This is interesting: Anthropic's models suffer more from "speaking" other languages in terms of token usage. In Spanish, the number of tokens used multiplies by 1.62x compared to OpenAI's model in English.
RESEARCH
-
Can LLMs trained on 1930s text learn to code?
By
–
Codex is doing AI research for me on Talkie using its in-app browser research question: can an LLM trained only on 1930s text learn to code?
-
AI Foundation Model Detects Multiple Diseases from Retinal Photos
By
–
What superhuman vision can detect from the retinal photo, which human eyes cannot, is stunning. A new foundation AI model screening for diabetes hypertension, hyperlipidemia, gout, osteoporosis, and thyroid disease @NatureMedicine https://
nature.com/articles/s4159
1-026-04359-w
… -
AI Scales Human Ingenuity and Decision-Making Capabilities
By
–
SAS CTO Bryan Harris: “From the very beginning, human ingenuity has been the engine behind every breakthrough that changed the world… And now AI will scale human observation and decision-making. There’s no question that this is a pivotal moment.” #SASInnovate
-
India Advances Global AI Collaboration with UN Leadership
By
–
IndiaAI engaged with H.E. Annalena Baerbock, President of the UN General Assembly, to advance global collaboration on inclusive and responsible AI. Shri S. Krishnan, @SecretaryMEITY
, shared India’s approach towards building scalable, safe and sovereign AI under the IndiaAI -
Testing Time LLM from Industry Legends
By
–
Test of Time LLM from some legends…pretty fun! https://t.co/ZsH7oLkUSM
— Soumith Chintala (@soumithchintala) 28 avril 2026Test of Time LLM from some legends…pretty fun!
-
Microsoft Releases TRELLIS.2: Open-Source Image-to-3D Model
By
–
Microsoft Presents "TRELLIS.2": An Open-Source, 4B-Parameter, Image-to-3D Model producing up to 1536³ PBR textured assets.
— Chubby♨️ (@kimmonismus) 28 avril 2026
Built On Native 3D VAES With 16× Spatial compression, delivering efficient, scalable, high-fidelity asset generation.
Ngl, pretty cool! pic.twitter.com/p2sGxFm616Microsoft Presents "TRELLIS.2": An Open-Source, 4B-Parameter, Image-to-3D Model producing up to 1536³ PBR textured assets. Built On Native 3D VAES With 16× Spatial compression, delivering efficient, scalable, high-fidelity asset generation. Ngl, pretty cool!
-
Transformer Architecture Blind Spot Fixed After 10 Years
By
–
The Transformer's biggest blind spot just got fixed after 10 years. Deep models have a silent problem. Signal from early layers gets buried as it climbs the stack. By the top, the original information is nearly gone. Past attempts blended layer outputs with smarter weights.
-
Efficiency-Driven Choices Compress Optionality With Strategic Costs
By
–
Your parallel is relevant.
Efficiency-driven choices tend to compress optionality over time, and rebuilding it later comes at a higher cost, but with a very different strategic value. -
LLM Safety Bias: Why AI Refuses Harmless Requests
By
–
Why do AI assistants suddenly say "I can't help with that" to perfectly harmless requests? Researchers from Hefei University of Technology and iFLYTEK Research have identified the root cause: LLMs develop a cognitive bias in their internal safety encoding, where even innocent