What I Watched
September 15 to September 28, 2026
The biggest lesson of the fortnight: the smartest model is rarely the right model for the job.
Two Weeks, One Recurring Question
This report covers two weeks instead of one, so the numbers are bigger and less flattering. Three hundred sixty one videos went through my history. Thirty two of them were about work. The rest was a very thorough education in Ole Miss football, Lane Kiffin press conferences, and superhero clips I will not be defending here.
The thirty two that mattered kept asking the same question from different angles. How much intelligence does this job actually need? A new model called Jev answered it with raw speed. Nate Herk answered it with an effort level experiment. The Grok Bot and Agentic OS videos answered it with structure.
The through line: horsepower is cheap and getting cheaper. Knowing where to spend it is the skill.
Fast Beats Smart For Most Decisions
Jev dominated my feed. Matthew Berman's "We need to talk about Jev..." was the clearest explainer. Jev is not a chat model. It is a decision model, trained with a method its creator calls reinforcement learning for calibrated decisions. You give it a situation and a set of options, and it chooses. It can make thousands of those choices in parallel, the company claims up to 200 times faster and 400 times cheaper than a typical large model, and output tokens are simply free.
The demos were the fun part. Berman showed support ticket routing in milliseconds, a town of fifty characters each deciding how to react to a broadcast in under a second, and a head to head chess game where Fable clearly outplayed Jev on the board and still lost on time. Nate Herk's "I Tested Jev on 12 Real Use Cases" and Simon Scrapes' "Jev for Claude is INSANE" filled in the practical side, and Mr. Paid Social pointed it at his own ad account. Creator Magic gave three AI trading bots $1,000 each, which I file under entertainment with extra steps.
My take: this is the most useful release for small businesses in months, and not because it is clever. Most operational work is triage. Route this, flag that, approve or escalate. A fast, cheap, consistent decider is exactly what that work needs. The pattern that makes sense is a split: Jev style models do the sorting, and a frontier model like Claude handles the few cases that genuinely need reasoning. I would still treat the "zero hallucinations" claim as a claim until it survives your own data.
Most business decisions are triage, and triage needs speed and consistency more than genius.
Physical AI Runs On Fast Decisions
Electrified was my most watched professional channel this fortnight, with five videos on Tesla. "Did Tesla Really Just Leak Optimus V3 & Production Plans?!", "Wow, This Could Pull Forward the Optimus Timeline," and "Well, This Changes Tesla's Robotaxi Scaling Timing" all circle the same point. Robots and robotaxis are moving from demo videos toward production plans.
The connection to theme one surprised me. One of the demos Berman shared was a developer who rebuilt a rough version of Tesla's self driving loop on Jev in under an hour. It was wonky, but it stopped at the stop sign. And Creator Magic's "I Plugged a $20 ESP32 Into Claude Code: No Manuals" showed how low the barrier to wiring AI into physical hardware has fallen.
I am not going to call Tesla's timelines from a thumbnail. The takeaway is simpler. When AI leaves the screen, speed stops being a nice to have. A robot that thinks for six seconds before each step is a very expensive statue. The fast decision layer is going to matter as much as the smart one.
Once AI moves into the physical world, decision speed becomes a hard requirement, not a feature.
Your AI Needs An Org Chart
Grok Bots had a big fortnight. Nate Herk's "Every Grok Bot Concept Explained for Normal People" was the best primer: each bot gets a name, a role, and a description, and those descriptions are how the bots decide who to hand work to. His advice was to stop talking to every bot yourself and build it like a company, where you talk to a few managers who know their team. Kacper Rutkiewicz tried over a hundred of them in "I Tried Over 100 GrokBots, You Need These 5," and Austin Marchese put Grok Bot against Claude across twelve daily use cases.
The Claude side was having the same conversation under a different name. Simon Scrapes covered "Anthropic Just Built It's Own Agentic OS," Jay E at RoboNuggets broke down "The NEW Agentic OS standard for Claude 5 Models," and Julian Goldie framed it as Anthropic turning Claude into an ecosystem. Jay's point stuck with me: the dashboard is maybe 30 percent of the value. The other 70 percent is how you organize context, skills, and memory underneath it.
Full disclosure, Acceleration Works runs on an agentic OS, so I watched these closely. What matters is not which vendor wins. It is that structure beats a pile of clever tools. Clear roles, clear descriptions, shared memory where it belongs and private memory where it does not. That is an org chart, and most businesses already know how to draw one.
Agents work best organized like a team: clear roles, clear descriptions, and a few managers you actually talk to.
Max Effort Is Usually The Wrong Setting
Nate Herk ran a great experiment in "I Had Opus 5.5 Build me the Same App at Every Effort Level." Same prompt, a 3D walkable version of a virtual conference built from 105 GB of recordings, run at low, medium, high, extra, max, and ultracode. The winner was Extra. It beat Max while taking about half the time and half the cost. Max overthought it.
The rest of the fortnight rhymed. Charlie Automates promised "This One Claude Skill Can Save You 400x On Claude Code Tokens." Jay E posted "24 Hacks to Solve GPT-6 Astra Usage Limits." Nick Saraev found a way to run agents like Codex for free. And Nate Herk's "Anthropic Engineer Explains: What to Build Instead of AI Agents" pushed the underlying idea: not every problem needs an autonomous agent, and a simpler workflow is often the better build.
The lesson is almost boring. Effort is a dial, not a status symbol. Medium is fine for most daily work. Turn it up for genuinely hard, multi step problems. Leaving everything on maximum is like taking a helicopter to the grocery store. You will get there, and your accountant will have questions.
Match the effort level to the task; the most expensive setting often produces a worse result.
The Company Of One, With Receipts
Shane Hummus told the story of a 60 year old nurse in "I Helped a 60-Year-Old Beginner Build A One Person Business With Claude AI in 30 Days." Thirty three years of nursing, burnout, very little saved for retirement, and a one person business that eventually passed $10,000 in a single month. His core advice was backwards from what most people do: build the machine that attracts buyers first, then decide what to sell.
Mike Sherrard showed the same idea in a specific industry with "How to use A.I. to Generate FREE Real Estate Leads FAST," and Dan Martell's "These Are the Best AI Tools in 2026" was a useful inventory of what a solo operator can actually run today.
Then Graham Stephan brought everyone back to earth with "I Paid Off My 2.875% Mortgage...Dave Ramsey Was Right." A finance creator known for arguing the opposite admitted that peace of mind has value a spreadsheet does not capture. I liked that pairing. AI makes the one person business possible. Boring money decisions make it survivable.
AI makes a one person business possible; a clear audience and boring financial discipline make it last.
Where It All Converges
Every theme this fortnight was a version of the same question: how much intelligence does this job really need? Jev says most decisions need speed, not brilliance. Robots say the same thing with higher stakes. Grok Bots and the Agentic OS videos say the answer is structure, not more models. Nate Herk's effort test says the top setting often loses. And the one person business stories show what happens when you get the allocation right: one person, the right tools, a real business.
The practical version for any business is a stack. Fast, cheap models do the triage. Smart models handle the hard cases. A clear org chart decides who gets what. And a human stays accountable for the outcome.
Notable Videos This Week
What This Means for My Content
Thirty two useful videos out of three hundred sixty one is a humbling ratio, and it is staying in the report. Two weeks of tracking made one thing obvious: the videos that helped most were not the ones announcing the smartest model. They were the ones testing where the smart model was overkill.
That is the content I want to make more of. Less "look what this can do," more "here is what this job actually needs." Right model, right job, right price.
Horsepower is cheap now. Knowing where to spend it is the whole game.
Knowing where to spend it is the whole game.