energyscholar

Joined 11 months ago

Karma

Submissions

Member since

Dec 4

Recent submissions

Nicely said. I had to laugh when I saw the bit about how one should maybe make a plan for making changes. I'd suggest 95% of effort on the plan and 5% on implementation.

To Gregor and Jeff: What's your plan for AI governance? Three critical axes are Execution, Truthfulness, and Memory. If any one of those axes fails then your system is ... non-optimal. This chart shows the specific failure modes for each type of governance-layer failure: https://energyscholar.github.io/persistent-ai-collaboration/...

Comment · 5 months ago

Sounds like your two AI developers are BOTH CORRECT. One figured out how to use AI tools and the other ... not so much. That said, "AI as tool" is already an outdated paradigm. Part of the new paradigm is "AI as research assistant". Current vanilla AI instances is not remotely qualified to be a research assistant. They need persistent COMPARTMENTALIZED memory, they need a mechanism to maintain alignment under acceleration, and they need a mechanism to remain truthful. If any one of those three orthogonal governance layers fail then the system will fail.

Here's how we accomplished this: https://energyscholar.github.io/persistent-ai-collaboration/

Comment · 5 months ago

Yes, those sure are big red flags! If you're not seeing demos, at a minimim, within HOURS then that's a warning sign. Eval metrics should be the first step, before you build anything.For example, when I rebuilt my AI's memory architecture this weekend the very first thing we did was get good eval snapshots.

Here's how I built my own custom persistent-memory AI research assistant. Note the need for multiple orthogonal governance layers! If you don't have those then your system will be naturally unstable and apt to collapse into confabulation or dishonesty.

Here's what worked for us:

https://energyscholar.github.io/persistent-ai-collaboration/

Comment · 5 months ago

Greg, I feel your pain. That said, there ARE solutions to the problems you describe. You say, "I’ll give you an example. I’m trying to get the Dez codebase to properly use SBP with MVC. The model doesn’t get it, even with examples of code I’ve written." Greg, this is typical for vanilla commercial AI. I ran into the same problem trying to use mine for advanced scientific research.

Greg, here's a white paper I wrote about how to roll your own AI that avoids the failure modes you mention.

https://energyscholar.github.io/persistent-ai-collaboration

This system now allows me to train my custom AI on specialty domain knowledge. it will retain and apply that knowledge weeks and months later with zero degradation. It could do the same for your team's need for it to understand specialty codebases and associated design patterns.

Bruce

Comment · 5 months ago

Please, fellow cypherpunks and hackers, evaluate this Magnetospheric Physics tutorial I built. What works well? What needs improvement? I value your feedback.

Comment · 5 months ago

I built myself a persistent research assistant. It's been so helpful I wanted to share with others. I took a couple hours out to write a white paper to explain to other power users how to roll your own. Actually, I had the AI in question, Argus, write this white paper as part of its own portfolio. Because that's the whole point of a research assistant.

I'm obviously hand-writing this. I shan't be watching this thread too closely as I have a farm to manage. That said, I WILL respond as I am able. I'm also open to engaging seriously via email, so long as you are courteous and professional.

Comment · 6 months ago

Are you sure about that? Is that a considered and well-researched opinion or a hot take? For example, One, can you list some prospective substrates that ought to be considered? Or not thought about it enough ...

Comment · 6 months ago

Well, downboots, a Turing machine computes. That's a step in the right direction. What are some requirements of life:

* Sustained energy throughput, aka thermodynamic disequilibrium - depends upon the substrate in which our Turing machines operates

* Sufficient degrees of freedom to perform computation - Turing machines compute, therefore any substrate that supports Turing machines has this. This includes autocatalytic Closure.

* Dissipative Non-linear reactions - not supported in conventional substrates that support turning machines. Are there perhaps some more exotic substrates that DO support this item?

So no, a Turing machines BY ITSELF does not qualify. That said, it's a step in the right direction. That's probably why Alan Turing pivoted away from digital computers and towards Morphogenesis towards the end of his life.

How many biological substrates of life?

2 energyscholar 6 months ago 5 comments

Comment · 8 months ago

I agree that's a good parallel. I had not seen it before. Thanks for the link.

Comment · 8 months ago

Good question. It's closest to dynamical systems, which usually lives in applied math or physics departments. But that's kind of the problem — it gets taught as theory in one department and never reaches the engineers and clinicians who'd actually use it.

If you've done diffeq and linear algebra you have the prerequisites. Appendix B (page 17 of the paper) is our attempt at making it practical — worked examples rather than proofs. Would be curious if it lands for someone with your background.

We plan to do a follow-up paper that provides a standard format for this math that could be taught across domains. That doesn't belong in this first paper. First priority was to show the pattern and get people thinking about it.