What I Watched
June 30 to July 6, 2026
A new frontier model landed, and the ecosystem moved past it in 72 hours.
I Log Every Video I Watch. This Week a New Model Reset the Board.
I keep a running log of everything I watch. Not for the memories, for the patterns. Read a week back to back and the noise drops away and the real story stands up.
This week had an obvious headline: a new frontier model, Fable 5, landed, with Sonnet 5 right behind it. But the headline was the least interesting part. What the ecosystem did in the 72 hours after launch is the actual story.
Here are the five themes that defined the week, and why each one matters if you run a business that wants to move faster.
Fable 5 Landed, and Everyone Rushed the Field
The week's headline was a new model. Fable 5 dropped, and within a day the testing videos were everywhere. Chase AI ran through the top use cases, Nate Herk showed how Anthropic engineers actually prompt it, and Trevor Lowrey posted the part nobody was telling you about. Sonnet 5 shipped alongside it, and DIY Smart Code called updating Claude Code the number one thing to do right now.
What struck me is how fast the shine wore off. By day two the conversation was not how smart is it. It was what do I build with it. Tom over at ICOR tested Sonnet 5 in a real work folder instead of a toy demo, which is exactly the right instinct.
A new model is a bigger engine. Useful, but an engine alone does not move a business. The rest of the week was about the car around it.
Everyone gets the new model on day one. The advantage is what you build around it.
Stop Chatting With AI. Start Running It.
The sharpest shift this week was from conversation to automation. Sabrina Ramonov showed Claude running agents on a schedule with no babysitting, which is a genuinely different way to work. You do not sit and prompt. You set the job and it runs on a timer.
Austin Marchese backed it up with eight Claude loops to build ten times faster, and Anthropic's own channel published a clean explainer of the agent loop. The pattern is consistent: the value is in the loop, not the single question.
For anyone running a business, this is the unlock. The recurring work, the reports, the follow ups, the content, does not need you in the chair. It needs a system set up once and left to run.
The value is in the loop, not the single prompt. Set the job once and let it run.
A Smart Model With No Memory Still Forgets You
Right behind automation came memory. Nate Herk laid out every level of a Claude second brain, from simple notes to a full knowledge system. Systems by Vic showed an easy mode version built with Claude Code, and Sebastian Hardy shared a neat trick for wiring Claude Code straight into an Obsidian vault.
The point underneath all of it is simple. A model that resets every morning is a very smart stranger. Memory is what turns it into a colleague that knows your business.
This is the layer most people skip because it is not flashy. It is also the layer that decides whether AI actually compounds for you or just answers one off questions forever.
Memory turns a smart stranger into a colleague that knows your business.
The Open Source Stack Is Eating the Software Bill
Theme four was about assembling instead of buying. The AI Engineer channel published a missing manual for building great agent skills, which is the craft of teaching Claude a repeatable job. Then the repo roundups landed: Hyperautomation Labs showed ten open repos good enough to replace $55,000 a year of SaaS, plus a separate free repo packing 232 AI agents.
The message is that a lot of what you used to pay a monthly fee for is now a free repo and a skill away. Nuno Tavares covered six Claude Code repos in the same vein.
I am not saying rip out every tool you own. I am saying the build versus buy math just shifted hard toward build, and the businesses that notice will run leaner than the ones that do not.
Build versus buy just shifted toward build. Skills plus open repos replace whole SaaS bills.
The Local AI Race Is About Memory, Not Price Tags
The hardware thread kept running. Alex Ziskind had a busy week, on shrinking an AI workstation, on how your operating system changes local AI performance, and on lashing three PCs together to run one giant model. The Stack showed a 748GB RAM desktop making local AI genuinely good.
There is a lot of noise around cheap boxes right now, with headlines about a big model running on a $1,499 mini PC. The honest version is that memory is what matters. The entry price box does not have the RAM for the good open source models. The capable local machine is the 128GB class, like the GMKtec EVO-X2, and even then the real work is in the tuning.
For a business the takeaway holds from last week. Owning your compute is getting realistic, but buy for memory and plan for setup, not for the sticker price on a short.
Local AI is decided by memory, not sticker price. Buy the 128GB class, not the $1,499 hype.
Where It All Converges
Line the five themes up and they tell one story. A new model landed, and the whole ecosystem immediately moved past it to the system around it: routines that run the work, memory that holds your context, skills and repos that replace the software bill, and local hardware to run it all. The model was the spark. The system is the product.
That is the pattern worth internalizing. Model upgrades will keep coming, faster than any of us can track. The durable advantage is not catching each one. It is having a system good enough that a new model just drops in and makes everything you already built better.
Notable Videos This Week
What This Means for My Content
Tracking what I watch keeps me honest about the difference between news and signal. A new model is news. What people build with it is signal. This week the signal was loud and consistent.
If I had to hand a business one instruction from this week it would be this. Do not chase the model. Build the system. The upgrade will come to you.
The model was the spark. The system is the product.
The system is the product.