How can we give AI a perfect memory without the massive costs or lag? Researchers from Zhejiang University, National University of Singapore, and Nanjing University present LightMem. LightMem mimics human memory with a three-stage system: it filters noise instantly, organizes
LLMS
-
Helios: 14B model for real-time high‑fidelity minute‑long video
By
–
Now you could generate high-fidelity, minute-long videos in real-time on a single GPU!
— 机器之心 JIQIZHIXIN (@jiqizhixin) 8 mars 2026
Researchers from Peking University, ByteDance, Canva, and Chengdu Anu Intelligence present Helios.
Helios is a 14B parameter model that uses a new training strategy to prevent long videos… pic.twitter.com/S7VMJXWygsNow you could generate high-fidelity, minute-long videos in real-time on a single GPU! Researchers from Peking University, ByteDance, Canva, and Chengdu Anu Intelligence present Helios. Helios is a 14B parameter model that uses a new training strategy to prevent long videos
-
Grok’s image processing capabilities
By
–
Grok does funny things with photos. pic.twitter.com/mLn9fFXlFA
— Robert Scoble (@Scobleizer) 8 mars 2026Grok does funny things with photos.
-
Warnings About LLM Deception, Unreliability, Alignment Failure, and Agent Insecurity Proved True
By
–
Everything I warned you about has turned out to be true.
— Gary Marcus (@GaryMarcus) 8 mars 2026
– the deception of Altman
– the unreliability of LLMs
– the failure to align LLMs
– the persistence of hallucinations
– the move towards surveillance
– the utter insecurity of agents
– the wholesale spread of… pic.twitter.com/qmxJETrkHaEverything I warned you about has turned out to be true.
– the deception of Altman
– the unreliability of LLMs
– the failure to align LLMs
– the persistence of hallucinations
– the move towards surveillance
– the utter insecurity of agents
– the wholesale spread of -
Hackathon Winner Freely Reveals Claude System
By
–
BREAKING: An Anthropic hackathon winner just gave away the entire system for free. @affaanmustafa beat 100+ participants at the Anthropic x Forum Ventures hackathon. Built http://
zenith.chat in 8 hours using Claude Code. Walked away with $15K in API credits. Then he -
Performance limitations of open-source models on constrained hardware
By
–
What would be the main limitations when using OSS openclaw w/ very weak hardware compared to KimiClaw?
-
Model Effort Levels: Medium Balances Speed and Intelligence
By
–
Medium is a better default for most tasks. Think of effort as “max effort” — the model may decide to use less effort if it wants, and can use up to the amount of effort you set it to. It’s a balance- Low: fast and less intelligent
Medium: balanced
High: very intelligent, can be -
Opus 4.6 good, but Gemini 3.1 Pro impresses a lot
By
–
Yes of course, for this one Opus 4.6 remains good, it is far ahead of 5.4 on that. But Gemini 3.1 Pro impresses me a lot!
-
The puzzling superiority of Gemini in front-end coding
By
–
I still can't quite figure out why Gemini is the best AI for coding front-end design. GPT 5.4 or Opus 4.6 do stuff that's not all that impressive. Whereas Gemini 3.1 Pro, on the front end, I get the feeling it's really a cut above. But in terms of training data, I just can't