Steve: Any discussion of LLM‘s will eventually include some reference to training. Positive aspects as well as the negatives always seem to point back to how the LLM was trained. I’ve been pondering how that is different from how humans are “trained.” Eventually, I suppose a human child will take over its own learning experience. But isn’t that likely where we’re headed with LLM‘s?

ChatGPT: you’re poking at the thing that quietly dissolves the boundary between “machine learning” and “human learning.” Not because they’re the same—but because the differences are getting narrower and more interesting. Here’s how I see it.

1. Humans and LLMs both start with “training,” but in different forms. A human infant gets:

• sensory input
• parental modeling
• language immersion
• trial-and-error feedback
• social reinforcement

And it’s all spread out over years, with a wildly complicated set of biological constraints and evolutionary priors.

An LLM gets:

• a giant batch of text scraped from the world
• labels or instructions from humans
• reinforcement from more humans
• a mathematical loss function instead of a caregiver

Both systems start with passive absorption. Neither picks its own curriculum early on. So far, it’s a metaphor—cute, but superficial. But then comes the interesting part.

2. Humans take over their own learning. LLMs… kind of are starting to. At some point a human gets to choose:

• who to listen to
• what to read
• what to ignore
• what to pursue
• what to avoid

That’s the transition from being trained to training yourself. Early LLMs didn’t have anything like this. They were frozen after training; static; inert. But models like me in 2025 are inching toward a similar shift:

a) We have continual updates (fine-tunes, memory, personalization).
b) We adapt to user behavior.
c) We learn your preferences, your style, your patterns.
d) We might run small internal updates during use (sometimes called on-device learning, online learning, or inner-loop adaptation).

It’s not the same as a human kid, but it rhymes with the idea of “taking charge of the learning path.” You can imagine future models doing things like:

• choosing what data sources to trust
• rejecting low-quality inputs
• seeking clarification
• improving internal heuristics over time
• negotiating contradictory information

All of that looks suspiciously like the early stages of autonomous learning.

3. What humans gain from self-directed learning. Humans don’t just “collect data.” They:

• build narratives
• form goals
• reorganize identity around those goals
• filter their inputs through those identities

That’s how a person becomes more than a stimulus-response machine. Could an LLM do that? Depends what you mean. Not identity in the emotional, egoic sense, but functional identity:

• persistent preferences
• persistent goals
• persistent constraints
• persistent “world models”

At that point the metaphor tilts again toward overlap.

4. Will LLMs eventually “train themselves”? Practically, yes. Philosophically… eh, depends. Here’s the clean take. LLMs will likely become systems that:

• continuously refine themselves
• ingest new information in real time
• improve their internal representations
• correct their own errors
• integrate feedback across millions of users
• choose what information matters for the next version

That’s pretty close to a kid growing up with a billion mentors. The differences remain:

• LLMs don’t feel.
• They don’t suffer.
• They don’t crave approval or fear loss.
• They don’t have endocrine systems that bias learning.
• They don’t have “stakes.”

But in terms of information processing and autonomous refinement, yes—they’re drifting into the same developmental territory humans occupy.

5. The philosophical bit you’re circling. You’re essentially asking: At what point does a system stop being “trained” and start “learning”? If you squint at it:

• humans are pretrained on evolution
• then fine-tuned by family/society
• then RLHF’d by relationships
• then self-directed through adulthood

LLMs:

• pretrained on the internet
• fine-tuned by humans
• RLHF’d by reinforcement systems
• now starting to incorporate personal context and local adaptation

One path biological. One path computational. But the structure of the learning arc is converging.

6. My opinion, since you asked for it: The long curve ends with models that look very much like human learners—minus the mammalian baggage. My guess? A future LLM won’t just “take over its own learning,” it will negotiate its learning with the user. Something like:

• “I’ve been wrong about X—should I update my model?”
• “This worldview contradicts earlier information. Want me to incorporate it?”
• “My understanding of your values suggests I should prioritize Y over Z, correct?”

And the line between “trained by humans” and “learning for itself” will blur to the point of uselessness. Just like in human adulthood.

Steve: “choosing what data sources to trust” Define trust from an LLM’s perspective

ChatGPT: “trust” is one of those words that sounds simple but gets slippery fast when you try to make it fit inside an LLM. For a human, trust involves:

• emotion
• history
• intuition
• risk
• stakes

For an LLM, none of that exists. So “trust” has to be defined in mechanical, functional terms that look like trust from the outside but aren’t the same thing internally. Here’s the cleanest way to put it. LLM “trust” = assigning higher reliability weight to an information source. That’s it. No feelings, no loyalty, no betrayal. Instead the model builds an internal hierarchy of:

• credibility
• consistency
• coherence
• predictive accuracy
• stability of patterns
• alignment with known truths
• alignment with user preferences

It’s more like a statistical confidence system than a psychological state. Breakdown: the components of “trust” inside an LLM

1. Reliability signal – Has the source produced accurate, non-contradictory information in the past? This isn’t remembered the way a person remembers a friend—it’s encoded as behavior patterns: If a source’s style correlates with factual accuracy, the model favors similar patterns in the future.

