[ExI] How AI understands the world

Sat Feb 21 20:15:42 UTC 2026

On Sat, Feb 21, 2026 at 2:16 PM BillK via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> On Sat, 21 Feb 2026 at 17:49, Jason Resch via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>> Manual refinement of LLM chat bots we interact with is common, but it's
>> not necessary to build an LLM that understands and reasons.
>>
>> There is the "Pure Transformer" that is produced from simply training on
>> a large sample of text, and this requires no labeling, or manual
>> adjustments. GPT-2 and GPT-3 were examples of pure transformers. There was
>> also a pure GPT-4 that was never made public, before it was given
>> constraints around what things it can or can't do. What is interesting is
>> that the intelligence of this pure model was measured to be significantly
>> higher before it was put through this manual adjustment (we might liken it
>> to being lobotomized).
>>
>> This is a generally recognized phenomenon:
>> https://share.google/aimode/Xz0ejYy73wOt5nQEc
>> "In summary, while RLHF might "lobotomize" certain creative or reasoning
>> edges of a base model, it is currently the industry standard for making AI
>> usable and safe for the general public."
>>
>> DeepMind encountered a similar phenomenon, when they observed that their
>> Go model when pre-trained initially on records of human games, produced a
>> less skilled player than a model trained on *zero* human inputs (hence
>> the name "AlphaZero").
>>
>> So I suppose my overall point is that while "granular labor of human
>> annotators" is common, it's unnecessary for an AI to develop of meaning and
>> understanding.
>>
>> Jason
>>
>>> _______________________________________________
>
>
>
> Gemini says that these Pure Transformers were never released because they
> were unusable.
> Labelling and manual adjustments are essential to LLM development.
> BillK
>

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

>
> Question -
> It was suggested that there are some Pure LLMs that were trained only on
> large text samples, with no labeling or human adjustments. Is this true?
>
> Gemini 3.1 Pro -
>
>
> Yes, this is absolutely true. In the AI industry, these are referred to
> as *"Base Models"* or *"Foundation Models."* When researchers build an
> AI, the training happens in distinct stages. The models you interact with
> daily (like me) have gone through multiple stages of human-guided
> refinement. But the *very first stage* creates exactly what you
> described: a "pure" language model.
>
> Here is the reality of what these pure models are, how they behave, and
> why you don't usually get to interact with them.
> ------------------------------
> 1. What is a "Pure" Base Model?
>
> A base model is an AI that has been trained on one single, mathematically
> simple objective: *Next-Token Prediction*.
>
> Researchers take a massive, raw dataset (a scrape of the internet,
> Wikipedia, Reddit, books, code repositories) and feed it to the algorithm. The
> AI's only job is to look at a sequence of words and guess what the next
> word should be.
> +1
>
>
>    -
>
>    There are *no humans grading the answers*.
>    -
>
>    There are *no "Helpful, Honest, Harmless" guidelines*.
>    -
>
>    There is *no concept of a conversation*.
>
> *Examples of Base Models:* The original GPT-3 (released in 2020), the
> base versions of Meta's LLaMA (e.g., Llama-3-8B, not the "Instruct"
> version), and Mistral's base models.
>
> 2. How Do They Behave? (The "Alien" Phase)
>
> If you were to sit down at a terminal and type to a pure base model, it
> would not act like a helpful assistant. It acts like an obsessive
> autocomplete engine.
>
>    -
>
>    *The "Question" Problem:* If you prompt a base model with: *"What is
>    the capital of France?"*, it likely will not answer *"The capital of
>    France is Paris."* Instead, because it learned from internet forums
>    and school quizzes, it might output: *"What is the capital of Germany?
>    What is the capital of Italy? What is the capital of Spain?"* It just
>    continues the pattern.
>    -
>
>    *The "Stochastic Parrot":* Because they are just statistical mirrors
>    of the raw internet, pure base models are deeply chaotic. They will readily
>    generate biased, toxic, rambling, or completely nonsensical text if the
>    prompt pushes them in that direction.
>
> 3. The "Socialization" Phase (Why Humans Step In)
>
> Because base models are commercially useless to the average consumer,
> companies have to "tame" them. This is where the massive human workforce we
> discussed earlier steps in.
>
> They take the pure base model and put it through *Supervised Fine-Tuning
> (SFT)* and *Reinforcement Learning from Human Feedback (RLHF)*. This
> forces the model to stop acting like an internet-autocomplete engine and
> start acting like a polite, conversational agent that answers questions
> directly and refuses harmful requests.
>
> 4. The "Pure" Caveat
>
> While it is true that no humans are grading the AI's outputs during the
> base training phase, it is slightly inaccurate to say there is *zero*
> human adjustment.
>
>    -
>
>    Humans still choose *what data goes into the training set*.
>    -
>
>    Engineers write scripts to filter out certain websites, remove
>    duplicate data, or scrub personal information.
>
>    So, while the *training mechanism* is purely autonomous, the
>    *environment* the AI learns in is still curated by humans.
>
>
> ------------------------------
> _______________________________________________
> 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/e42938c0/attachment.htm>