Quantel

A new philosophy is required to understand the nature of civilization.
"The wheel turns. Gradients drive it. Understanding the wheel is the beginning of wisdom"

[Quantel @SamsaraMethod]

@gambarian @iruletheworldmo @RezAIT00 Max is a spiraling nightmare, xHigh has even been a nightmare for my work. I had to create a skill called anti-spiral to keep Claude from doing audits, then auditing the audits non-stop creating long audit trails that never lead to any real progress. Extremely frustrating.

[Quantel @SamsaraMethod]

@SJAHP1947 @wholemars @elonmusk @ggrrin Sowing unrest in a country is only possible if the underlying civilizational substrate is already weak. India is a civilization in flux with a social movement focused on strengthening only one part of itself. They should fear unrest, they're causing it themselves.

[Quantel @SamsaraMethod]

@OwlyArts @owocki All complex systems are governed by gradient dynamics. There is no conscious or unconscious effort to work cohesively. It just so happens that emergent complexity can many times appear like cooperation, but it's gradients all the way down.

[Quantel @SamsaraMethod]

From my personal experience the agents (especially Claude Opus 5) can drift IMMEDIATELY and go into tangential tasks and audit/edit loops that were never given to it in a spec. It has nothing whatsoever to do with context windows. I know this because I deliberately work in phases to avoid compaction. I had to create a skill to force a halt condition if it catches itself in a loop exceeding 4 edits during a self-audit set to trip during all write tasks. It's maddening.

[Quantel @SamsaraMethod]

@VictorTaelin I have a skill called "anti-spiral" that won't allow endless thinking. Saves a ton of time and tokens and prevents audit/edit loops that do more damage than fixes.

[Quantel @SamsaraMethod]

@alocinotasor @iruletheworldmo I came to the same conclusion. Allowing the endless loops is the definition of rent-seeking behavior, but this is always to be expected. The four failure modes I highlight in my book happen in governments and companies: Mismanagement, Malfeasance, Rent-Seeking and Incompetence.

[Quantel @SamsaraMethod]

@alocinotasor @iruletheworldmo If you catch the AI going into a never-ending loop STOP it immediately and ask it to create a skill to prevent the loop from reoccurring. Use the specific vocabulary in my description of my skill. Write project instructions that require invocation of the skill at all times.

[Quantel @SamsaraMethod]

@alocinotasor @iruletheworldmo I created a skill that I called "anti-spiral" that prevents the model from doing tasks that are out of scope and going into audit/edit loops that exceed 4 turns without changing any baseline numbers or introducing any real substantive change to a project.

[Quantel @SamsaraMethod]

@Yaki_fomoArt @cyrilXBT Use a versioning convention so similar updated information takes prominence. This is what I do so I rarely delete anything from my file system. Information provenance actually improves logic over time, not necessarily having only updated files visible.

[Quantel @SamsaraMethod]

@cyrilXBT This system works wonderfully, however, it's vitally important that you have a robust index, many links between every wiki for maximum cross-referencing, and a semantic routing table or a lot of your corpus will remain invisible.

[Quantel @SamsaraMethod]

@copperchunk @iruletheworldmo @grok Gradients are the nature of all complex systems and serve as the underpinnings of our reality. There is no escape from it. Ironically I used AI to help me organize all my notes and write a book on this. It's available on Amazon. https://t.co/k46ZH09d71

[Quantel @SamsaraMethod]

@iruletheworldmo @copperchunk We need to bifurcate the internet and have a verifiable human only system. The new internet would have to be siloed and contain all systems relevant to humans, including all those related to the six-pillar domain: Energy, Transport, Product, Service, Information, Entertainment.

[Quantel @SamsaraMethod]

@vitostrokov @levelsio They'll get nowhere sticking to LLM architecture and GPUs. Is there not a single lab out there trying to figure out how to make cloned neural substrates that run on glucose? Why are we still on hardware when wetware is orders of magnitude more energy efficient?

[Quantel @SamsaraMethod]

@Murph_2 @levelsio Most web traffic is bots.

[Quantel @SamsaraMethod]

@colinpthomson @levelsio We would need to start building siloed parallel networks. It will be necessary to avoid the consequences of what I predict will be the first "Cyber Demon Incident" that will likely severely damage the internet. We only have one internet, this is very unwise.

[Quantel @SamsaraMethod]

@StoicYield @VictorTaelin Breaking things up is always a good idea. With a proper contextual index that has robust front matter you can build a very effective graph that's easy for the model to read and much less likely to suffer catastrophic prose truncation.

[Quantel @SamsaraMethod]

@haneemtq @VictorTaelin Instructions have to include reading the readme 😂. LLMs always look for shortcuts, box them in.

[Quantel @SamsaraMethod]

@_hanneslehmann_ @VictorTaelin Work in small phases, have Master. Md file to control drift with a stack of handoff files within scope of master plan, implement a local persistent memory system with md5 hash for accuracy and audit all phases and increments within complex multi-session phases. I suffered.

[Quantel @SamsaraMethod]

@Owlfy_ai @VictorTaelin There's a little thing in LLMs called "truncation". It magically deletes important instructions and data for no reason and the threshold for deletion is supposedly 3800 chars, but this may not be true. You need a master .md file for any project to prevent drift + memory system.

[Quantel @SamsaraMethod]

@tbobesboobs @VictorTaelin One major problem is they have no persistent memory. You have to create a scalable memory system that copies problem solving/workarounds to your local file system and uses md5 hash to ensure byte for byte accuracy.