looking forward to seeing Kuramoto Oscillator models implemented on new hardware.
Thanks also for citing our work on Kuramoto Orientation Diffusion Models (arxiv.org/abs/2509.15328) ♥️
Michael Carbin retweeted
A drifting model using coupled oscillators 👀
It manages to produce half-decent ImageNet 64x64 samples, which may seem like a toy experiment nowadays, but was basically SoTA only about a decade ago.
Love the detailed blog post! Takes me back to my reservoir computing roots😁
🧠 Today we introduce Un-0 from @unconvAI : the first large-scale generative model build on physics as a compute primitive. This represents a “hello world” moment for physics-based models. We use the inherent time-varying behavior of physical systems to do compute for us. The result is a new way to build a computer that can be VASTLY more power efficient. 🧵
unconv.ai/blog/introducing-u…
Michael Carbin retweeted
Still working on this but
Here is a first look at my efforts to run an EBM on pulse (not phase) coupled oscillators.
This model was trained to reconstruct a glyph. It’s basically hello, world.
Not perfect but concept good!
Link to my original post in thread.
1/
Michael Carbin retweeted
super impressive for unconventional to develop the best coupled oscillator image generation model in 6 months!
if you want to give coupled oscillator computing a try for free using FPGAs, check out my open-source architecture, DIMPLE ⬇️
blog: unconv.ai/blog/introducing-u…
repo: train or extend it yourself. github.com/unconv-ai/Un-0
And the official @unconvAI post!
