what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

How I AI Mental Models, Failure Modes, Long Horizon Workflows

presenter: Mark Ruvald Pedersen IDA: Driving IT 2026 date: Friday 30 October 2026 blog: blog.radix63.dk GitHub: eisbaw LinkedIn: ruvald
Sketch: a person at a bus stop holding a map; the bus route on the sign is a straight line of stops, while the real route winds through the city. All models are wrong, but some are useful. George Box.
Image: Jono Hey, Sketchplanations · CC BY-NC 4.0
what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

whoami

mpedersen

  • Senior Firmware Engineer, NVIDIA · Copenhagen
  • Electronics to compilers: firmware, real-time, algorithms, tooling
  • Concepts first: decomposition, architecture, theory
  • Long-time NixOS ❄️ user
  • Husband, father, evening hacker, often with AI

Mark Ruvald Pedersen in sunglasses, his son blurred on his back

Opinions are my own, not those of my employer.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

AI Usage Levels

Five ways to use AI: conversation, editor assistance, delegated engineering, backlog-driven execution with intermittent babysitting, and unprompted loops for autonomous work and maintenance

Start small and build up: learn to walk, then run, then scale.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Neural Networks as Universal Function Approximators

Functions describe the worldPato Molina · videoyoutube.com/watch?v=BWZTlfrneD8

Functions describe the world · Pato Molina ↗

Composition: x through g then f to y, small swappable parts, marked with a green check

Latent space visualized with t-SNE: training and testing data fall into clusters, the features the network has learned

Latent space, t-SNE · Source: GeeksforGeeks

  • Composition at massive scale

Composition: greater than the sum of its parts. Lego bricks compose; a moulded toy does not.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Training Is Compression

  • Compression is pattern discovery
  • Patterns become weights: islands of latent space
  • Your EXACT tools and your domain are not in the weights

Training data compressed into weights, with a brace zooming into a pattern pyramid from syntax through semantics and intent to domain

You’re talking with ./humanity.exe. Be specific about role, context and intent.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Layers from Pressure

  • MPED conjecture: anything with enough complexity and selection pressure separates into layers of abstraction
  • Gravity shapes rock; training loss shapes representations

Earth crust layers compared with neural network abstractions

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Think in Distributions of Outcomes

  • The same prompt gives different answers
  • Every answer is one branch of many
  • Context steers the distribution

Many-worlds-style branching as a metaphor for possible token continuations, with one emitted path highlighted and a Fusion comparison below

Curate the context: you are rolling dice within that distribution.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Open Loop vs Closed Loop: Vibe Coders vs Engineers

Free climbing versus climbing with a safety line and feedback

“Give Claude a way to verify its work… it will 2-3x the quality of the final result.”Boris Cherny

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Failure Modes

  • Too much or too little context: it cannot read your mind
  • Hallucination: plausible nonsense. Falsify early
  • Overconfidence: fluent ≠ correct
  • Context poisoning: early errors propagate
  • Stale false memories: a MEMORY.md outside the repo
  • Tool friction: confusing cruft builds up

An early incorrect fact poisons downstream context like falling dominoes

The calculator principle: deterministic work goes to deterministic tools.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

The Context Window Is a Soup

Exploration is expensive.
Distillation is cheap.
Keep context clean.

Context as a soup of instructions, tools, history, and task information

One task per context window. Rewind, don’t correct.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Agents Onboard Fresh Every Session

  • No memory between sessions: a newly hired professor every morning
  • Long-term: PRD, design, plan, conventions. Short-term: the task and how to verify it
  • Memory lives in the repo: reviewed, versioned, shared

A context window behaves like volatile RAM between sessions

PRD = destination · design = vehicle · plan = route map

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Codify Your {Methodology, Values, Philosophy, Ethos, Wisdom}

  • Role-cast the AI as your mini-me: values and invariants, not snake_case
  • “We do not block on the main loop thread”
  • Every instruction costs reliability: prune

One overloaded agent compared with calm focused agents, one concern each, branched and joined

Agents have focus and bias; skills are objective and transferable.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

A stdlib of Skills

  • Workflows compose skills like UNIX pipes
  • Sandwiches: deterministic steps where it matters, judgement where it helps
Neo in The Matrix, just after a skill was loaded into his head: I know kung fu
“I know kung fu.” · The Matrix (1999)

Four layers: tool-usage primitives, experts with breadcrumb indices into extracted PDFs, abstract workflows, and agents on top

Program the LLM’s behaviour with standards and conventions stored in your repo.after Geoffrey Huntley

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Sub-Agent Telescoping and Fork-Join

A parent agent forks focused sub-agents and joins their distilled conclusions

  • Divide and conquer
  • Fork a task, join a conclusion
  • A hundred fresh professors need supervision

If a human is in the inner loop, you will be overloaded.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Unreasonable Amounts of Testing

  • Falsify first, then ground in real data and hardware
  • Cheap and fast → adversarial → closest to reality
  • You get what you measure

A base plate, falsify first then ground, under a staircase of seven ever taller steps: TDD, fuzzing, property, mutation, simulation, formal, real end-to-end; grouped as cheap and fast, adversarial, expensive and true

Track hypotheses with hyp: a criterion and evidence for every suspicion · github.com/eisbaw/hyp

