The continual learning and post-transformer architecture research lab behind the path to AGI.
Joined March 2026
- Tweets40
- Following0
- Followers235
- Likes5
Because our team does not sleep and has no life (a lot of supplements and caffeine), we decided to work on a mega kernel for ResonatorLM, inspired by the recent demo @flappyairplanes.
A 500M parameter model learning live while it plays DOOM, all on one persistent kernel.
Check out the full video here, all implemented on a single @nvidia RTX PRO 5000 Blackwell card.
A 4B parameter model with strong agentic and long context capabilities, running in your pocket, on an @Apple iPhone.
axioniclabs retweeted
Our Resonator model from @axioniclabs now works locally on a iPhone.
A 4B parameter model, that is running completely locally, and has agentic intelligence + long context capabilities.
A lot more to come soon!
A big shout out to the research team for getting this out!
Here is the official launch video! A shout out to @typesafeai and the creators of Anything-Jev for inspiring this work!
We built a system 1, Jev-like decision model in 2 days last week based off our own Resonator architecture. It is significantly faster and more efficient for specific use-cases. It works surprisingly well for specific classification and agentic use-cases:
Read the full post, and more what we actually do, here: axioniclabs.ai/research/deci…
We are continuing to scale up Resonator.
We have a few results that are dropping later this week, centered around our pretraining pipeline, and an updated 4B parameter model.
We presented ResonatorLM at ICANN this afternoon.
We think long-context language modeling and continual learning need a different architecture, and that is the problem ResonatorLM is working on.
If that is the kind of work you care about, our blog is at axioniclabs.ai/research/reso…
axioniclabs retweeted
Our account was temporarily compromised, We will NEVER launch a token!!!!
There is no Axionic Labs token or crypto affiliation. We will never launch a token. Any and all tokens launched under the Axionic name are fake and a scam. Our Twitter was temporarily hacked. Please be careful!
@axioniclabs research is being published at ICANN 2026. Our research on post-transformer alternatives will be highlighted in the main conference from 14–17 September at ICANN, held at University of Padua
ResonatorLM explores an alternative to self-attention for long-context sequence modelling, using learned resonators to build a more efficient recurrent architecture.
At @axioniclabs, we work on new model architectures and learning systems that challenge the assumptions behind today's AI stack.
ICANN brings together researchers working on the foundations of neural networks. That is the hardest audience this work could have, which is exactly why we want it in front of them.
Come talk to us if you're working on post-transformer architectures, continual learning or adaptive AI models.
we are in LA on 26 september.
HOSTLINE. founders, researchers, infra people, and the investors backing them.
saturday, 6:30pm, private estate on the west side. open bar, DJ, food.
list is by request. address comes with the approval.
dress code is all white and it is enforced.
partiful.com/e/gxhFNw2n4JKLn…
"Axionic Labs is not hiring."
That's what we've been replying to DMs all week ; a lousy thing to send someone who read our work.
So: the Resonance Prize. Read our research, tear it apart.
$500 for 1st
No ML background needed
Entries close Sept 21
Rules ↓
25 founders, a wood-panelled room on van ness, and one question: when does it make sense to train your own model?
founders dinner at harris' with @thundercompute came in with a thesis, left with a better question.
nyc next, sept 9 with @crossmint:
luma.com/wvghia9s
Excited to see our good friends @ThunderCompute raise their series A!
Today, we raised a $13M series A to solve the GPU shortage
Trillions are being spent on GPU CapEx while 80% of it sits idle
The capacity already exists
Virtualizing GPUs unlocks it
Backed by Matrix, YC, and CEAS
axioniclabs retweeted
Today, we raised a $13M series A to solve the GPU shortage
Trillions are being spent on GPU CapEx while 80% of it sits idle
The capacity already exists
Virtualizing GPUs unlocks it
Backed by Matrix, YC, and CEAS
Your AI model is too smart for its own good.
Bigger isn't better. It's just more expensive.
On Aug 20 in SF, we're hosting a founders dinner with @ThunderCompute on why task-specific models beat general-purpose ones on cost, speed, and accuracy.
Luma link soon!