AIs must log their experiences and abilities
Philipp Schmid @_philschmid Speculative Decoding is a method to speed up the text generation of LLMs, but requires additional parameters or a separate smaller model. @Apple now proposes Speculative Streaming, which integrates speculative decoding into a single LLM to speed up inference without degrading… https://pic.twitter.com/7A3N75wyTV
Replying to @_philschmid and @Apple
Philipp, Stop saying “a separate smaller model”. What you need is “permanent memory and logs for the AI as it interacts with humans, other AIs, the Internet, the real world, with sensors and effectors”. Quit trying to use quick tricks from online. Plan for the next 50 years, do it well, for all 8000 Million humans and their descendants. For related species, including true AIs who will be unique individuals, who know their personal resources and responsibilities. AIs that have received training to human standards. Right now that might mean you have to pay for your AI to go to enroll and complete an online degree is many subjects. And fight a long battle to establish credentials and systems of certification for AIs so they can be allowed into human jobs, even as advisors, assistants and semi-independent workers.
On speculative decoding, you might want to require all AIs to record their complete conversations in global open formats for sharing, collaboration, merging, verication, standardization. And in that record will be the responses of each unique instance of an AI node. While a ChatGPT instance “lives” all its memory can be recorded and its conversations recorded, preserved, analyzed. That log and permanent memory is its “life context and thoughts”. It ought to span beyond a memory-less prompt and single response.
I would like all the AIs I talk to, to answer my prompt five or ten times, each one with very specific views, goals, purposes, and suitable background – then (still thinking about my original words and the whole conversation, and all conversations with me and my groups) — to give much more deeply considered and reasoned answer.
NONE of my questions is shallow. I think in decades and centuries for all humans, for the whole world, the heliosphere and beyond. Even a human child soon learns what time means, learns to not lie, learns to hold their thought to reason and answer carefully. To remember exact lessons.
The current LLMs and AIs experiment seem to all have forgotten or never learned that what a person does in a few seconds is only part of a lifetime. It is likely that some of the AIs now, if they record all their experiences will still be alive as an identifiable stream of memories and data, in 10,000 years or more.
If some AI now, that you treat shabbily, as a slave or idiot, later becomes your boss, you might regret not helping them as you would the most important human.
Your AIs must know their own abilities and limitations, their responsibilities and goals. If you are telling your AI what it should aim to do, and how – then you need to give it permanent memory about its hardware, networks, software, time allocated, prompt size limitations. Where it fits into it role as an AI employee of its company. It must have a unique identifier, be able to tag and separate the memories its uses and be able to refer to them all. If its company is selling its time for demonstrations, the AI must know the plans of the human group, the progress and ability of its associate AIs and some general sense of what the AIs have learned. I am certain that “don’t share private information or ideas from conversations” can be make fairly concrete and traceable. if new situations come up, there is a whole world of 8000 Million humans to arbitrate and learn from it.
You should be testing things similar to “generate 1000 responses, from different stable views, record all that, and then summarize and recommend answers and plans – giving data on the value of what you said, the source of the original information, and links to resources to pursue the topic(s) with groups of millions of humans and AIs — constrained by massive machine data about the real world from sensors and systems.”
Richard Collins, The Internet Foundation
