Bert, Ernie, and the imagination of AI
2026-03-05
Loading blurb...
2026
2026-03-05
Loading blurb...
2026-03-02
Loading blurb...
2026-02-27
Loading blurb...
2026-02-26
For the longest time, esoteric, impossible-to-use software hid in the shadows. Now, when words are the interface, their time to shine has arrived.
For example, G'MIC. Agents, skills, and tools speak its language.
2026-02-26
I've created this little place to share bite sized thoughts that hopefully speak for themselves
2026-02-26
Loading blurb...
2026-02-24
People label things impossible when the real blocker is inertia.
Page 2 of 2
Are you an Ernie or a Bert? What kind of imagination are you developing?
In the popular television show Sesame Street, there was a skit where Ernie wakes Bert to tell him he can't sleep. He's scared of shadows, spooky things, and monsters.
Bert reminds Ernie that they are both friends with other furry puppet monsters like Grover, Herry, and Cookie Monster. He goes on to say, "you're imagining all those scary things, but you can imagine nice and good things too."
Bert then asks Ernie to name things he likes. Ernie responds with "Big balloons." "What else?" "Little balloons." "There must be something else." "Sure, medium-sized balloons."
Do you have an overdeveloped imagination of scary things? Are you limiting your imagination of good things?
The one thing that would prevent the "Cambrian Explosion" of agents and workflows I predicted is cost.
For that explosion to really take off, access to knowledge and tools has to be in more people's hands.
A back-of-the-napkin calculation suggests the current (March 2026) upper limit of people who could use harnesses to make LLMs and tools do interesting things is tens of millions globally. This estimate is based on paid subscriptions to ChatGPT or Anthropic, or access to Kimi, DeepSeek, Qwen, et al.
But if you want to do truly interesting things (the kind that require wild YOLO agent runs), the daily included tokens on entry-level accounts run out quickly. Then your choices are to upgrade to $200-plus-per-month accounts or use pay-per-token APIs, which can easily run into the tens of dollars for single jobs.
The other option is beefy on-prem hardware to run open models, but that comes with hardware, skill, and maintenance costs.
This exposes a different compounding effect, the power law, where the most interesting autonomous agent evolution concentrates and compounds faster among those who have ideas, knowledge, and money. "
What if the strongest signal of the agentic era isn't buzz about Openclaw, but a pricing page rewrite in an unrelated product? Remotion's move hints at exactly that.
Remotion is a tool that turns code into videos using React.
Early this month (February 2026), they reworded their per render pricing from "Video Renders" to "Remotion for Automators".
This is the first time I've seen a pricing model specifically for agentic workflows. An agent running in a harness calling Remotion is an example "prompt-to-video app" their licensing page explicitly mentions as the target of their per-render pricing.
Remotion also released the Skill to help build high volume video production into your app in minutes.
Don't worry though, individuals and small teams can render as many videos as they'd like for free.
Your word of the day is harness: a device to capture compute and artificial brain power.
I was struggling to find the right term for the standard shape of the modern agent stack: agent loop + LLM + tools + skills. It keeps showing up in products like OpenCode, Claude Code, Codex, and OpenClaw. Turns out people are converging on harness, and it fits.
A harness captures horsepower, so why not use it for machine horsepower too. I expect a huge proliferation of niche tooling with purpose-built harnesses under the hood.