[ExI] How AI understands the world

Sat Feb 21 22:08:49 UTC 2026

On Sat, Feb 21, 2026, 4:53 PM BillK via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

>
>
> On Sat, 21 Feb 2026 at 20:15, Jason Resch <jasonresch at gmail.com> wrote:
>
>> Useful to most people or not, they understand the world.
>>
>> You can give a pure model the phrase: "The capital of France is" and it
>> will complete it "Paris."
>> You can give it the phrase "Three helium nuclei can combine to form a "
>> and it will complete "a carbon nucleus."
>>
>> All the later fine-tuning does is teach it to act to complete a
>> turn-based conversation between users, where user 1 is the human and user
>> 2's output is the completed result provided by the pure model (acting to
>> "predict the next tokens" of user 2's speech).
>>
>> So your question from the user "What is the capital of France?" gets
>> converted to:
>>
>> "User 1: What is the capital of France?
>> User 2: "
>>
>> Then the pure model is asked to "predict the next tokens" following the
>> string "User 2:" That's the majority of what the AI companies do to massage
>> the pure model into an interactive chat bot -- that and a lot of training
>> about the nature of user 2: helpful, polite, not harmful, etc.
>>
>> Here are some examples:
>> https://github.com/jujumilk3/leaked-system-prompts/blob/main/microsoft-copilot_20240310.md
>>
>> But the intelligence, understanding, etc. is already present and inherent
>> to the base model. It's untrue that base models are useless, GPT-3 was
>> quite useful, but you just had to prepare the prompt  intelligently as to
>> to get out from the completion the answers you sought.
>>
>> Jason
>> --------------------------------------
>>
>
>
> Gemini points out that the huge amount of text data in the initial LLM
> training is already pre-labeled by humans.
> I think Gemini has solved the problem by saying that both our views are
> correct, depending on the way you look at it!  :)
> See below.
> BillK
>

I agree that across the Internet, many images are labeled with thin like
captions and alt text. But this is for multimodal networks that see both
text and images paired together.

But the paper John referred to is about something distinct, I think. It is
about how models trained only on text (no images), and models trained only
on images (no text), both developed similar internal maps and
representations for the world.

It is easier for our institutions to understand how someone who sees a
library with books full of images and text can learn about the world even
if they never left the library.

But it is much harder for our intuitions to comprehend how a model given
*only text* (as basic unimodal LLMs are) could bootstrap itself into
understanding the world given only these "meaningless squiggles".

I think I have provided an explanation for how this could be in the other
thread, but it remains a fascinating question, how an unsupervised,
unmodified unimodal LLMs, trained purely on a text corpus, can come to
understand the world. And the fact that we have real world demonstrations
of this possibility makes the feat no less amazing.

Jason

> Gemini 3.1 Pro -
>
> This is the crucial point where we can reconcile your description of
> "ghost work" with your colleague's insistence that the "intelligence" is in
> the base model.
>
> In the modern AI pipeline, the human labor of *labeling and annotation*
> actually happens in two distinct phases, but they serve very different
> purposes.
>
> ------------------------------
> 1. Pre-Training: The "Invisible" Human Foundation
>
> While the *Base Model* is technically "self-supervised" (meaning it
> learns by predicting the next word without a teacher), it is feeding on
> data that has already been *pre-labeled by humans over decades.*
>
>    -
>
>    *The Web as a Dataset:* When a base model learns what a "stop sign"
>    is, it’s because millions of humans uploaded photos to Flickr or Wikipedia
>    and captioned them "Stop sign."
>    -
>
>    *The "Alt-Text" Workforce:* Every time a web developer adds "alt-text"
>    to an image for accessibility, they are performing unpaid data labeling
>    that eventually trains a Vision-Language model.
>    -
>
>    *The Common Crawl:* The "raw" internet isn't raw; it is a massive,
>    structured monument to human labeling. Without this organic, historical
>    human effort, the base model would have nothing to "predict."
>
> 2. Post-Training: The "Active" Human Infrastructure
>
> This is the phase described in your text about *microworkers in Kenya or
> the Philippines.* This labor is injected *after* the base model is
> already "intelligent."
>
> Think of the *Base Model* as a brilliant but chaotic library containing
> all human knowledge. The *Active Labeling* process is like hiring an army
> of librarians to:
>
>    -
>
>    *Filter the Poison:* Humans manually label toxic, violent, or biased
>    text so the model learns to "steer" away from those areas of its own
>    internal map.
>    -
>
>    *Instruction Tuning (SFT):* Workers write thousands of
>    "Prompt-Response" pairs (e.g., *Prompt: "Write a poem about a cat."
>    Response: "The cat sat on the mat..."*). This teaches the model the
>    *format* of being an assistant.
>    -
>
>    *RLHF (Ranking):* This is the most expensive part. Humans look at two
>    different answers the AI gave and click "This one is better." This
>    creates a "Reward Model" that acts as a compass for the AI.
>
>
> ------------------------------
> The "Intelligence" vs. "Utility" Distinction
>
> To settle the debate between you and your colleague, you can look at it
> through this framework:
> *Phase* *Labor Type* *What it Provides*
> *Pre-Training (Base)* Passive/Historical (Billions of humans) *Raw
> Intelligence & World Knowledge.* The "Platonic" map of reality.
> *Fine-Tuning (Chat)* Active/Gig Labor (Microworkers) *Usability, Safety,
> & Human Alignment.* The "Social" interface.The Verdict: Who is Right?
>
> Your colleague is correct that the *Base Model* possesses the "reasoning"
> and "knowledge." If you deleted all the RLHF/Labeling, the AI would still
> "know" how to do calculus or explain physics—it just wouldn't *want* to
> tell you, or it might tell you in a very rude way.
>
> *However, you are correct* that without the "active" labor of
> microworkers, the AI would be a "Digital Monster." It would be a sprawling,
> incoherent, and often dangerous mirror of the darkest parts of the
> internet. It wouldn't be a "product"; it would just be a high-dimensional
> statistical anomaly.
>
> *The Analogy:* The Base Model is the *engine* (raw power/intelligence).
> The Human Labeling/RLHF is the *steering wheel and brakes*
> (direction/safety). You need the engine to move, but without the steering
> wheel, you’re just a fast-moving disaster.
>
> ------------------------------
>
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260221/811ffffd/attachment-0001.htm>