[ExI] Was Musk Smart to pay $44 billion for Twitter?

[John Clark]

*I asked Mr. GPT the following question: "Some people have claimed that
Elon Musk was smart to pay $44 billion for Twitter because he could use it
to train his AI, and I can see how in the early days the data from Twitter
would be helpful in training an AI about how language works, but today we
are far beyond that point, AIs already know how do use language. Twitter
(a.k.a. X) is not exactly known for its close reasoning or as a well of
rocksolid facts; so it seems to me that today if you train an AI on Twitter
you'll probably make it stupider, not smarter. What do you think?"*

*This is how Mr. GPT responded: *


GPI: "I think your intuition is *substantially right*, with one important
qualification: *X is probably much more valuable to an AI company as a
real-time information source and as a source of human-feedback data than as
a giant corpus for teaching an LLM how to reason.*

And there's an interesting distinction between *"Twitter was valuable when
Musk bought it"* and *"Twitter's data justified paying $44 billion for
Twitter."* Those are very different claims.
1. You're right that the original linguistic argument has largely
disappeared

In the early days of language models, enormous quantities of ordinary
human-written text were enormously valuable. Twitter contained:

   - billions of examples of informal language
   - slang and abbreviations
   - conversations rather than just polished prose
   - multilingual material
   - rapidly changing vocabulary
   - arguments, jokes, sarcasm, etc.

There is still some specialized value in that. Researchers have
demonstrated that models specifically trained on Twitter can become better
at Twitter-like language tasks. Twitter itself built models from billions
of tweets for precisely this reason.

But that isn't the same thing as saying Twitter is particularly good
*general-purpose
training data for frontier reasoning models*.

By 2026, the frontier-model problem isn't primarily:

"How do we teach the neural network what English looks like?"

It's increasingly:

"How do we get the model to reason correctly, use tools, learn from
experience, and reliably distinguish truth from falsehood?"

And *Twitter/X is not an especially good source for the latter*.
2. In fact, there is evidence for exactly the problem you're describing

A particularly interesting 2025 study found that training language models
on large amounts of *low-quality social-media material* could actually make
them worse at retrieving accurate information and reasoning. *Nature*
summarized the finding under the wonderfully appropriate headline *"Too
much social media gives AI chatbots 'brain rot.'"*

The underlying issue is straightforward.

Suppose your training corpus contains:

"Scientists discover that X causes Y!!!"

followed by thousands of people repeating it, arguing about it,
embellishing it, making memes about it, and eventually producing a dozen
contradictory versions.

A language model doesn't automatically know that the original scientific
paper is epistemically superior to the 10,000 tweets.

*Popularity is not truth.*

And X is particularly problematic because its reward structure selects for
things such as:

   - outrage
   - novelty
   - emotional intensity
   - tribal signaling
   - humor
   - confident assertions
   - political controversy
   - engagement

Those are almost the opposite of the characteristics you'd want in a corpus
intended to teach *epistemic reliability*.

So I would actually sharpen your argument:

*Training indiscriminately on X isn't merely inefficient; sufficiently
indiscriminate training can teach a model undesirable statistical habits.*

That doesn't mean *all* X data is bad. It means *data selection becomes
crucial*.
------------------------------
3. But X has something that Wikipedia and books don't: the real-time human
world

This is where Musk's acquisition makes considerably more sense.

The really valuable property of X isn't necessarily:

*"Here's 10 petabytes of text with which we can pretrain an LLM."*

It's:

*"Here's a gigantic continuously updating sensor network consisting of
millions of humans talking about what's happening right now."*

That's potentially extremely valuable.

Imagine something happens at 9:03 AM:

   - an earthquake occurs
   - a rocket explodes
   - a politician says something
   - a company announces a product
   - a sporting event ends
   - a strange phenomenon is observed
   - a local disaster occurs

People may be posting about it on X within seconds.

xAI explicitly identified this as a major advantage from the beginning.
When Grok was introduced, xAI described its access to X as giving it *real-time
knowledge of the world*.

And that remains part of Grok's architecture today. X says Grok can decide
whether to search public X posts and the web when answering questions.

