Online Self-Distillation and Dreaming Could Solve AI's Forgetting Problem
Google's new multi-token prediction architecture boosts on-device inference speed by over 50% while freeing up 130MB of RAM on Pixel 9 and 10 devices.
AI
Most Recent
Online Self-Distillation and Dreaming Could Solve LLM Lifelong Learning
Google's new multi-token prediction architecture boosts on-device inference speed by over 50% while freeing up 130MB of RAM on Pixel 9 and 10 devices.
OpenRouter: Open-source AI models are now just 3–6 months behind frontier
Google's new multi-token prediction architecture boosts on-device inference speed by over 50% while freeing up 130MB of RAM on Pixel 9 and 10 devices.
SemiAnalysis: AI tokens cost $0.99/million and account for 30% of salaries
Google's new multi-token prediction architecture boosts on-device inference speed by over 50% while freeing up 130MB of RAM on Pixel 9 and 10 devices.
GLM 5.2 Climbs to Second in Long
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.
DeepSeek and Gemini Get First-Author Bylines in Chinese Journal
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.
Open-source Ornith 397B model beats Claude Opus 4.7 in coding benchmarks
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.
Italian AI Startup Domyn Takes on US and China with Open-Source 400B Model
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.
Google Halts Showcase Payments to Coerce Media into AI Training Deals
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.
Six in ten AI bug fixes are just copy-pasted from the web, Cursor finds
Andon Labs' Vending-Bench 2 benchmark pits AI models against a year-long vending machine business simulation. Open-source GLM 5.2 rises to second, while Kimi and Minimax show divergent results.