Dan Beksha

Private audio notes

Learning podcast

Lessons I create to understand AI, energy, and housing. Each starts with a question and works toward a useful explanation, with research and excerpts from people who know the subject. Episode notes are public; listening access is shared individually.

Request access ↗
How to listen in Apple Podcasts

Once approved, you’ll receive a private RSS link. In Apple Podcasts, open Library, choose Follow a Show by URL from the menu, and paste that link. Use your personal feed link, rather than this page’s address.

Start with a question

AI · Start with 01How does a language model learn anything? →Energy · Start with 08Why can’t new power plants just connect to the grid? →Housing · Start with 14Does building housing really lower rents? →

Jul 10, 202613:20

07 — AI Technical — Can an AI model stand in for a customer?

What synthetic consumers reproduce, what they miss, and why aggregates hide segment failures.

Available in the private feed.

Episode notes & sources

The impetus

The striking claim about synthetic purchase intent ↗

The dramatic ‘90% accuracy’ framing creates the real question: what did the study actually measure, and when can aggregate resemblance conceal segment-level failure?

Source audio in this episode

Evidence and further reading

Jul 9, 202615:26

06 — AI Technical — Why will AI costs feel more like cloud costs than software seats?

Why tokens create variable infrastructure spend—and what finance teams need to manage.

Available in the private feed.

Episode notes & sources

The impetus

Ramp’s emerging pattern in enterprise AI spend ↗

The pattern connects the cloud-cost analogy to a wider market shift: AI spending spreads into monitoring, integration, and use-specific infrastructure rather than staying a predictable software seat.

Source audio in this episode

Evidence and further reading

Jul 8, 202612:40

05 — AI Technical — How do you know an agent works?

Task evals, trajectory inspection, production monitoring, and business outcomes.

Available in the private feed.

Episode notes & sources

The impetus

Ara Kharazian on AI trust and observability spend ↗

A Ramp spending observation suggested that enterprise value may accrue around proving and monitoring performance, not only buying model access.

Source audio in this episode

Evidence and further reading

Jul 7, 202612:07

04 — AI Technical — What belongs in the model, the context, a skill, or a tool?

An architectural rule based on change rate and authority.

Available in the private feed.

Episode notes & sources

The impetus

Anthropic Skills and the emerging agent stack ↗

The architectural question underneath Skills is more durable than the announcement: what belongs in a reusable procedure, and what should remain in context, a model, or a tool?

Source audio in this episode

Evidence and further reading

Jul 6, 202613:27

03 — AI Technical — Why does Claude Code feel different from chat?

How repositories, tools, tests, and reversible diffs close the agent feedback loop.

Available in the private feed.

Episode notes & sources

The impetus

Danielle Morrill: make Claude Code your chief of staff ↗

The chief-of-staff framing raises the central problem: why does an agent become more capable when work is organized inside a repository with tools and feedback?

Source audio in this episode

Boris Cherny on why Claude Code began in the terminal ↗

Boris Cherny, speaking with Dan Shipper on AI & I. 10 seconds, about 3:03 into this episode.

More source audio

Evidence and further reading

Jul 5, 202614:41

02 — AI Technical — What is an AI agent, actually?

The model, tools, state, permissions, feedback, stopping rules, and ownership inside an agentic loop.

Available in the private feed.

Episode notes & sources

The impetus

Aaron Levie on the scarcity of people who can deploy agents ↗

The hard part is not saying ‘agent’; it is understanding the loop well enough to deploy one inside a real organization.

Source audio in this episode

Evidence and further reading

Jul 4, 202615:09

01 — AI Technical — How does a language model learn anything?

Tokens, prediction error, parameters, gradient descent, attention, post-training, and inference.

Available in the private feed.

Episode notes & sources

The impetus

A rigorous, accessible route into how ChatGPT works ↗

Karpathy’s from-scratch approach suggests that language-model training can be explained without treating the system as either a database or a mysterious mind.

Source audio in this episode

Evidence and further reading