https://arxiv.org/abs/2308.08998
@aibreakfast
-
DeepMind’s Reinforced Self-Training removes human from RLHF
By
–
DeepMind showcases iterative self-improvement for Natural Language Generation. They basically took the "human" out of Reinforcement Learning from Human Feedback (RLHF) cycle and are calling it Reinforced Self-Training (ReST) [link to paper in ALT text]
-
Real-time AI phone calls are almost ready, with a few tweaks needed.
By
–
Still some tweaking to be done, but real-time AI phone calls are getting good pic.twitter.com/nmfsAw231m
— AI Breakfast (@AiBreakfast) 20 août 2023Still some tweaking to be done, but real-time AI phone calls are getting good
-
Google DeepMind’s Gemini to surpass GPT-4 with AlphaGo-LLM combination
By
–
Google DeepMind’s CEO @demishassabis says their new AI model "Gemini" will far surpass the capabilities GPT-4 Gemini (still several months from release) is a multi-modal AI system that combines their Go-winning program, AlphaGo, with LLM capabilities. Google acquired
-
AI surpasses humans in tasks, Kurzweil’s 2045 prediction underestimated
By
–
AI has now surpassed humans at a number of tasks, and the rate at which humans are being surpassed at new tasks is increasing. Kurzweil may have underestimated when he predicted 2045. Data from @ContextualAI and graphic from @TIME
-
AGI likely to have fewer parameters than GPT-4
By
–
AGI will likely have fewer parameters than GPT-4.
-
AI answers biased by left-leaning online text training data
By
–
The AI’s answers are purely a result of the training data it ingests (and tech companies aren’t specific about what was precisely included in the training corpus) You could argue that a preponderance of online text is slightly left-leaning, as those who are compelled to write
-
OpenAI’s secret project to uncover covert AI breakthroughs
By
–
Hmm… OpenAI has a special project where they try to find covert AI breakthroughs people use in secret
