here's some free alpha: if we do RL for too long after pretraining, we will surely overwrite parameters and start to forget things in the original instructGPT paper, their best model mixed RLHF with pretraining gradients to avoid exactly this model drift issue yet no one is
RESEARCH
-
SIGGRAPH 2025 Coverage: Creative AI and Visual Computing
By
–
Looking forward to your coverage of #SIGGRAPH2025!
-

Efficient Differentially Private LLM Fine-Tuning via Reinforcement Learning
By
–
Efficient Differentially Private Fine-Tuning of LLMs via Reinforcement Learning
-
Reasoning Model Generalization: Math to Creative Writing Transfer
By
–
i'm looking for good examples of reasoning model generalization for example, a model incentivized via RL to think for a while and solve math problems gets better at creative writing is this common?
-
Manus AI launches Wide Research for parallel processing
By
–
Manus AI introduced Wide Research 🤖
— 🚨 AI News | TestingCatalog (@testingcatalog) 31 juillet 2025
"Wide Research is a system-level mechanism for parallel processing, and a protocol for agent-to-agent collaboration." https://t.co/z6DH8z5sFR pic.twitter.com/bjrTgCThsLManus AI introduced Wide Research "Wide Research is a system-level mechanism for parallel processing, and a protocol for agent-to-agent collaboration."
-
China’s AI Momentum: Path to Surpass US Leadership
By
–
There is now a path for China to surpass the U.S. in AI. Even though the U.S. is still ahead, China has tremendous momentum with its vibrant open-weights model ecosystem and aggressive moves in semiconductor design and manufacturing. In the startup world, we know momentum
-
FLUX.1 Krea [dev] : SOTA open-weights model for photorealism
By
–
Black Forest Labs released FLUX.1 Krea [dev] – a new SOTA open-weights FLUX model, tailored for photorealism! https://t.co/vf6d1EH3dq pic.twitter.com/RSyC7BVNgV
— 🚨 AI News | TestingCatalog (@testingcatalog) 31 juillet 2025Black Forest Labs released FLUX.1 Krea [dev] – a new SOTA open-weights FLUX model, tailored for photorealism!
-
Three New AI Models Released: Qwen, Cohere, Mistral
By
–
Three new models this morning (Qwen, Cohere, Mistral) and our power is out so I'm a bit behind now!
-
AI Content Labels Impact on Message Persuasiveness Study
By
–
New policy brief: Does labeling AI content affect its persuasiveness? Scholars evaluate the impact of authorship labels on how people perceive AI-written policy messages. https://
hai.stanford.edu/policy/labelin
g-ai-generated-content-may-not-change-its-persuasiveness
… -
Microsoft Study on AI Models References Original Research
By
–
Microsoft study: https://
arxiv.org/abs/2507.07935 Original: https://
arxiv.org/abs/2303.10130
