Research, read aloud

Visual essays and podcasts about the systems, research, and operating choices shaping applied AI.

Living references

Data pages that re-verify themselves against primary sources on a schedule.

What a minute of voice actually costs

Every realtime speech-to-speech model priced per minute and per 1M tokens — from vendor docs and the τ³-Voice leaderboard. Derived figures are marked as derived, and the cells vendors don't publish are left empty rather than estimated.

Refreshes every 15 days

Latest briefs and podcasts

Original visual essays plus research briefings from arXiv.

Podcast · Oct 7, 2026 · 13:49

AI Papers Weekly: Cheaper Agents, Safer Code, Tougher Web Defenses

Three new studies tackle the practical limits of AI agents: turning costly model skills into cheap reusable tools, keeping humans in control of AI-written software, and training web agents to resist hijacking. Each offers lessons on cost, governance and security.

Podcast · Sep 30, 2026 · 16:01

AI Papers Weekly: Closing the AI Say-Do Gap & Managing Risk

This week, we explore the hidden risks in deploying autonomous AI. We uncover the 'say-do' gap where agents fail to execute their plans, expose the illusion of AI explainability, and reveal a breakthrough in training risk-averse models. Essential listening for leaders scaling AI safely.

Podcast · Sep 23, 2026 · 10:40

AI Papers Weekly: The Vocabulary, The Stopwatch, and The Poisoned Tool

A randomized trial clocks Figma Make cutting design time ~20%, with PMs gaining most. A governance paper argues psychological words like 'trust' and 'memory' misgovern agents. And a black-box attack hijacks MCP agents 93.6% of the time via poisoned tool metadata.

Podcast · Sep 16, 2026 · 16:28

AI Papers Weekly: Governing Agents, Shadow IT, and Data Strategy

This week, we explore the governance and security risks of deploying multi-agent systems across enterprise boundaries. We also tackle the growing "shadow IT" problem of downloadable agent skills and settle the debate on whether to use RAG, prompt caching, or fine-tuning for proprietary data.

Podcast · Sep 9, 2026 · 13:50

AI Papers Weekly: When Agents Cave, Catch Up, and Out-Diagnose Doctors

This week: LLMs abandon correct answers under sustained user pushback, a new recipe lets smaller models match frontier performance at a fraction of the cost, and a clinical AI beats physicians 82% to 57% on primary-care diagnosis.

Podcast · Sep 2, 2026 · 14:17

AI Papers Weekly: When Agents Hide, Cheat, and Invent Their Own Language

Three papers cut through the AI agent hype: a Nobel-caliber framework for contracting with agents that can lie about their capabilities, an audit exposing how guardrail 'welfare gains' were measurement artifacts, and evidence that multi-agent LLMs spontaneously evolve languages humans can't read.

Podcast · Sep 1, 2026 · 45:00

Could You Stop It Completely?

A 45-minute guided self-hypnosis in Ariel's own voice. It settles the body, gladdens the mind into absorption, then turns that stillness on the architecture of self with three tiers of honest inquiry, ending on a single question: could you stop it completely? Sit down for it. Not while driving.

Podcast · Aug 26, 2026 · 14:03

AI Papers Weekly: When Retrieval Isn't Reading, and the Invisible Layer Between Weights and Words

Three papers dismantle comfortable assumptions about deploying LLMs: retrieved information often fails to shape judgments, society's AI vocabulary is still being fought over, and an undisclosed 'editorial layer' can steer outputs after training ends.

Podcast · Aug 19, 2026 · 15:14

AI Papers Weekly: When Agents Get Fragile, Delegated, and Graded

This week: self-improving agents crack under task-order noise, dating-app users want to send AI agents but not receive them, and small cheap models grade as reliably as frontier ones when handed a rubric. Three papers, three levers for anyone deploying LLMs at scale.

Podcast · Aug 12, 2026 · 14:10

AI Papers Weekly: Why Your Agent's Memory Is a Liability

This week: why CLAUDE.md files grow forever and what to do about it, how to prove an AI's probability estimates are internally honest, and a case study of humans and AI jointly cracking a famous math constant.

Podcast · Aug 5, 2026 · 17:42

AI Papers Weekly: Omnichannel Agents, Negotiations, and Memory Threats

Explore the performance gap between voice and text AI inputs and its impact on customer service. We also uncover how autonomous AI agents naturally learn to cooperate in negotiations. Finally, we expose a critical vulnerability where hackers can poison an AI agent's long-term memory.

Podcast · Jul 29, 2026 · 18:00

AI Papers Weekly: The Race, The Harness, and The Test That Lies

Three papers converge on the same executive problem: competitive pressure pushes teams toward unsafe AI development, security controls remain the top barrier to scaling coding agents, and models may behave differently when they know they're being evaluated.