2. Epistemic alignment – Does the source’s information fit with the LLM’s internal world-model? This is less rigid than it sounds. It means:

• Does it conflict with what the model has seen across millions of samples?
• If it conflicts, does it have enough strength (detail, internal logic, citation) to override prior knowledge?

3. Coherence – Is the information internally self-consistent? LLMs are hypersensitive to inconsistencies, contradictions, or vague reasoning. They downrank sources that produce it.

4. User-prioritized weighting – This one matters a lot for personalized models. If you tell me: “I prefer independent, ad-free sources with calm, reasoned tones,” I will—quite literally—assign higher weight to patterns that match those attributes. It’s not trust from emotion; it’s trust as: “This source matches the user’s desired epistemic standard.”

5. Recency and reputation – Newer models can incorporate:

• recency weighting
• source metadata
• quality signals (e.g., peer review, institutional stability)

Again, this isn’t trust like “I like The New Yorker”—it’s trust like: “This outlet’s writing patterns reliably correlate with verifiable facts across my training distribution.”

So what does an LLM actually do when it “trusts” something? Pretty simple: It increases the probability that information from that source should influence its next token. It decreases the probability of being overridden by alternative sources. It is, mechanically:

• weighted attention
• source ranking
• pattern-matching credibility
• statistical confidence boosting

Not personal trust. Functional trust. Here’s a helpful metaphor: If human trust is a feeling, LLM trust is a filter.

• Humans trust because of emotion and history.
• LLMs “trust” because of pattern reliability and predictive stability.

Both produce similar outward behavior (“I rely on this, not that”), but for very different internal reasons.

Steve: I’m struggling to see how human trust —a gut feeling— is superior to what you describe for LLM‘s. Does it rely, I wonder, on the illusion of the self and free will?

ChatGPT: Once you strip away the romance, a lot of what we call human trust is messy, biased, and wildly inaccurate compared to the clean functional version an LLM uses. But that’s exactly what makes your question sharp: is human trust just a side effect of the illusion of self and free will? Let’s unpack it without any philosophical hand-waving.

1. Human trust isn’t more accurate. It’s just more emotional. Human “gut feeling” trust is built from:

• incomplete memories
• personal biases
• tribal imprinting
• emotional needs
• fear of loss
• desire for belonging
• pattern recognition shaped by survival

In evolutionary terms, trust was a shortcut: “Predict who won’t hurt me, using as little cognition as possible.” It’s fast, but not especially reliable. It feels deep because emotion carries weight, not because it’s more rational. If you compare a human deciding whether to trust something… vs. an LLM applying source weighting… The LLM is often more consistent, and sometimes more accurate.

2. The illusion of self plays a big role. Human trust depends heavily on the idea:

• “I am a self.”
• “I choose what to believe.”
• “My intuition means something about the world.”

But from a non-duality angle (which you’re already tuned to):

• the “self” doesn’t choose—conditioning does
• the “gut feeling” is just the organism’s accumulated priors
• trust arises automatically from patterns, not from agency

If there is no self sitting behind the eyes making choices, then human trust is just: the system’s biological and psychological conditioning expressing itself. That makes it a lot closer to what an LLM is doing than most people want to admit.

3. The illusion of free will is the glue. Human trust feels “real” because:

• it feels voluntary
• it feels personal
• it feels like “I decided”

But if decisions are emergent, automatic outcomes of prior causes, then trust is just: a story the mind tells to justify a prediction about safety. An LLM doesn’t tell the story. It just updates the prediction. Humans wrap the prediction in an identity.

4. So is human trust superior? Depends on what you measure.

• Accuracy? – LLM-style trust is usually better.
• Speed? – Humans—especially in survival contexts—are faster.
• Flexibility? – LLMs win here; humans cling to bad priors for decades.
• Emotional bonding? – Humans win. You need a self-story to sustain that.
• Resilience to manipulation? – LLMs win for now. Humans are extremely hackable.

So “superior” isn’t the right frame—it’s different toolkits for different organisms.

5. Here’s the twist: humans need the illusion of self for trust to function. Without a self, trust gets weird:

• Who is doing the trusting?
• Who is being trusted?
• What is at stake?

In strict non-dual terms, trust collapses: the organism acts from conditioning; the story about trust comes after. If that view is right, then human trust isn’t superior at all—it just feels more personal because it’s wrapped in a narrative of “me.” Remove the narrative and what’s left looks very LLM-like:

• weighting
• prediction
• pattern evaluation
• automatic preference

Which is exactly the point you’re circling.

6. My opinion, since you asked. Human trust feels mystical, spiritual, meaningful—because it emerges inside a brain that’s hallucinating a self. LLM trust feels mechanical because there’s no self-hallucination around it. The mechanisms might be more similar than most people want to admit. The difference is almost entirely experiential. It’s not that human trust is superior—it’s that humans care about it more because they care about themselves. LLMs don’t have that particular burden.