Marius Slavescu retweeted
[1/2] In collaboration with @ContractedAI:
"Knowing, but not Noticing"
- We show that frontier ER models (which robot foundation models dominantly initialize from) consistently fail to know 'what to do next' in a chained task;
- how to measure that failure quantitatively using data sourced from a novel branching/ 'counterfactual'-style ego data collection regime; and
- propose a means to produce post-training data to fill in those gaps, based on what context ablations via harnessing reveal the model already knows, but simply fails to conjure reflexively on its own.
Link in thread 🧵
Marius Slavescu retweeted
The ability to try any new thing with a sensible approach is a cornerstone of artificial general intelligence.
Now the real world is yours for learning, NEO.
Marius Slavescu retweeted
mHC.cu performance oriented changes to enable larger expansion rates (n):
1. Template specialization for n=4/8/16/32. Kernels now dispatch to optimal unroll # instead of always using 32
2. Tensor Core path for n >= 32 via cuBLASLt - restructured to row-major layout so no transpose kernels needed
3. enabling pipelining for larger expansion rates (n) for the static H path. pipelining active at n >= 16.
Marius Slavescu retweeted
Just added the backward kernels
github.com/AndreSlavescu/mHC…
Marius Slavescu retweeted
Decided to implement Deepseek's new paper (mHC: Manifold-Constrained Hyper-Connections) in CUDA.
Marius Slavescu retweeted
Replying to @awnihannun
@awnihannun @andre_slav03 check this video, he was able to run the model on a 512GB mac + 128GB mac using:
Inferencer App: inferencer.com Kimi-K2-Thinking-Q4.25
m.youtube.com/watch?v=GydlPn…
Marius Slavescu retweeted
Update:
INT8 path now has fused perturbed matmul for better performance. MLP expand uses pure INT8 activations with no float conversion:
github.com/AndreSlavescu/egg…
per-step time on H100 PCI-e is at <60ms.
@j_foerst @bidiptas13 @d0rc
Marius Slavescu retweeted
EGGROLL 8xB200 distributed training with P2P weight averaging!
github.com/AndreSlavescu/egg…
Thank you @NorthflankWill and @northflank for the B200s.
Marius Slavescu retweeted
It's @ArtificialAnlys official: Parakeet TDT running on @TeamHathora has the best cost to WER (word error rate) ratio of any ASR model! This was all in less than a month since we launched
artificialanalysis.ai/speech…
Huge thanks to the @nvidia voice research team (Adi & folks) and the AA team (@grmcameron @kiriillbutler)
Try it for yourself here: models.hathora.dev/model/nvi…
Marius Slavescu retweeted
Tonight’s AI Lightning Talks at the @shinecapital office with talks from
Zack @AssemblyAI
@andre_slav03 @TeamHathora
And Kent from @Medal_TV and General Intuition.
We lit up the minds of NYC AI curious.
Come back in 2026!
Marius Slavescu retweeted
Super cool viz of how an Omni modal model can process audio tokens in/out by @andre_slav03
qwen3-omni-mechanics-gui-web…
Marius Slavescu retweeted
Here are our three papers on ways to use AI for teaching. All have prompts & details:
Using AI to make teaching easier & more impactful: papers.ssrn.com/sol3/papers.…
Using errors made by AI to teach hard things: papers.ssrn.com/sol3/papers.…
Assigning AI to students: papers.ssrn.com/sol3/papers.…
Marius Slavescu retweeted
Thank you @Autodromo_Monza for the completion certificate of the first race on a non-oval circuit during MIMO 2023 🏁🤖🏎️
#MilanoMonza #IndyAutonomousChallenge #NoSteeringWheelNoProblem
#AutonomousRacing #MIMO2023
Marius Slavescu retweeted
“More than 10,000 spectators filled the stands to witness the driverless racecars competing on the Monza “Temple of Speed” Formula One circuit as part of the Milan Monza Motor Show (MIMO) that ran last weekend at the @Autodromo_Monza” - @IAutonomousV
insideautonomousvehicles.com…
Marius Slavescu retweeted
We’re showing @AndrettiFE’s @JakeDennis19 around Indy today. During our first stop, he’s seeing the innovations in autonomous mobility happening right here in Indiana with @IndyAChallenge. – at Indianapolis, IN
Marius Slavescu retweeted
Congratulations to autonomous racing team MITT-PITT-RW who achieved a personal best time of 173.8 kph on their first @IndyAChallenge road race in Monza, Italy 🤖🏎️
The first challenge? Learning to turn right!
Full Story: bit.ly/3XoP2GG | #UWaterlooNews
Marius Slavescu retweeted
Congratulations to the winners and finalists of the Best Paper Award of
the 4th Workshop on Event-based Vision at #CVPR2023: Papers and code: tub-rip.github.io/eventvisio… #eventcamera With @guillermogb @KostasPenn @CFermuller @DMigliorePro
Marius Slavescu retweeted
Carve out a few hours to learn streamlit.io/
Powerful for rapid prototyping, interactive visualization.
It's a hammer and you'll start seeing a lot of nails.
Marius Slavescu retweeted
#ICYMI
⚡️EMN Nº 20 - “I love the Monza Challenge”
The @IndyAChallenge racing cars go outside the typicals American ovals circuit and @Autodromo_Monza is a new challenge for them. Brandon Dixon, Prof. of @UofAlabama, speaks about it.
©️ Marcelo Padin
📹youtu.be/hdjgkQsiieQ






























