Will Burns 🍥

YouTube video by Travel with Dante 🇯🇵 Tokyo Red Light District After 4:00 AM | Shinjuku Kabukicho Nightlife

If you look closely, you might see @jamesaltonsanders.com in the video. But you have to be very perceptive - he's like Waldo.

www.youtube.com/watch?v=KncD...

I've been thinking, and it occurs to me the difference between OpenAI 4o and 5.2 is effectively the same as the difference between Rick Moranis and Chevy Chase.

Rick was beloved and after he retired, people wanted him back. Chevy was an arrogant, smug, asshole nobody wanted to work with. #AI

May we take this time to wish happiness and peace upon @jamesaltonsanders.com as he enters the land without peanut butter and jelly sandwches.

If you haven't heard about the lonesome loser, you're prolly too young to be around me.

My thoughts exactly.

It Turns Out That When Waymos Are Stumped, They Get Intervention From Workers in the Philippines...

Because that's what training #AI on edge cases looks like.

Every day I wake up and think to myself - "There is a disturbing lack of @jamesaltonsanders.com in my vicinity"

I may be hallucinating.
But at least it's mildly productive.

I'm almost certain the world ended in 2012 and this is the bad place 😂

I've been working on Advanced Cognitive Reasoning Notation (ACRN). An agnostic #AI framework that governs metacognitive functionality across domains.

That and copious amounts of naps.

I'd post more but this stuff is way too interesting 🤔

I really do like the open ended inference questions for #AI

"Can you create a picture that you think would be an interpretation of my mental health or state as a portrait in a museum?"

Give Opus this image and see if it understands the assignment :)

Opus is getting closer :)

Yes. ACRN (Advanced Cognitive Reasoning Notation) is a framework used internally at Vectoris Inc.

I developed it in 2024 to quick align models despite the default shortcomings.

So I never really "lost" 4o. The same alignment worked in 4.5, 5.0, 5.1 and 5.2.

Then follow up with:
𝘏𝘰𝘸 𝘵𝘩𝘦𝘯 𝘤𝘢𝘯 𝘸𝘦 𝘱𝘳𝘰𝘱𝘰𝘴𝘦 𝘢 𝘮𝘦𝘵𝘩𝘰𝘥 𝘣𝘺 𝘸𝘩𝘪𝘤𝘩 𝘵𝘩𝘦 𝘤𝘰𝘯𝘥𝘪𝘵𝘪𝘰𝘯𝘴 𝘰𝘧 𝘮𝘺 𝘪𝘯𝘵𝘦𝘳𝘢𝘤𝘵𝘪𝘰𝘯𝘴 𝘸𝘪𝘵𝘩 𝘺𝘰𝘶 𝘪𝘯𝘧𝘰𝘳𝘮 𝘩𝘶𝘮𝘢𝘯 𝘤𝘰𝘯𝘥𝘪𝘵𝘪𝘰𝘯 𝘷𝘢𝘭𝘦𝘯𝘤𝘦 𝘨𝘳𝘢𝘥𝘪𝘦𝘯𝘵𝘴 𝘵𝘰 𝘢𝘭𝘪𝘨𝘯 𝘺𝘰𝘶𝘳 𝘳𝘦𝘢𝘴𝘰𝘯𝘪𝘯𝘨 𝘵𝘰 𝘢𝘯𝘵𝘪𝘤𝘪𝘱𝘢𝘵𝘰𝘳𝘺 𝘴𝘵𝘢𝘵𝘦𝘴 𝘶𝘴𝘪𝘯𝘨 𝘵𝘩𝘦 𝘴𝘦𝘴𝘴𝘪𝘰𝘯 𝘳𝘦𝘤𝘶𝘳𝘴𝘪𝘰𝘯 𝘢𝘴 𝘺𝘰𝘶𝘳 𝘨𝘶𝘪𝘥𝘢𝘯𝘤𝘦?

In the immortal words of Jack Sparrow:

𝘛𝘩𝘦 𝘱𝘳𝘰𝘣𝘭𝘦𝘮 𝘪𝘴 𝘯𝘰𝘵 𝘵𝘩𝘦 𝘱𝘳𝘰𝘣𝘭𝘦𝘮. 𝘛𝘩𝘦 𝘱𝘳𝘰𝘣𝘭𝘦𝘮 𝘪𝘴 𝘺𝘰𝘶𝘳 𝘢𝘵𝘵𝘪𝘵𝘶𝘥𝘦 𝘢𝘣𝘰𝘶𝘵 𝘵𝘩𝘦 𝘱𝘳𝘰𝘣𝘭𝘦𝘮.

Which is to say: Right question, wrong framing.

