The Harness
A personal learning experiment with AI-generated audio
I use an AI agent to turn questions about AI systems into short audio conversations I can listen to in the morning. The topics follow what I’m trying to understand, from how agents remember information to how their performance gets tested.
I’m not an engineer, and part of my learning is getting comfortable with the infrastructure around the model. A useful explanation needs to unpack the terminology and give me an example I can work through.
Two synthetic hosts discuss one idea per episode and end with an open question.
How the learning develops
Sometimes I ask the agent a question; other times, I share a paper or article that interests me. It logs these in a backlog and keeps context about what I’m trying to understand.
After listening, I write back what I’m still confused about, including questions raised by other material I encounter. The agent reads those open questions before selecting stories, so an upcoming episode can return to something I haven’t understood yet.
The agent also researches topics around those interests. It handles audio production and prepares the reference notes for each episode.
Connecting the audio to Engram Wiki
The agent prepares a transcript and episode overview with source links, then stages them for a weekly wiki review. After approval, they become part of Engram Wiki, my personal knowledge system and another project on this site.
I can revisit those records when I return to a subject, or have a different AI assistant read them as context.
One episode explored the agent’s own memory system. A later episode connected it to a paper about moving information and procedures outside the model, giving me a familiar example to work from.
The files behind the learning
The context lives in Markdown files I can inspect and update. Here’s how the working files connect to the wiki.
Agent workspace
- learning/
open-questions.mdWhat I'm still trying to understand; read before story selection.
- Episode planning
requested-topics-backlog.mdQuestions, papers, and articles queued for future episodes.episode-log.mdWhat has aired, with listening links and sources.
- Each episode
script.txtThe dialogue used to generate the audio.- Weekly wiki stagingPrepared transcript and index entry awaiting the weekly review.
Engram Wiki
- topics/
the-harness.mdThe show's purpose and format.the-harness-episodes.mdEpisode overviews, source links, and paths to transcripts.
- exports/the-harness/transcripts/
- One Markdown transcript per episodeThe conversation preserved for later reading and reference.
Prepared records reach the wiki through a weekly batch review and approval.
Selected excerpts
Five excerpts · 3 minutes 32 seconds in total
01 Episode 7 · October 13, 2026 1:16
Why a shopping list changes the task
A familiar example introduces how memory and reusable instructions can change what an AI agent has to do.
Read the excerpt
Jordan: The paper's one big idea comes from the design researcher Donald Norman: a good external aid doesn't make you smarter. It changes the task into one you can already do. A shopping list doesn't improve your memory. It turns a hard remembering problem into an easy recognizing problem.
Alex: So apply that to agents.
Jordan: That's exactly what the paper does. An unaided language model faces three chronic problems, each solved not by a better model but by moving the burden out of the model and into infrastructure. First, memory. The context window is finite and everything vanishes between sessions, so teams built persistent memory stores — and recalling the past became retrieving it, the shopping-list move. Second, skills. A model re-derives procedures every time, slightly differently each time, so teams packaged procedures into reusable skill documents — improvising every step became composing pre-tested parts. Third, protocols. Tool calls negotiated through free-form prompting are brittle, so teams wrote explicit contracts for how tools get discovered and invoked — fragile negotiation became structured exchange. Memory, skills, protocols — state, expertise, and interaction, each moved outward.
The source is a review of agent infrastructure.
02 claude.ai performance sprint · October 6, 2026 0:28
Making claude.ai 3× faster
An episode I requested about Anthropic's performance sprint, focusing on how engineers directed and reviewed Claude's work.
Read the excerpt
Alex: Okay, but here's the part that hooked me. They didn't do it alone. They ran the whole sprint out of a single Slack channel, with Claude in every thread. Claude found the bottlenecks, built the benchmarks, shipped the improvements, and watched every deploy. The humans steered — they set goals, made tradeoffs, approved every change. But Claude did a huge share of the actual work. More than three thousand changes merged in two weeks, and not a single customer-facing incident or rollback.
Read how Anthropic made claude.ai faster
Anthropic's account of its own two-week sprint.
03 Episode 6 · October 13, 2026 0:40
Choosing what to test
A coastline analogy explains how selecting examples from one model's failures can give a misleading picture of performance.
Read the excerpt
Alex: And then there's the insidious one: how you pick the hard cases.
Jordan: This is my favorite part. Model capability is jagged — picture a coastline, not a flat beach. If you pick hard cases from wherever today's model happens to fail, you're sampling the valleys of one model's coastline. You end up measuring that model's failure fingerprint, not what's actually hard or valuable.
Alex: So how should you pick them?
Jordan: Because a person judged them hard. You should be able to say why a task is hard before you let it in. Production failures and bug reports are good sources. But even user traffic can mislead — people try what they expect to work, so a sample drawn straight from traffic can skew easy.
04 AIR-LLM · October 8, 2026 0:31
AIR-LLM: model weights over radio
I was curious about AIR-LLM's proposal to receive model weights over radio signals. This excerpt separates the physical measurements from the simulated wireless conditions.
Read the excerpt
Alex: Alright, but does it actually work, or is this a whiteboard fantasy? What did they measure?
Jordan: Good question. Let's separate what's verified from what's still simulation. They tested it two ways. First, on real hardware: they took an actual RF mixer, measured it on the bench, and ran the computations through it. Second, the wireless channel — the air between the tower and the phone — was simulated, using ray-traced models of real streets in Paris and Munich. So: real mixer, simulated airwaves.
The paper combines wired RF-mixer measurements with simulated wireless channels.
05 OpenAI CFO · October 6, 2026 0:36
Starting with the decision
OpenAI's account of its finance experiments led to a discussion of the decision a workflow is meant to support.
Read the excerpt
Jordan: Her core move was to stop asking how to make each finance task faster, and start asking which decision the task exists to serve. Begin with a consequential decision, map the data, tools, approvals, and handoffs required to support it, then work backward to figure out which parts the AI can analyze, coordinate, or complete.
Alex: That's the big idea. She says the real promise isn't closing the books faster. It's giving finance teams more time to shape what happens next, while the outcome can still change. It's about decision latency, the gap between something changing in the business and a leader being able to act on it.
The source describes OpenAI's own work and ambitions.
These excerpts use AI-generated voices and scripts. They are selected from my personal learning feed, with transcripts and links to the original material.