We were able to reproduce the strong findings of the HRM paper on ARC-AGI-1. Further, we ran a series of ablation experiments to get to the bottom of what's behind it. Key findings: 1. The HRM model architecture itself (the centerpiece of the paper) is not an important factor.
RESEARCH
-
SuperNinja AI Research Platform Launches on Cerebras Infrastructure
By
–
SuperNinja is launching today:
Try SuperNinja: https://
super.myninja.ai
Read the research blog: https://
ninjatech.ai/blog/superninj
a-cerebras-worlds-fastest-deep-research
…
Build on Cerebras Inference: http://
inference.cerebras.ai -

SuperNinja Achieves 5x Speedup on Cerebras Hardware
By
–
Deploying SuperNinja deep research on Cerebras resulted in an instant 5x speedup. While leading closed-source research agents took minutes to complete a task, SuperNinja finished in 57 seconds.
-
NinjaTech Achieves One-Minute Frontier Research with Cerebras
By
–
In June, the NinjaTech team reached out to Cerebras with the ambitious goal of cutting deep research time down to 1min at frontier quality. We reached the speed, but the models at the time weren’t quite smart enough vs. o3.
-
SuperNinja: 5x Faster Deep Research AI Powered by Cerebras
By
–
Introducing SuperNinja – the world’s fastest deep research by @NinjaTechAI powered by Cerebras:
— Cerebras (@cerebras) 15 août 2025
‣ 1-2min for full answer instead of 10min
‣ 5x faster than o3, Gemini Pro etc.
‣ Same accuracy as measured by GAIA pic.twitter.com/rXDoamiksmIntroducing SuperNinja – the world’s fastest deep research by @NinjaTechAI powered by Cerebras:
‣ 1-2min for full answer instead of 10min
‣ 5x faster than o3, Gemini Pro etc.
‣ Same accuracy as measured by GAIA -

Character-Level Tokenization Success in DNA Language Models
By
–
watch me realize in realtime that @gdb is deadly serious about character-level tokenization and it's already working well for @pdhsu, @SKonermann, and @patrickc's @ArcInstitute 's 40B model that is the "GPT-1 of DNA"
— swyx 🐣 (@swyx) 15 août 2025
with biological language models (instead of natural language),… https://t.co/SMA81JYX0b pic.twitter.com/XXsMmDEZWewatch me realize in realtime that @gdb is deadly serious about character-level tokenization and it's already working well for @pdhsu
, @SKonermann
, and @patrickc
's @ArcInstitute 's 40B model that is the "GPT-1 of DNA" with biological language models (instead of natural language), -
AI System Fails Basic Problem Solving Tasks Consistently
By
–
Je souscris : l’erreur devient la norme ! Il est incapable de résoudre un problème de Maternelle et il est bête comme un foin. On recule, on n’avance pas avec cette version. Il faut tout lui expliquer comme à un enfant de 5 ans.
-
Denny Zhou’s Pioneering Contributions to Modern LLM Reasoning
By
–
Many aspects of modern LLM reasoning were invented by @denny_zhou and his team. Just putting it out there because many people do not know what goes on in frontier labs.
-
Warmth and Factuality: An Inverse Trade-off
By
–
Optimizing for increased warmth produces decreased factuality.
-

CulturalGround Open-Source Multilingual Cultural VQA Dataset Released
By
–
The code used for curating CulturalGround, the largest open-source multilingual cultural VQA dataset, is now publicly available. The proposed pipeline could greatly enhance factual understanding of a wide array of real-world entities in multimodal LLMs, at zero cost. 30M VQA