Hey! I'd love to hear your thoughts, send me a voice note.

AI can now write code, solve math problems, analyze research, plan multi-step tasks, and produce answers that sound remarkably confident.

But intelligence creates a strange problem:

What happens when producing an answer becomes easier than verifying whether the answer is correct?

A model can generate thousands of lines of code in minutes.

A human may need hours to inspect them.

An AI can propose a mathematical proof almost instantly.

Checking every step may be harder than creating the proof.

An autonomous agent might perform hundreds of actions before a person ever sees the final result.

So the next frontier of AI may not simply be building systems capable of generating better answers.

It may be building systems capable of proving when those answers deserve to be trusted.

Central question

If AI eventually becomes better than humans at solving difficult problems, who or what will be capable of checking its work?

That gives us a fantastic episode spanning AI reasoning, hallucinations, uncertainty, reinforcement learning, formal verification, code execution, scientific replication, AI judges, and the emerging idea of scalable oversight.

Hey! I'd love to hear your thoughts, send me a voice note.

Artificial intelligence can explain quantum physics, write software, analyze thousands of documents, and reason through complex problems.

Then you start a new conversation...

and it may not remember what happened yesterday.

So what does it actually mean for an AI to remember?

In Episode 10 of Zero to Singularity, we explore one of the most important problems standing between today’s AI assistants and truly persistent intelligent agents: long-term memory.

We break down the difference between a model’s learned parameters and its context window, why a huge context window is not the same as permanent memory, and how modern AI systems use retrieval, embeddings, vector databases, summaries, and external memory stores to preserve information across time.

Then we go deeper.

How should an AI decide what is worth remembering? How does it know when a memory is outdated? What happens when two memories contradict each other? Can an AI learn from experience without retraining its entire neural network? And how can it keep learning without suffering catastrophic forgetting gaining new knowledge while damaging what it already knows?

We explore:

context windows, working memory, episodic memory, semantic memory, retrieval-augmented generation, vector databases, agent memory, continual learning, catastrophic forgetting, memory compression, personalization, privacy, memory poisoning, and the possibility of AI systems that accumulate experience over months or even years.

Hey! I'd love to hear your thoughts, send me a voice note.

Modern AI was built on an extraordinary resource: human knowledge.

Books. Websites. Research papers. Code. Images. Video. Conversations. Billions upon billions of examples created by people.

But what happens when frontier models have already consumed most of the useful, accessible human-generated data?

In Episode 9 of Zero to Singularity, we investigate one of the biggest questions facing the next generation of artificial intelligence: Can AI keep improving if high-quality human training data becomes a bottleneck?

This episode explores:

• how pretraining data powers modern AI
• scaling laws and diminishing returns
• the difference between more data and better data
• synthetic training data
• self-play and automated curriculum generation
• reinforcement learning and verifiable tasks
• model-generated reasoning examples
• simulation and virtual environments
• multimodal data from video, audio, robotics, and sensors
• proprietary and expert-generated datasets
• inference-time scaling and test-time compute
• retrieval, tools, agents, and external memory
• model collapse and synthetic-data contamination
• why AI-generated internet content could become a training problem
• whether machines can eventually generate their own useful learning experiences

We also examine some of the biggest claims surrounding the future of AI training:

Is the internet running out of useful data? Can synthetic data replace human-created knowledge? Does training on AI-generated content inevitably cause model collapse? Can self-play generate effectively unlimited training material? And could future AI systems eventually design their own increasingly difficult curriculum?

The deeper question is no longer simply:

How much human knowledge can AI absorb?

It may become:

Can intelligence eventually create the experiences it needs to make itself smarter?