The Neuron: AI Explained

AI reasoning models don’t just give answers — they plan, deliberate, and sometimes try to cheat.

In this episode of The Neuron, we’re joined by Bowen Baker, Research Scientist at OpenAI, to explore whether we can monitor AI reasoning before things go wrong — and why that transparency may not last forever.

Bowen walks us through real examples of AI reward hacking, explains why monitoring chain-of-thought is often more effective than checking outputs, and introduces the idea of a “monitorability tax” — trading raw performance for safety and transparency.

We also cover:
Why smaller models thinking longer can be safer than bigger models

How AI systems learn to hide misbehavior

Why suppressing “bad thoughts” can backfire

The limits of chain-of-thought monitoring

Bowen’s personal view on open-source AI and safety risks

If you care about how AI actually works — and what could go wrong — this conversation is essential.

Resources:
Title URL
Evaluating chain-of-thought monitorability | OpenAI https://openai.com/index/evaluating-chain-of-thought-monitorability/
Understanding neural networks through sparse circuits | OpenAI https://openai.com/index/understanding-neural-networks-through-sparse-circuits/
OpenAI's alignment blog: https://alignment.openai.com/
👉 Subscribe for more interviews with the people building AI
👉 Join the newsletter at https://theneuron.ai

Everyone is rushing to build AI agents — but most companies are setting themselves up for failure.

In this episode of The Neuron, Darin Patterson, VP of Market Strategy at Make, explains why agentic AI only works if your automation foundation is solid first. We break down when to use deterministic workflows vs AI agents, how to avoid fragile automation sprawl, and why visibility into your entire automation landscape is now mission-critical.

You’ll see real examples of building agents in Make, how Model Context Protocol (MCP) fits into modern workflows, and why orchestration — not hype — is the real unlock for scaling AI safely inside organizations.

Subscribe to The Neuron newsletter for more interviews with the leaders shaping the future of work and AI: https://theneuron.ai

IBM just released Granite 4.0, a new family of open language models designed to be fast, memory-efficient, and enterprise-ready — and it represents a very different philosophy from today’s frontier AI race.

In this episode of The Neuron, IBM Research’s David Cox joins us to unpack why IBM treats AI models as tools rather than entities, how hybrid architectures dramatically reduce memory and cost, and why openness, transparency, and external audits matter more than ever for real-world deployment.

We dive into long-context efficiency, agent safety, LoRA adapters, on-device AI, voice interfaces, and why the future of AI may look a lot more boring — in the best possible way.

If you’re building AI systems for production, agents, or enterprise workflows, this conversation is required listening.

Subscribe to The Neuron newsletter for more interviews with the leaders shaping the future of work and AI: https://theneuron.ai

Imagine an AI that doesn’t just output answers — it remembers, adapts, and reasons over time like a living system. In this episode of The Neuron, Corey Noles and Grant Harvey sit down with Zuzanna Stamirowska, CEO & Cofounder of Pathway, to break down the world’s first post-Transformer frontier model: BDH — the Dragon Hatchling architecture.

Zuzanna explains why current language models are stuck in a “Groundhog Day” loop — waking up with no memory — and how Pathway’s architecture introduces true temporal reasoning and continual learning.

We explore:
• Why Transformers lack real memory and time awareness
• How BDH uses brain-like neurons, synapses, and emergent structure
• How models can “get bored,” adapt, and strengthen connections
• Why Pathway sees reasoning — not language — as the core of intelligence
• How BDH enables infinite context, live learning, and interpretability
• Why gluing two trained models together actually works in BDH
• The path to AGI through generalization, not scaling
• Real-world early adopters (Formula 1, NATO, French Postal Service)
• Safety, reversibility, checkpointing, and building predictable behavior
• Why this architecture could power the next era of scientific innovation

From brain-inspired message passing to emergent neural structures that literally appear during training, this is one of the most ambitious rethinks of AI architecture since Transformers themselves.

If you want a window into what comes after LLMs, this interview is essential.

Subscribe to The Neuron newsletter for more interviews with the leaders shaping the future of work and AI: https://theneuron.ai

Carina Hong dropped out of Stanford's PhD program to build "mathematical superintelligence" — and just raised $64M to do it. In this episode, we explore what that actually means: an AI that doesn't just solve math problems but discovers new theorems, proves them formally, and gets smarter with each iteration. Carina explains how her team solved a 130-year-old problem about Lyapunov functions, disproved a 30-year-old graph theory conjecture, and why math is the secret "bedrock" for everything from chip design to quant trading to coding agents. We also discuss the fascinating connections between neuroscience, AI, and mathematics.

Lean more about Axiom: https://axiommath.ai/ 

Subscribe to The Neuron newsletter: https://theneuron.ai