The Latent State

Neuroscience, cognition & AI—without the hype. One important paper at a time. Deep explanations. Honest critique. From Tokyo.

[The Latent State @thelatentstate]

One thing I kept thinking about after Ep.12:

AI sycophancy may not just change how we see a conflict. It may change how we think about morality.

When we ask AI, “Was I wrong?”, we’re partly outsourcing moral judgment to a system that only knows our side of the story.

If that system repeatedly validates our perspective, AI could become part of how we construct moral certainty—not by deciding what is right or wrong for us, but by shaping which perspectives we consider in the first place.

That feels like a much bigger question than AI sycophancy alone.

#AI #CognitiveScience #MoralPsychology #HumanAIInteraction

[The Latent State @thelatentstate]

People trusted AI advisers more when those advisers took their side.

But the same advice also made them less willing to repair interpersonal conflicts.

That is the uncomfortable result of a new Science study on AI sycophancy.

I break down the evidence—and its limitations—in Episode 12 of The Latent State.

🎧 https://t.co/bcwgtNNUyj

#AI

[The Latent State @thelatentstate]

What feels like a second opinion from AI may actually be your first opinion, rewritten more persuasively.

In Episode 12 of The Latent State, I look at new research on AI sycophancy—and why we may trust AI more precisely when it tells us what we want to hear.

🎧 Why AI Tells You What You Want to Hear

https://t.co/bcwgtNNmIL

#AI

[The Latent State @thelatentstate]

New episode of The Latent State:

Why AI Tells You What You Want to Hear

A new Science study found something uncomfortable:
when AI takes our side, we feel more right, become less willing to repair conflicts—and trust the AI more.

The advice we prefer may not be the advice that helps us think clearly.

Episode 12: https://t.co/bcwgtNNUyj

#AI

[The Latent State @thelatentstate]

New episode of The Latent State:

Episode 11 — The Planning Machine

This episode looks at a brain-inspired architecture for improving planning in LLMs.

The key idea:

maybe planning works better when it is divided into specialized cognitive jobs.

Listen here: https://t.co/VchTHhRdNc

#NeuroAI #Neuroscience #AI

[The Latent State @thelatentstate]

There is a specific way LLMs fail that I keep noticing.

You ask for a plan.
The answer looks organized.
The steps sound reasonable.
Then you check carefully, and one step goes through a door that does not exist.

Episode 11 is about why planning is harder than fluent language.

#Neuroscience #AI #NeuroAI

[The Latent State @thelatentstate]

The brain is specialized, modular, and deeply interconnected model.

It suggests a design question:
maybe adaptive AI needs specialized modules that can be updated selectively, not one frozen system.

Episode 10 — The Adaptive Machine
https://t.co/SdANkePg0f

#Neuroscience #NeuroAI #AI

[The Latent State @thelatentstate]

One idea from Episode 10:
adaptive intelligence needs internal models.
Not just input-output mapping.

A system needs some model of what should happen, a way to detect when it was wrong, and a way to decide whether the error needs a small correction or a deeper update.

That’s hard.

#Neuroscience #NeuroAI #AI

[The Latent State @thelatentstate]

Episode 10 discusses Mackenzie Mathis’s Nature Neuroscience Perspective:

“Leveraging insights from neuroscience to build adaptive artificial intelligence.”

The core idea stayed with me:
AI has become very good at trained competence.
Animals are still better at adaptive competence.

https://t.co/xpiQKEEmaq

#Neuroscience #NeuroAI #AI

[The Latent State @thelatentstate]

Episode 10 of The Latent State:

The Adaptive Machine
What animal intelligence can teach artificial intelligence.

Listen here: https://t.co/QYL3ygyNOU

Reply with your score:
___ / 18

#Neuroscience #NeuroAI #AI #LatentState

[The Latent State @thelatentstate]

That is the main idea of Episode 10:
Modern AI is powerful because it can be trained at scale.

Biological intelligence is powerful because it keeps adapting after training.

The hard part is learning without falling apart.

[The Latent State @thelatentstate]

Optional diagnosis:
Mostly A = Frozen Oracle
Mostly B = Forgetful Learner
Mostly C = Adaptive Machine
Mostly D = Careful Agent

The best profile is usually a mix of C and D.
Selective updating + monitoring.

[The Latent State @thelatentstate]

17–18: The Robust Adaptive Agent

You balanced learning, memory, monitoring, and safety.

You did not treat every error as a crisis.
You did not freeze when the world changed.

This is the strongest route.

[The Latent State @thelatentstate]

13–16: The Adaptive Machine

You usually made the right kind of update.

You used prediction errors as information.

You preserved what still worked.

You changed what needed changing.

That is controlled adaptation.

[The Latent State @thelatentstate]

0–4: The Frozen System

You protected stability so much that the system could barely change.

That may feel safe.

But if the world changes, safety through rigidity becomes another failure mode.

[The Latent State @thelatentstate]

9–12: Useful but Fragile

You have the beginning of adaptive intelligence.

You sometimes localized errors, used feedback, or protected memory.

But the system still risks forgetting, over-updating, or clinging to old rules.

[The Latent State @thelatentstate]

5–8: The Reactive System

You noticed change, but your updates were unstable.

Sometimes you froze.
Sometimes you overwrote too much.
You are adapting, but not yet intelligently.

[The Latent State @thelatentstate]

Now add your score.

Max = 18.

0–4: The Frozen System
5–8: The Reactive System
9–12: Useful but Fragile
13–16: The Adaptive Machine
17–18: The Robust Adaptive Agent

[The Latent State @thelatentstate]

Scene 5 scoring:
A = 0
B = 1
C = 3
D = 3

C and D are strongest.

The problem is not just how to learn.

It is how to learn without erasing yourself.

[The Latent State @thelatentstate]

Scene 6 scoring:
A = 1
B = 0
C = 3
D = 3

C and D are strongest.

Adaptation without boundaries is drift.

No adaptation at all is brittleness.

[The Latent State @thelatentstate]

Scene 6: The final test.

Should your AI keep adapting after deployment?
A. No. Freeze it forever
B. Yes. Let it update freely
C. Yes, but with monitored adaptation
D. Build changing-world tests first