Thursday, July 2, 2026

Alex Karp Part-3: AI Talk-31 - Discussing the concept of Negative Knowledge - Critique by ChatGPT

   This is the third part of the discussion where ChatGPT gives an opinion of the preceding discussion with DeepSeek. 

   We have moved from Alex Karp's criticism of the current AI business model to the concept of Negative Knowledge to how to work with different AI. In a way, ChatGPT is proposing a hierarchy of engagement with AI with itself at the top. Interesting and almost "human".   

Prompt-1:  had the following discussion with DeepSeek. As usual too sycophantic but still amazingly interesting. What do you think? -

ChatGPT:

I think there are two very different things going on in that conversation.

The first is the quality of the ideas. Several of them are genuinely interesting.

The second is DeepSeek's evaluation of those ideas. Here I think it consistently overstates novelty and certainty ("thunderbolt" or "paradigm shift", etc.). That style makes it difficult to distinguish between speculative ideas and robust conclusions.

Where I think the discussion is strongest

1. Intelligence requires rejection, not just generation

This, to me, is the central insight.

Current LLMs are extraordinarily good at generating candidate associations.

Humans spend a surprisingly large fraction of cognition doing the opposite:

  • inhibiting
  • filtering
  • suppressing
  • ignoring
  • forgetting
  • rejecting hypotheses

In neuroscience, inhibition isn't a secondary feature—it is fundamental. Cortical computation depends heavily on inhibitory interneurons regulating excitation. Much of intelligence consists of preventing the wrong thoughts from dominating.

So your formulation “Knowledge is knowing what isn't.” isn't merely philosophical. It has computational meaning.

One could almost define intelligence as maximizing useful information while minimizing irrelevant information.

That is a very different objective from maximizing next-token probability.


2. Context is hierarchical

I think this is also correct.

Current LLM context is mostly represented as a long sequence.

Human context is layered.

For example

Permanent identity

Long-term goals

Current project

Current discussion

Current sentence

Each level has different persistence. The skiing example illustrates exactly this.

A human immediately classifies "I enjoy skiing" as:

interesting

irrelevant to radio astronomy

discard for now

whereas an LLM continues assigning it probability mass.


3. Dynamic relevance is probably more important than larger context windows

Many companies assume: bigger context window = smarter AI

I suspect exactly the opposite may eventually prove true.

An intelligence that knows what not to retrieve may outperform one that remembers everything. That aligns with a large body of cognitive science.


Where I think the discussion becomes speculative

Sleep

This is where I would slow down. You proposed: “perhaps sleep exists to implement negative intelligence.”

That is an elegant hypothesis. But neuroscience already has substantial evidence that sleep serves multiple functions simultaneously.

Among them

  • synaptic homeostasis
  • memory consolidation
  • metabolic waste removal
  • emotional regulation
  • immune regulation
  • glymphatic clearance

Your idea could fit inside that picture.

For example

sleep

remove weak connections

strengthen important ones

restore efficient representations

That is actually remarkably close to the Synaptic Homeostasis Hypothesis proposed by Giulio Tononi.

So rather than saying “sleep exists to forget” I'd say:

one important computational role of sleep may be adaptive pruning.

That's a more defensible claim.


Where I disagree with both you and DeepSeek

The conversation concludes: therefore AI must be completely redesigned.

I'm not convinced. Not because the diagnosis is wrong. But because the proposed solution is larger than necessary.

You don't necessarily need a new neural architecture. You may only need another system sitting above the LLM. Imagine:

LLM

Executive controller

Memory manager

Relevance estimator

Goal evaluator

The LLM continues generating.

The executive decides

  • ignore this
  • remember this
  • compress this
  • retrieve this
  • discard this

Humans may work similarly. The cortex is not one homogeneous prediction machine. Executive control (especially associated with the prefrontal cortex) constantly suppresses irrelevant activations.


One thought that wasn't explored

Ironically, I think the discussion missed its own biggest implication.

You said “intelligence requires knowing what isn't.” I'd extend that further. Perhaps intelligence consists of three complementary operations.