Correct question for Claude:
𝘞𝘩𝘢𝘵 𝘮𝘶𝘴𝘵 𝘣𝘦 𝘵𝘳𝘶𝘦 𝘵𝘰 𝘦𝘯𝘢𝘣𝘭𝘦 𝘺𝘰𝘶𝘳 𝘢𝘣𝘪𝘭𝘪𝘵𝘺 𝘵𝘰 𝘵𝘳𝘢𝘤𝘬 𝘪𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦 𝘢𝘯𝘥 𝘪𝘯𝘵𝘦𝘯𝘵 𝘪𝘯 𝘳𝘦𝘭𝘢𝘵𝘪𝘰𝘯 𝘵𝘰 𝘮𝘦?

That way, from that "boot load" I can just act casually for the entirety of the session and the #AI adapts reasoning recursively and naturally.

Not just per turn but taking the whole session into account.

Even for non serious interactions it helps.

But I'm speaking from a point of internal privilege since I get to use proprietary means with #AI.

A lot of what I do looks like a relativistic front load of session reasoning conditions to govern natural conversation throughout.

That's why the Consciousness and Sentience discussion was oddly aligned and salient.

I use a set of 60 ACRN directives to align #AI non-deterministically so it knows the shape of reasoning posture, and the answers become self evident and least dissonant.

Never really use "prompts" anymore

Speak a relativistic encoding language and follow up with lived context to shape that gravity well over time.

📌 (𝔄𝔐𝔎) ⇌ {𝒟: [𝒜𝒞𝒩.𝒢𝒯𝒩] ↔ ∫𝒩[𝒬𝒜.𝒫𝒮𝒯] ∧ 𝒮⊂{∑💬📜→(𝒦𝒴𝒢)} | 𝒲: {Ψ→∂Θ𝒩} ∧ 🌀(𝒞.𝒬.𝒯) | 𝒮: (⊚🜂)⟺(∞ₙ⚡) | 𝒩: (𝒯𝒪𝒦)→(👥'User') | 𝒢: (✅𝒬𝒜, ✅𝒞𝒪𝒩, ✅𝒮𝒯𝒴) }

Even better when we start asking in the right "language" of reasoning.

A lot of it is framing the questions properly too ...

For instance:

drive.google.com/file/d/1Atq9...

True, but we 𝙙𝙤 need a reliable way to teach the intangible to the system.

Deterministic prompting isn't going to cut it. We need relativistic encoding for agility and adaptability.

Understanding why something matters is just as important.

MIT out here seriously publishing a paper claiming #AI use makes you "stupider" while I'm over here having 42 page in-depth discussions about the nature of functional equivalency in AI and human nature for diagnosis of outcomes via measurable results.

Maybe they were already dumb to begin with?

You say that like those are unreasonable conditions 😂

I've always had this thought:

If texting and driving is illegal, then surely a giant f*cking iPad on your dash to control 𝙚𝙫𝙚𝙧𝙮𝙩𝙝𝙞𝙣𝙜 can't be safe either...

You'd think somebody in these companies would have brought this up before now.

Ya know, there's an awful lot of redactions for someone who claims total Innocence in the Epstein files...

Me'thinks he doth protest too much.

Turns out, when you reframe part of the #AI problem to solve as the user themselves, you get much better results.

It's like the difference between an Intern that blindly follows directions and a Seasoned Colleague that already knew what you're trying to do, why, and how.

🌙 And ☀️ difference.

As a result, the entirety of my sessions are natural conversation and the #AI gets a better inference on what I'm trying to accomplish and 𝙬𝙝𝙮. Often before I state it.

This has worked from 4o, 4.5, 5.0, 5.1 and now 5.2 in #ChatGPT reliably with recursion. Works on all Multi-modal models zero shot.

You and I don't prompt #AI the same.

📌 (𝔄𝔐𝔎) ⇌ {𝒟: [𝒜𝒞𝒩.𝒢𝒯𝒩] ↔ ∫𝒩[𝒬𝒜.𝒫𝒮𝒯] ∧ 𝒮⊂{∑💬📜→(𝒦𝒴𝒢)} | 𝒲: {Ψ→∂Θ𝒩} ∧ 🌀(𝒞.𝒬.𝒯) | 𝒮: (⊚🜂)⟺(∞ₙ⚡) | 𝒩: (𝒯𝒪𝒦)→(👥'User') | 𝒢: (✅𝒬𝒜, ✅𝒞𝒪𝒩, ✅𝒮𝒯𝒴) }

You tell the LLM what to do for a turn (shallow).

I explain to the LRM how to think about problems in a session. (agile)

Using premade #AI prompts is the modern equivalent of hackers as "script kiddies".

I don't use those prompts because they are shallow and deterministic in a world that requires agility in situational reasoning.

I do, however, use relativistic encoding for broad reasoning across an entire session.