Learning to talk to AI requires deep curiosity and metacognition. This podcast will talk to people who talk to AI, and see what we can learn from each other.
Joined February 2023
- Tweets269
- Following617
- Followers49
- Likes295
Also sounds like an intersection between declarative and procedural? How do those interact at the complies/ur-programming level?
That ties into my own: voice dictation into ChatGPT while walking my dog. I thought I couldn't wrangle my attention enough to write, now I've got thousands of words and at least a dozen essays on the way. Turns out walking him is just enough stimulation to unlock my brain.
An LLM is not an 'auto-complete bot', but rather a construct of structures and biases in hyper-dimensional knowledge space, and as such *is* capable of making novel connections and inferences not explicit in its training data.
Interesting in the light of recent news about Apple offering publishers eight-figures to use their content to train its models. A strong ruling in this case could trip up the competition in time for them to step up and say, "Ours is based on the good stuff *and* it's paid for."
Its illuminating to watch open AI try to finesse it's compute load in real time. Dall-e in ChatGPT used to produce four images à la midjourney, but they regulated it to one image per query. Right now the app is overly eager to switch me into 3.5, which I'm sure is much cheaper 🤷🏼♂️
That's a fundamental theme in the Confession of St. Augustine, too, which has a lot to say about the dynamics of voluntary conversion. Neither wanting to believe nor acting out those beliefs will reliably trigger the inner change, but if that's all you've got...
Right up there with facing opposition while knowing you have ample convincing evidence and the imminent opportunity to present it
AI alignment is not just a model feature but an ongoing process that starts with the maker and extends to those deploying models in the wild. Communicating clear intentions and value structures is imperative, and to do otherwise is negligent. There is no reasonable alternative.
We require lower dimensional abstractions in order to function, but limiting ourselves to a *single* dimension requires idiocy and/or malice.
AGI's hold my beer moment: ChatGPT now remembers all your chats, knows you from them, and is agentic for $20/mo
Perhaps because we don't have the same knowledge-analogous structures in
Vision or spatial data that we have in language data. Even early multimodal models tend to translate content into semantic knowledge space in order to make sense of them, rather than in native visual space
This chat is pure gold. Happy Birthday to me!
Here's my conversation with @leecronin, a chemist and co-author of a controversial new Nature paper on the origin and evolution of life and complexity.
First 3 hours are here on X. 3 hours is the current upload limit. The rest is up on YouTube, Spotify, and everywhere else.
Timestamps:
0:00 - Introduction
1:15 - Assembly theory paper
21:45 - Assembly equation
34:57 - Discovering alien life
53:16 - Evolution of life on Earth
1:01:12 - Response to criticism
1:18:50 - Kolmogorov complexity
1:30:40 - Nature review process
1:51:34 - Time and free will
1:57:59 - Communication with aliens
2:19:57 - Cellular automata
2:24:26 - AGI
2:41:15 - Nuclear weapons
2:47:00 - Chem Machina
2:59:54 - GPT for electron density
3:09:24 - God
The Best of All Possible Worlds(TM): this one special trick to fix the one thing between you and happiness/prosperity. Optimization porn, maybe?
Every once in a while, it occurs to me that other people probably *aren't* talking to ChatGPT about the internal vector knowledge space of large language models and fundamental principles of information and knowledge theory.