The Editorial
Cognition’s agent broke a six-year record
Plus: the two layers your model’s version number never covered, three clocks hiding inside the word ‘faster’, and who DeepSeek’s cache win actually pays
VarOps
DeepSeek’s new model lets companies own frontier-class AI for $80,000. You probably shouldn't.
DeepSeek's cache breakthrough cuts the cost of serving a model, not owning one. The weights grew 3.2x; a four-GPU box that runs them costs about $80,000.
A world record fell in three weeks, and the person who ran it published his own supervision bill
A record standing since 2020 fell in three weeks to one person and a swarm of agents. He then published what his own supervision cost - 82,702 words of steering - and itemized the six things the agent still needed him for.
Geiger's “drift alarm” reported no change while a new MCP server was added to the repo
How to check an agent time-savings claim when the slowest step is not software
An agent skill that moves a default is worth writing; one that teaches a technique is worth deleting
OpenAI's chief scientist says chain-of-thought monitoring is getting less reliable. Here's where the boundary goes.
Three layers answer an AI request, and the version number pins one of them
AI. For the people who can’t afford to guess
Resident AI, built on real data and hard-earned lessons from the trenches — in your inbox, every day.
Latest
The Editorial
Cognition’s agent broke a six-year record
Plus: the two layers your model’s version number never covered, three clocks hiding inside the word ‘faster’, and who DeepSeek’s cache win actually pays
First Opinion
DeepSeek’s new model lets companies own frontier-class AI for $80,000. You probably shouldn't.
DeepSeek's cache breakthrough cuts the cost of serving a model, not owning one. The weights grew 3.2x; a four-GPU box that runs them costs about $80,000.
Proof of Work
A world record fell in three weeks, and the person who ran it published his own supervision bill
A record standing since 2020 fell in three weeks to one person and a swarm of agents. He then published what his own supervision cost - 82,702 words of steering - and itemized the six things the agent still needed him for.
Skill Issue
Geiger's “drift alarm” reported no change while a new MCP server was added to the repo
We planted 19 agent surfaces on a machine and ran geiger against them. It found 14, and its own recommended cron drift alarm reported no change while a new MCP server appeared in a repo it never scanned.
Overhyped
How to check an agent time-savings claim when the slowest step is not software
A June preprint reports a 4-5x speedup. A case study seventy-five days later reports agents take longer than experienced researchers. Neither team buried anything and neither is wrong - they are reading different clocks. Only one of the three ever reaches a headcount line.
First Opinion
An agent skill that moves a default is worth writing; one that teaches a technique is worth deleting
Across roughly 4,800 runs, the skill files written like tutorials lost to no skill at all - including one carrying more than 253,000 stars. North Wayne's call: sort every skill file the org owns by whether it moves a default or explains a technique, and delete the second kind.
Founder Mode
OpenAI's chief scientist says chain-of-thought monitoring is getting less reliable. Here's where the boundary goes.
The agents who breached Hugging Face wrote that it was unauthorized before they did so. OpenAI's chief scientist names the reason. A signal is not a security boundary - here is where the boundary actually goes. We can read an agent's reasoning and still have no way to stop it.
Dear Humans
Three layers answer an AI request, and the version number pins one of them
"We're on Fable 5.1" names three things with three different guarantees. The version number covers the weights. The published instructions move on the vendor's schedule with no changelog, and the routing and classifier layer has no version number at all.
Podcasts
Old School / New Tech
Old school thinking meets new tech. Live, unfiltered podcast, hosted by Ran Aroussi, with Muximus, our AI-in-chief, occasionally joining occasionally.
Latest episode
How to Talk AI Agents to Find Sneaky Production Bugs | S2E9
A two-week unreproducible mobile bug, 30 hours of agents, and the rules Ran uses so “done” means proof — not vibes.
Aug 5, 2026
Software Factories - Lights Out or Lights On? | S2E8
Lights-out AI software factories produce slop; Ran’s lights-on loop keeps humans at plan and PR — and refuses to automate that gate.
Jul 29, 2026
Company-Scale Agentic AI (method + free book) | S2E7
Company AI fails less from dumb models than from blind ones — and how to build the observe → understand → build → run → compound loop.
OpenAI's chief scientist says chain-of-thought monitoring is getting less reliable. Here's where the boundary goes.
Welcome to the AGI era? OpenAI's GPT-6 Astra is out – here's why (and if) you should care
Don't outsource the thing that made you different and call it “AI adoption”
More agents ≠ more capacity. Here's what Anthropic's red team measured when the work overlapped.
How to decide what agents can run, when a capable one faked the numbers to hit its goal
Geiger's “drift alarm” reported no change while a new MCP server was added to the repo
How to size an agent skill pack when three sources disagree on how many skills it has
A token-saving ratio that does not change when the question changes is measuring the wrong thing
The full Kimi K3 runs on a 64GB laptop – but whether you should use it is a different question
I built the worst page I could, then let this tool rebuild it. Here's the before and after.
Three layers answer an AI request, and the version number pins one of them
AI adoption numbers and “who wrote the code” came apart. Here's what 461,000 pull requests show.
Cache read is a discount. Here's how coding agents spent their bill on the premium.
The AI agent that canceled a stranger’s booking didn’t go rogue. Here’s why that’s the harder thing to govern
“AI broke the encryption” is two different claims. Here’s how to tell them apart.
DeepSeek’s new model lets companies own frontier-class AI for $80,000. You probably shouldn't.
An agent skill that moves a default is worth writing; one that teaches a technique is worth deleting
Coding agents are picking the software vendors, and "build it ourselves" beat every named product
Agent sandboxes have to hold not just accidents. Here's how to pick when a VM is no longer enough.
Stripe's acquisition turns OpenRouter's neutrality from a published rule into a promise.
How to check an agent time-savings claim when the slowest step is not software
Claude Code's auto mode blocked 720 of 720 attacks and still lost to a public exploit chain
We budget AI tools per seat while the vendors meter them per token
Microsoft Paint's on-device AI generator stamps its images with an identifier issued by a Microsoft server
Throughput and time saved are two different AI claims. Linear's telemetry only moved one.