Скачать или смотреть How do Advanced AI Agents learn?

AI Safety Episode 20: How do Advanced Autonomous AI agents (AAAIs) learn?
In the collective intelligence approach to Artificial General Intelligence (AGI), Dr. Craig A. Kaplan first
distinguishes between learning at the Network level where many individual human and AI
agents cooperate, and learning at the individual AI Agent (aka AAAI) level. Then he reviews
some of the common approaches for training individual AI agents, such as the Large Language
Models (LLMs) produced by OpenAI (GPT-4) and Google (PaLM-2). LLMs are first trained, then
tuned, and finally prompted to respond in desired ways.

Training is what companies like OpenAI and Google do when they first produce their LLMs. It
typically requires huge amounts of data from the internet, significant time, and millions of dollars
in computing and data center expense.

Tuning uses smaller domain-specific datasets to give a trained LLM expertise in specific
areas or adjust its behavior. Models can also be tuned by having the models interact with
humans or other AIs. Reinforcement Learning with Human Feedback (RLHF) is a type of tuning
whereby typically many humans provide feedback to an LLM (e.g. telling it: “That’s good” or “That’s
bad”) in order to shape the LLM’s behavior.

Constitutional AI is a tuning approach whereby researchers first train one AI on what is right or
wrong based on a “constitution” written by the researchers. Then that trained AI teaches other
AIs, thereby taking humans out of the loop and increasing efficiency. Although this approach
scales well, Kaplan believes it can be potentially dangerous.

Prompting is what the end-users of the tuned LLMs do when they interact with the trained and
tuned LLM. An example of a prompt might be telling the LLM to: “behave and answer the
following questions as a lawyer might do.”

Included are clips of Yenjin Choi (Professor, University of Washington), Jensen Huang (CEO,
NVIDIA), Stuart Russell (Professor, UC Berkeley), Satya Nadella (CEO, Microsoft) and others.

[email protected]
Superintelligence.com https://www.superintelligence.com/ - Patents on developing safe AGI and SI
iQ Studios https://www.iqstudios.net/ - Video production
iQ Company https://www.iqco.com/ – AGI consulting
LinkedIn -   / craigakaplan  
X - https://x.com/Aligned_SI

Video Source Quotes:

S3 Ep 10 Yejin Choi: Teaching AI Common Sense and Morality (Host: Pieter Abbeel):    • S3 Ep 10 MacArthur Genius Prof Yejin Choi:...  

COMPUTEX 2023 Keynote NVIDIA Keynote: Jensen Huang:    • COMPUTEX 2023 Keynote NVIDIA Keynote  

Beyond ChatGPT: Stuart Russell on the Risks and Rewards of A.I.:    • Beyond ChatGPT: Stuart Russell on the Risk...  

Special Edition with Andrew Ross Sorkin: AI and Microsoft CEO Satya Nadella:    • Special Edition with Andrew Ross Sorkin: A...  

#AI #artificialintelligence #AGI #artificialgeneralintelligence #aiagents #deeplearning
#machinelearning #neuralnetworks #AAAI #tuningai #tuningllm #google #jensenhuang #nvidia #satyanadella #microsoft #openai #customai #aiagent #aiagents #ucberkeley
#universityofwashington #craigakaplan #iqcompany #iqstudios #llm #superintelligence

This video was edited by Kris Hawes, https://krishawes.com/.