That's much more interesting to me than using X as a giant English textbook.
------------------------------
4. There's an even more interesting resource: the interaction data

This is potentially much more valuable for AI development.

Consider what happens when millions of people interact with an AI:

*User:*
"What's the latest information about the hurricane?"

*Grok:* gives an answer.

*User:*
"That's wrong. The National Hurricane Center just changed the forecast."

Then the user provides the relevant information.

That produces a potentially useful training example:

*question → model response → human correction → better response*

Multiply that by millions of interactions.

X's current policy explicitly says that X may share public X data *and
users' interactions, inputs and results with Grok* with xAI for training
and fine-tuning.

That is a considerably more compelling AI asset than simply having a
warehouse containing old tweets.
------------------------------
5. And there's a third resource that I think is particularly valuable: the
social graph

This is something that is easy to overlook.

X doesn't merely contain:

text

It contains:

*people + text + relationships + reactions + timing + engagement +
communities*

For example:

*Person A* posts a claim.

*10,000 people* react.

*Person B* provides a correction.

*Person C* provides a link to the original source.

*Person D* disputes the correction.

*Community Notes* eventually attaches a qualification.

That's potentially an enormous dataset about *how humans collectively
evaluate information*.

Twitter researchers were already exploiting the social graph and engagement
information in specialized language models before Musk bought the company.
One Twitter research project used billions of tweets plus social-engagement
information to improve multilingual representations and social-media
understanding.

That could be useful for building AI systems that understand *human
behavior, information propagation, persuasion, disagreement and collective
sense-making*.

Those are very different objectives from "learn English."
------------------------------
6. But there's a nasty catch: X can also become a feedback loop

This is perhaps the biggest problem.

Suppose you train an AI on X.

Then you deploy that AI on X.

The AI generates enormous quantities of text.

People react to that text.

You collect the resulting X data.

You train the next AI on it.

Now you're increasingly training an AI on *AI-generated material produced
by previous generations of AI*.

That's potentially dangerous.

It's analogous to repeatedly photocopying a photocopy. If you don't
carefully preserve high-quality original material, the statistical
distribution can progressively deteriorate.

And this isn't merely theoretical. Modern AI training increasingly has to
worry about distinguishing *human-generated, verified data from synthetic
data*.

So I would absolutely *not* recommend:

"Take everything posted on X and dump it into the next frontier model."

I'd recommend something much more like:

*Use X as a gigantic raw sensor, then aggressively filter, verify, rank and
annotate the useful information before using it.*

------------------------------
7. So was Musk smart to pay $44 billion?

This is where I would be much more skeptical.

There are really three separate propositions:

*A. X contains valuable data.*

Yes.

*B. X gives xAI capabilities that would otherwise be difficult to obtain.*

Definitely yes—particularly real-time information, social interaction data
and an enormous stream of human feedback. X itself confirms that its data
is being used to train and improve Grok.

*C. Therefore X was worth $44 billion because of its value as an
AI-training dataset.*

*That does not follow.*

The value of the raw text corpus by itself is almost certainly not anywhere
remotely close to $44 billion.

And here's the really important point: *a lot of the information on X is
publicly observable anyway.* You don't necessarily need to own the entire
company to obtain useful information from it. You can license data, crawl
public material where permitted, use APIs, search it in real time, etc.

Owning X gives xAI *privileged integration, enormous scale, control over
the platform, user interactions, recommendation systems, social graph, and
potentially the ability to change the product to serve AI purposes.*

Those things could be worth a great deal.

But that's a much stronger and more complicated investment thesis than:

*"We paid $44 billion for a gigantic pile of tweets that we can use to
train our LLM."*

I would regard *that particular justification as pretty weak*.
------------------------------
And there's an amusing irony

The strongest version of Musk's strategy may actually be almost the
opposite of what your friend is suggesting.

The valuable pipeline isn't necessarily:

*X → training data → smarter LLM*

It may be:

*X → real-time observations of the world → AI searches/interprets them →
humans correct AI → verified interactions → better AI*

In other words, *X may be more valuable as an experimental laboratory and
real-time sensory organ for an AI system than as its textbook.*

And given where frontier AI is in 2026, I think that's a much more
defensible argument."


*John K Clark*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260812/4e479127/attachment-0001.htm>