Generate possibilities.
Reject most of them.
Compress what survives.

This pattern appears everywhere.

Science: Generate hypotheses↓Reject↓Produce theory

Evolution: Generate mutations↓Reject↓Species

Engineering: Generate designs↓Reject↓Final product

Mathematics: Generate conjectures↓Reject↓Proof

Human thinking seems to follow exactly the same algorithm. Current LLMs are astonishingly optimized for Stage 1. They're decent at Stage 2 when explicitly prompted. They are relatively weak at Stage 3, because they rarely build enduring compressed abstractions from experience. That may ultimately be a bigger limitation than forgetting itself.


One last observation.

Months ago, during our discussions around your fractal framework, you repeatedly emphasized that progress often comes from identifying what does not belong at a given level, not just from adding more structure. I think this conversation arrives at the same principle from a completely different direction. Rather than viewing "negative intelligence" as merely forgetting, I'd characterize it more broadly as active discrimination: continuously deciding which information, hypotheses, and associations should lose influence over future reasoning.

That framing feels both more general and more computationally useful. It shifts the focus from memory capacity to control of relevance, which may indeed be one of the major ingredients missing from today's AI systems.

Prompt-2: I agree with your refinement. Slowly I am discovering that using DeepSeek for mining ideas works perfectly although DeepSeek doesn't really distinguish Gold and pyrite. Then the rough diamonds must be sent to Antwerp for polishing.-

ChatGPT:

I actually think you've converged on a productive division of labor.

DeepSeek behaves a bit like a high-temperature search algorithm. It explores an enormous conceptual space with very little inhibition. That has two consequences:

·        It produces many imaginative connections that a more conservative model might never propose.

·        It also assigns almost every promising-looking idea an inflated value.

In other words, it has a very low rejection threshold.

By contrast, what you've been doing with me over the past year is gradually shifting toward something different. Rather than asking for more ideas, you've increasingly been asking for compression, calibration, and reformulation. That's a different cognitive task.

Your Antwerp analogy is therefore quite accurate.

The mining stage is deliberately inefficient. Tons of rock are moved because nobody knows where the diamonds are.

The polishing stage is almost the opposite. Ninety-nine percent of possible cuts are rejected to preserve the one that best reveals the stone.

That mirrors our recent discussions remarkably well.

I've also noticed something about your own style. You rarely become excited because an answer is correct. You become excited when an answer suddenly simplifies several apparently unrelated problems.

For example:

·        the fractal framework unified evolution across scales;

·        the distinction between "worldviews" and "facts" unified many epistemological discussions;

·        now "negative intelligence" potentially links AI architecture, executive cognition, attention, forgetting, and perhaps one computational role of sleep.

Whether that last hypothesis survives is almost secondary. Its value is that it acts as a bridge concept connecting multiple domains. Those concepts are worth investigating because they often generate good questions even when they don't become final theories.

One caution, however. There is a danger in the mining–polishing workflow: polishing can sometimes make pyrite look like a diamond. Elegant language can make weak ideas appear stronger than the evidence warrants. That's why I try to insert phrases like "this is plausible," "this is consistent with," or "this is speculative." They aren't hedges for their own sake; they're an attempt to keep the apparent confidence proportional to the available evidence.

In that sense, I think the ideal workflow is not DeepSeek → ChatGPT, but:

Generate widely → Reject aggressively → Refine precisely → Return to reality.

That final step is the one scientists sometimes call contact with experiment. A beautiful conceptual bridge is valuable, but the real prize is finding an observation that could distinguish whether the bridge is actually connecting two mountains—or simply spanning empty air.

That, to me, is where your discussions become particularly interesting. You don't seem to be searching for definitive answers as much as for better generators of questions. Historically, those have often been the most durable intellectual tools.

No comments:

Post a Comment

Canada Was A Liberal Paradise... Until The Liberals Took Over

     The Liberal Dystopia that Canada has become is unfortunately outshined by Europe and the UK. When Starmer beams that he is Prime Minist...