The modern MoE architecture is insane: > Mixtral 8x7B: 47B total params, only 13B active per token
> DeepSeek-V3: 671B params, 37B active – beats GPT-4 at 1/10th cost
> Grok-1: 314B params, trained faster than any dense model of similar quality Pattern: 5-10x more parameters.
MACHINE LEARNING
-
MoE Architecture: 5-10x More Parameters
By
–
-
FSD now safer and smoother than human driving
By
–
And FSD now is way safer than letting a human drive. Certainly is way smoother.
-
Tesla asserts its self-driving is earlier than any 2008 or 2018 car
By
–
Show me a 2008 car that can drive itself. Even a 2018. My Tesla does. Everywhere. No one else was earlier.
-
AI Learning Roadmap for 2026: Your Path to Success
By
–
What will be your learning roadmap to Learn AI in 2026?
→ View original post on X — @krishnaik06, 2026-01-03 06:02 UTC
-
SpaceTimePilot: Generative Rendering for Dynamic Scenes
By
–
SpaceTimePilot
— AK (@_akhaliq) 3 janvier 2026
Generative Rendering of Dynamic Scenes Across Space and Time pic.twitter.com/d4pHH1i53JSpaceTimePilot Generative Rendering of Dynamic Scenes Across Space and Time
-
PyTorch’s Signed 64-bit Index Standardization Decision
By
–
Pytorch made the right call standardizing on signed 64 bit indexes. I would probably still be rather pointlessly making case by case decisions to use int32 if it were an option. Some old habits linger.
-
LoRA Efficiency for Specific AI Tasks and Datasets
By
–
No not the opposite – they just looked at a particular subsets of tasks, datasets, and metrics. There are some things that LoRA is sufficient for.
-
TMLR Celebrates Peer Review Community for AI Research
By
–
Finally, a big thank you to our advisory board, our first managing editor @fpedregosa
, the 1000s of reviewers and 100s of action editors for TMLR, who share our belief in scientific peer review for AI research. Witnessing so many people work with us has been truly heartwarming. -
Conferences Partner on Journal-to-Conference Track Model
By
–
We partnered with many conferences (AutoML, CoLLAs, RLC, MLRC, LoG) and now also partner with the NeurIPS/ICML/ICLR journal-to-conference track. I personally believe separating publishing from conference presentation is promising for the future of scientific dissemination in AI.
-
TMLR Scales Publishing Innovation in Machine Learning Research
By
–
TMLR went from publishing less than 200 papers in 2022 to over 1300 in 2025. And along the way we've experimented with various innovations, such as certifications, the Expert Reviewer status and the infinite conference: https://
tmlr.infinite-conf.org/index.html