▲ Signal
The AI Engineering Skills Map for Knowledge Workers
Coding agents like Claude Code are the clearest clue to where knowledge work is heading. Listen from minute 12.
The reading log
I read obsessively about AI so you don't have to. Every piece here survived two cuts: it made my weekly round-up, and it was still worth your time when I refreshed this page. AI content ages fast — I prune this list monthly, so the shelf stays honest.
58 reads · updated Aug 26, 2026
Where are you on the curve?
The “if you only read three things” picks from recent round-ups.
Coding agents like Claude Code are the clearest clue to where knowledge work is heading. Listen from minute 12.
Benedict Evans on why the domain experts are rarely the people who can diagnose the problems worth solving. Enter the consultants.
Should your company own the model, harness, context and memory — or just buy a vertical solution?
LLMs are probability machines — a real problem for tasks that demand 100% precision.
I'd read summaries of this hack, but the actual play-by-play is straight out of a sci-fi movie.
AI is much more than a 20% productivity upgrade once you build self-learning agents.
Coding agents like Claude Code are the clearest clue to where knowledge work is heading. Listen from minute 12.
Benedict Evans on why the domain experts are rarely the people who can diagnose the problems worth solving. Enter the consultants.
Should your company own the model, harness, context and memory — or just buy a vertical solution?
LLMs are probability machines — a real problem for tasks that demand 100% precision.
I'd read summaries of this hack, but the actual play-by-play is straight out of a sci-fi movie.
AI is much more than a 20% productivity upgrade once you build self-learning agents.
A very cool interactive demo of a hedge-fund-specific research agent built on internal data — with guardrails.
Is it time to stop using skill-creator and write our skills by hand?
I was late to Claude Design, but it looks promising for anyone who lives in decks.
Ethan Mollick explains the difference between web chat and agentic systems like Cowork and Codex. The best on-ramp I know.
The second-order effects of lawyers working faster with tools like Harvey and Legora.
As your context grows, a domain expert has to draw the maps — ontologies — of your company's knowledge.
Stop limiting agents with babysitting — govern them with clear, enforceable law. A glimpse of AI-run operations.
AI detectors judge patterns, not meaning. What matters is whether the writer means what they say.
Ryan Carson runs his company on a fleet of agents — and still keeps pen and paper for the human parts.
Altman on non-consensus bets, killing good ideas to protect great ones, and why human connection gets more valuable.
Frontier intelligence is overkill for routine work. The smartest firms will learn to use less of it over time.
Ben Thompson's market take: bullish but not blind, with some fantastic metaphors on TSMC.
Income inequality was a huge problem pre-AI, and I expect the gap to widen. The backlash is here — and it's bipartisan.
Karpathy compresses AI systems engineering into one hour: from the LLM as a raw computer to message-passing agent graphs.
The bottleneck has moved: problem definition is now the critical skill.
Sloppy voice capture in, browsable personal wiki out — a full pipeline for a second brain that maintains itself.
I'm a huge Matt Van Horn fan, especially his focus on using CLIs to ship fast.
Zuck's 6,500 words on decentralized, open-source AI. He's going to compete on price.
A primer on why the coveted personal agent is really just a few tool calls and some memory.
Fluency doesn't drive success — problem framing and challenging assumptions do.
Bullish on land, power, shells and applications. Bearish on neoclouds and silicon.
SF engineers can't understand why normies don't have agents yet. Here are the reasons.
It's easier than ever to scale yourself. But should you?
Claude went on a tear in 2026, but ChatGPT and Codex are getting really good.
Jensen and 50 AI companies sign a pro-open-source letter. Note the one conspicuous absence.
A detailed explainer on wiring Slack, shared drives, and GitHub into one company brain everyone can query.
Opus and Fable need less handholding — which means it's time to simplify our skills.
If you only read one Fable article, this should be it.
The new models do better work when you stop micromanaging them.
A collection of prompts that actually exploit what the frontier models can do.
Ben Thompson on why Chinese open-source models threaten U.S. labs — and why America should foster its own.
The Man Group on building routers, connecting unstructured data, and what they now look for in new hires.
The bigger shortage isn't engineers — it's domain experts who understand AI.
The norms on recording are changing fast — apparently people now record their dates.
They have a strong foothold on your most important context. What comes next?
Managing LLMs is creating a lot of invisible labor.
Connectors aren't magic — they miss a lot of data in their searches.
The open management question of the decade: what does delegation look like when the worker never sleeps?
Agents need memory that humans can read, update, and inspect. Files are the natural fit — I've been building on this idea.
Coinbase's CEO went viral explaining how they manage token spend.
Companies are pushing back on vendor lock-in from Anthropic and OpenAI. Here's what the new stack might look like.
AI now takes control of your computer, and it's damn good. But how does it differ from browser use?
Brex's Pedro Franceschi on running the most AI-forward company around — and why he feels so strongly about the harness.
Ben Thompson's take on the Anthropic vs. U.S. government showdown.
Automate the review of your own work sessions and let it improve your tools every day. Template included.
Court reporters were supposed to be automated away. There's now a shortage.
Only 10% of companies see real savings. The fix: leadership treating data access and process design as their problem, not IT's.
Use a frontier model to write the instructions a small local model follows — teach procedures, not facts.
How students and new grads actually feel about entering an AI-shaped job market.
Box CEO Aaron Levie is both AI-pilled and realistic: permissions, messy data, and constantly-updating models slow everything down.
Why startups win by owning the whole workflow — data, governance and all — instead of wrapping the model.
AI spend is now a real budget line — and nobody can yet link tokens to business outcomes.