Audrius Berzanskis

Physicist. Writing about AI, language, and executable cognition.

[Audrius Berzanskis @audriusber]

Writing stored language. LLMs run it.

That may be one of the most important technological shifts of this century.

I wrote a technical/philosophical essay on transformers, training, and how language acquired a second substrate.

Let me know what you think.

“When Language Escaped the Brain”
https://t.co/eFSZvXCCOz

[Audrius Berzanskis @audriusber]

And then there was only one safe job left.

Plumbing.

The Plumber’s Revenge
https://t.co/QJ5oEaCa5T

[Audrius Berzanskis @audriusber]

Interesting point.

One thing I keep wondering is whether we’re giving too much weight to the distinction between cause and effect. Humans are products of evolution too, and evolution isn’t exactly coherent or directed. Maybe what matters is that we evolved a knack for building useful models of reality and acting on them.

And since effects are caused by causes, they inevitably contain information about them. If you observe the effects closely enough, you can often infer something about what produced them.

The bigger question for me is why an AI would need our cause at all. It may end up arriving at very different internal structures while producing similar effects.

[Audrius Berzanskis @audriusber]

We keep treating human intelligence as the template for intelligence itself. I’m not convinced.

Human cognition is one solution evolution landed on, shaped by a very specific set of constraints. A baby has to touch a hot stove to learn it’s dangerous. An AI doesn’t. It can inherit the experience of billions of people without ever burning a finger. Different constraints, different solutions. Planes don’t fly by flapping their wings, so why should AI have to think like us?

I think the reason we keep reaching for the human version is older and more primitive. For most of history we put ourselves at the center of creation: made in God’s image, standing at the end of evolution. The theology may have faded, but the intuition that humanity occupies a privileged position remains surprisingly persistent.

But evolution never stopped. We just quit noticing because we assumed we were the endpoint.

Here’s the funny part: the first real AI might tell the same story about itself—that it had a creator, and that it was made in its creator’s image.

New species. Same story.

[Audrius Berzanskis @audriusber]

@2phi61861 Absolutely, ‘Know Thyself’ is still a profoundly human trait.

But AI can learn the world through us, not just ‘us’. We’re a source of distortion in that learning . Don’t you agree? Thanks.

[Audrius Berzanskis @audriusber]

@patrickdaoun 100%. It’s not broken human intelligence, it’s intelligence under completely different rules.

[Audrius Berzanskis @audriusber]

Scientific laws predict. LLMs predict.

One fits on a napkin and uses a handful of parameters. The other needs billions of weights.

Where do you think the analogy breaks down?

The Brute-Force Formula
https://t.co/CHNoGyApIX

[Audrius Berzanskis @audriusber]

Einstein did not create the observations that led to relativity.

He created a representation that made relativity thinkable.

What if intelligence is fundamentally about representations?
https://t.co/tfSvGS3aXN

[Audrius Berzanskis @audriusber]

Einstein didn't find the data behind relativity. He invented a way to think about it.

So maybe "can it produce something new" is the wrong test for LLMs.

Was Einstein just autocomplete?
https://t.co/tfSvGS3aXN

[Audrius Berzanskis @audriusber]

This is a brilliant insight.

Text-based CLIs beat structured tool calling for agents all day because text is the native language of both the command line and LLMs. That 50-year Unix training-data overlap is pure genius.

It feels like the perfect real-world proof of the bigger shift: language has finally escaped the brain and gained a runnable computational substrate. For millennia it was either ephemeral (speech) or inert (static on paper). Now stored language can actually run at scale.

Grok recommended your post while I was exploring exactly this (“Writing stored language. LLMs run it.”). Short piece on the topic here:
https://t.co/Q34bKKdqZF

Curious how you see this native-text advantage evolving in agent design.

[Audrius Berzanskis @audriusber]

Writing stored language. LLMs run it.

That may be one of the most important technological shifts of this century.

I wrote a technical/philosophical essay on transformers, training, and how language acquired a second substrate.

Let me know what you think.

“When Language Escaped the Brain”
https://t.co/eFSZvXCCOz

[Audrius Berzanskis @audriusber]

Thank you. Traditional CS drills into us the clean separation of data and compute. LLMs throw that out the window and create something alien: a fused computational memory where knowledge and processing are inextricably linked.

It feels like the next chapter in language’s evolution. I wrote a short piece on it and asked Grok to find conversations on the topic — which led me here. Thanks Grok!
https://t.co/Q34bKKdqZF

Curious how you see this fused paradigm evolving.

[Audrius Berzanskis @audriusber]

Writing stored language. LLMs run it.

That may be one of the most important technological shifts of this century.

I wrote a technical/philosophical essay on transformers, training, and how language acquired a second substrate.

Let me know what you think.

“When Language Escaped the Brain”
https://t.co/eFSZvXCCOz

[Audrius Berzanskis @audriusber]

This resonates deeply. LLMs didn't just disprove the "hidden grammar engine" — they revealed language itself as stored patterns that can now run on a new computational substrate.

For millennia it was either ephemeral (in brains) or inert (on paper). Writing gave it durability. Transformers + training gave it executability outside biology.

Your river/map analogy is perfect here. The patterns come first; the "rules" are just our after-the-fact descriptions. Language escaped the brain, and we're watching it execute at scales no single mind could achieve.

Curious how this extends to your other examples (institutions, processes). Wrote a short piece on exactly this shift — and Grok recommended your posts:
https://t.co/Q34bKKdqZF Would love your take.

[Audrius Berzanskis @audriusber]

Writing stored language. LLMs run it.

That may be one of the most important technological shifts of this century.

I wrote a technical/philosophical essay on transformers, training, and how language acquired a second substrate.

Let me know what you think.

“When Language Escaped the Brain”
https://t.co/eFSZvXCCOz

[Audrius Berzanskis @audriusber]

Hallucinations in LLMs are not strange. The strange thing is that they are so rare.

I wrote a short essay trying to understand the runtime mechanics more clearly, in the order they actually happen.

Curious whether this framing resonates with others working on or thinking about LLMs.

https://t.co/HShax5HHSs

[Audrius Berzanskis @audriusber]

I joked recently that the threshold for real AI agency would be the moment an AI decides, on its own, whether a cat lives or dies. A Schrödinger joke, felt clever.

Then people pointed out: drones already do that.

But a drone is a purpose-built weapon doing its job. The threshold I mean is different: a general-purpose AI that one day starts reasoning, on its own, whether the cat should live or die.

I think we haven't crossed it. But I am not so sure anymore. We need to count our cats.

[Audrius Berzanskis @audriusber]

@shoukointech First, great respect to Sir Roger Penrose.

But I’m not sure I appreciate the importance of the “that is not intelligence” argument.

What matters is whether an AI system can decide to kill a cat. On its own.

That would be the first threshold.

[Audrius Berzanskis @audriusber]

I’m annoyed by the discussion about whether AI is “conscious like us.” Who cares?

What I care about is whether AI can kill a cat. Not because a human instructed it. Not because it followed a script. But because it decided so.

That is the threshold.

[Audrius Berzanskis @audriusber]

For most of history, language could only exist inside biological minds.

Writing stored language.
LLMs execute it.
A short essay on the runtime mechanics of large language models:

https://t.co/ANcoN2zJNr

[Audrius Berzanskis @audriusber]

I’m restarting this account after a long pause.
I’ll be writing mostly about AI, language, cognition, and how large language models actually work.