Angelo Dalli

Chief Scientist at @UmnaiBase. Unleashing limitless human ingenuity through a revolutionary machine learning approach: Hybrid Intelligence. AI PhD

[Angelo Dalli @AngeloDalli]

Neurosymbolic AI is not a niche research topic.

It is the natural evolution of AI.

Neural systems learn.

Symbolic systems reason.

Real-world decision systems require both. @UmnaiBase

[Angelo Dalli @AngeloDalli]

Predictions inform.

Decisions act.

Most AI today is still optimized for the former.

[Angelo Dalli @AngeloDalli]

AI without structure struggles with reasoning.

AI without adaptability struggles with reality.

The future requires both. @UmnaiBase

[Angelo Dalli @AngeloDalli]

Now we know why Apple didn’t make the Apple Car: Jony was saving the Pherrari.

Italian translation: Madonna! 😱🤌🏻

[Angelo Dalli @AngeloDalli]

Scale improves capability.

It does not automatically produce reasoning.

That distinction is becoming increasingly important in AI system design. @UmnaiBase

[Angelo Dalli @AngeloDalli]

One gives flexibility.
The other gives structure.

This is why purely statistical systems struggle with:
•traceability
•causality
•defensibility

The future of AI is not neural versus symbolic.
It is neural plus symbolic.

[Angelo Dalli @AngeloDalli]

Neural systems approximate.
Symbolic systems explain.
Modern AI increasingly requires both.

[Angelo Dalli @AngeloDalli]

Symbolic AI is excellent at:
•logic
•constraints
•explicit reasoning

[Angelo Dalli @AngeloDalli]

@Grady_Booch @JimDMiller It is hard not to be judgmental when seeing such behaviour. Imagine a doctor telling you: oh do you expect me to actually research the side effects of all medicine I am prescribing you? The dumbification of research is a real concern.

[Angelo Dalli @AngeloDalli]

Most AI “reasoning” isn’t reasoning at all.
It’s just statistical compression. Here’s how real decision-making actually works in reliable systems

1. Identify
Spot the situation and pull out the key signals.

2. Assess
Analyze risks, constraints, trade-offs, and options.

3. Resolve
Choose an action — and clearly justify it.

Today’s LLMs try to mash all three steps into ONE forward pass.
It feels magical.
But it’s a black box.

Speed? Massive win.

Accountability, traceability & safety? Big loss.

The strongest AI systems today (advanced agents, thinking-style models, production-grade setups) keep these stages deliberately separate. Result:
→ Debuggable
→ Auditable
→ Trustworthy

This is what separates toy demos from systems you can actually bet your company (or life) on.

Do you want your AI to be fast… or actually reliable?  Quote RT with your take #AI #Reasoning #AgenticAI #xAI

[Angelo Dalli @AngeloDalli]

I see it more like your kludge idea than an actual neurosymbolic AI system. There are huge complex switch statements in the code, which is architecturally awful (guaranteed to give @Grady_Booch a headache). Bet you’re going to see more of these neurosymbolic systems in action soon :)

[Angelo Dalli @AngeloDalli]

So the real question:

Is test-time compute + process supervision enough… or do we need a deeper architectural shift?

What do you think?

[Angelo Dalli @AngeloDalli]

The uncomfortable truth about today’s AI:

Most models jump straight from input → output.

No assessment phase.
No guaranteed progress.
No real reasoning.

They’re just extremely good predictors.

[Angelo Dalli @AngeloDalli]

This is why we still see confident hallucinations, brittle agents, and systems that sound smart but fail when it matters.

We’re optimizing for fluency, not reliability.

[Angelo Dalli @AngeloDalli]

They compress:
•detection
•reasoning
•action

into one probabilistic process.

Structured AI systems separate these stages explicitly. That separation is critical for accountability and control.

[Angelo Dalli @AngeloDalli]

The first step in intelligence is not reasoning.
It is identification.

[Angelo Dalli @AngeloDalli]

Poor identification guarantees poor decisions.

This is one reason many AI systems fail in dynamic environments.