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

The Ralph Loop: Inner and Outer Loops

Repeated fresh agent sessions use a persistent workspace and backlog

A dashed orchestrator context spawns implementation and review/test/fix subagents below it, then receives their conclusions; human and CI feedback forms the outer loop

  • A long-lived orchestrator, many small implementer sub-agents
  • Inner loop: agents test and fix
  • Outer loop: human and CI try it
what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Ask for Visuals, Not Walls of Text

  • Prose takes time to read and to falsify
  • Human visual perception is a high-bandwidth channel
  • Ask for SVG figures, animations, explainer videos

A wall of text, slow to read, beside the same content as a drawn figure and an explainer video, structure at a glance

A wrong arrow in a figure is spotted in seconds, by human and by AI.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

AI Is an Amplifier

  • Not a text box: an army of fast, fallible synthetic humans
  • We are the shepherds of fallible AI: design the guardrails
  • AI is work: like a guitar, it takes practice
  • Review the process: let AI reflect on itself and mine its sessions
Quality ∝Human × Prompt × Model × Harness × WorkflowTool friction

An engineer shepherds a team of fast but fallible AI sled dogs

AI scales the good and the bad; bottlenecks become clear.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Tools That Cut the Friction

gitgit revertevery step a commit; any damage is a revert away Nixnix developone pinned toolchain for you, the agent and CI justjust testnamed recipes: the agent runs what you run Rustcargo clippya strict compiler: errors are fast negative feedback Backlog.mdbacklog task listtasks, acceptance criteria and notes, in the repo hyphyp inita criterion and evidence for every suspicion

Tool friction divides quality: give agents CLIs that fail loudly.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Coding Is No Longer the Bottleneck

  • 8× the code ships only 30% faster (Voss)
  • The constraint moves: review, then deciding what to build
  • Programming has always been communication

A flow from intent to shipped software: coding is a wide pipe now, refinement and review are the narrow ones

“Removing waste from your organization and processes is a bigger accelerator than AI itself.”Geoffrey Huntley

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

What Is Worth Building Moved

  • Months became days
  • De-risk early: small demos first, compose later
  • Verification is the bottleneck

An axis of cost to build from hours to person-years; the worth-it line moved from a week to past person-years, with a tool for one, a board twin and a Lean prototype in the opened zone; below, Dan Luu's rule, ambitious about what you attempt, conservative about what you trust

Be ambitious about what you attempt, conservative about what you trust.after Dan Luu

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

IKEA Software and the Skills of the Future

IKEA software against artisanal software on four scales: time to build, cost, how long it lasts, whether it solves today's problem

Canned wisdom: a shelf of tins labelled with schools of thought, one poured into a skill file and handed to an agent

  • IKEA software: cheap, fast, good enough
  • Schools of thought: codified, transferable

Cheap code makes its complements dear: verification, judgement, taste, trust.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Greenfield and Brownfield, Built With AI

Greenfield

Brownfield

Linux ath11kWi-Fi memory amplification, patched NanodaLean 4 proof checking, 12% faster mosh-tcpTCP transport where UDP is blocked NFS-Ganeshadirectory-entry decoding fixed Renodespurious I²C warnings fixed

A subset of private hobby projects, from the last year.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Case: How My Team Does Backlog Refinement Now

  • AI collects the state of every task, read-only
  • AI proposes the agenda; the team decides in the meeting
  • AI enacts the decisions and commits the record

Collect read-only, propose an agenda page, the team meets and decides and the meeting transcript carries the guidance back, AI enacts the updates; the dated page is committed

AI proposes the order, people decide, AI does the bookkeeping.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Case: GACIA, Code Improved While We Sleep

  • Nobody prompts tech debt: dead code, stale docs, flaky tests
  • Perpetual tasks: a YAML file, a cron trigger, a prompt
  • Runs at night on idle CI; small merge requests wait in the morning
  • Green pipeline, reviewers assigned: a human still says yes
GACIA explainervideoyoutube.com/watch?v=3evWKxtNulo

GACIA: GitLab Autonomous Code Improvement Agent ↗

The debt nobody schedules is the debt that compounds.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Monday Morning

Five steps as a staircase: ethos in CLAUDE.md, codify one skill, one closed loop, a Ralph loop for days, mine your sessions; then repeat

Ask not what AI can do for you, but what you can do for AI.

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Q&A

Two speech bubbles, a question and an answer, and the speaker's blog, GitHub and LinkedIn

This deck written in Marp · packaged by Nix Nix snowflake · made in concert with Claude, from my notes

References

what-ai-is.mdhow-it-fails.mdhow-i-work.mdwhat-it-changes.mdshowcase.mdmonday.md

Backup: Confine It, Then Let It Run

  • Damage is a DontCare: sandboxes, containers, VMs; commit and push
  • The biggest risk is leaking IP: no outbound network
  • Packages through a gateway that delays and quarantines: hallucinated package names get squatted

A sandbox holding the agent and a repo copy, outbound network crossed out, packages in through a delaying gateway, commits out to Git

Make errors unlikely, recoverable and confined.

The terminal character grid: ../tools/cover_grid.py writes it; assets/cc-chrome.js makes it flicker.