New white paper: Scholars from @StanfordHAI
, @Asia_Foundation
, and @UPTuks map the current landscape of technical approaches to developing LLMs that better perform for and represent low-resource languages. (1/4) https://
hai.stanford.edu/policy/mind-th
e-language-gap-mapping-the-challenges-of-llm-development-in-low-resource-language-contexts
…
MACHINE LEARNING
-
Mapping LLM Development Challenges in Low-Resource Languages
By
–
-
Multilingual AI Models: Trade-offs and Data Quality Solutions
By
–
The paper discusses the trade-offs of three approaches (massively multilingual models, regional multilingual models, and monolingual or monocultural models) and highlights ongoing initiatives to address underlying data scarcity and quality issues. (3/4)
-
AI-GA Meta-Evolution: Algorithms That Invent Algorithms
By
–
[ Algorithms That Invent Algorithms ] AI‑GA Meta‑Evolution Demo (v2): https://
github.com/MontrealAI/AGI
-Alpha-Agent-v0/tree/main/alpha_factory_v1/demos/aiga_meta_evolution
… #AGI #AGIALPHA #MetaLearning -
Open-Source AI Agents and LLM Apps with Step-by-Step Tutorials
By
–
I build open-source AI agents and LLM apps and share them through detailed step-by-step instructions. If you're interested in ML, LLMs, RAG, and AI agents and want to receive apps and tutorials every week, subscribe to AI Engineering for free:
-
Stanford FramePack Generates 1-Minute Video from Single Image on 6GB VRAM
By
–
Stanford a dévoilé FramePack.
— VISION IA (@vision_ia) 23 avril 2025
Cette IA peut tourner sur un GPU de pc portable avec seulement 6 Go de VRAM pour générer une vidéo d'une minute en 30 fps à partir d'une seule image.
Sans distillation, open source.
1 exemple fou : 👇 pic.twitter.com/XxK9OnfiwPStanford a dévoilé FramePack. Cette IA peut tourner sur un GPU de pc portable avec seulement 6 Go de VRAM pour générer une vidéo d'une minute en 30 fps à partir d'une seule image. Sans distillation, open source. 1 exemple fou :
-
Text embeddings rediscover BM25 core principles during training
By
–
-> Researchers show that text embeddings methods re-discover BM25 from scratch during training! Far from abandoning standard search techniques like BM25, it appears that neural search methods naturally re-implement their core principles. Let's you have a search query like
-
Veo 2 outperforms Sora with smoother motion and natural lighting
By
–
Bonus tip: Veo 2 has a better understanding of real-world motion than most models. In side-by-side tests with OpenAI’s Sora, it felt more real. Smooth camera pans. Natural lighting. No weird glitches.
-
Google AI Studio video generation playground for developers and anyone
By
–
Go to: http://
aistudio.google.com/generate-video This is Google’s playground for testing new AI models. It’s meant for developers but anyone can use it. You just need a free Google account. -
One-Step RL with Policy Gradients and Majority Voting
By
–
From a quick look, this is 1-step RL using policy gradients to regress towards the majority voter using 0-1 loss.