Blog / WEEKLY #004
#004 Opus 5.5 pushes boundaries, Muse takes off, and does China have AI product managers?
Flash off-the-shelf firmware, rework the interface into Chinese, bolt on a voice output path, and a small desktop dev board goes from toy to an assistant terminal that gets used every day; as for taking over the whole house, that is still just a vision.
Phone performance is already more than enough, yet an Agent can barely access anything on it: from permissions and sandboxes to app isolation, the mobile bottleneck lies beyond raw compute.
As the cost of generating content approaches zero, what becomes scarce is no longer output but a reason for people to keep coming back. Works, a consistent set of preferences, and feedback from real users are where trust comes from.
Switching to open-source or self-trained models to cut costs only works once the business is stable and quality can be verified independently. This article discusses the conditions under which that decision should be made.
Moving plugins from one agent to another, format conversion is no longer the hard part; the real threshold is that each platform exposes different capabilities, and upstream keeps changing fast.
As models start acting on intent rather than literal instructions, the center of gravity in collaboration shifts from prescribing every step to stating goals and acceptance criteria. Which constraints still need to be written down explicitly is worth discussing separately.
The ceiling of what a life assistant can do often is not the model — it is where the information it can see lives: the same school notice is readily at hand in email, but hard to reach inside a group chat.
Personal agents have lowered the bar to "being able to chat", but what decides retention is not the chatting — it is the first useful task: scenario entry points, an execution process you can see, and a conversation the user never has to manage.
Resale discounts on model APIs keep thinning; profit has to be earned by turning tokens into usable services. This article maps the cost structure of AI applications and the opportunity in "experience arbitrage."
Whether logins, software, and files survive after a task ends is splitting agent execution environments into two generations. Persistent environments make scheduled tasks and long-running collaboration possible — and push cost and multi-machine management onto the user.
Once visuals can generate in real time and shots can be produced in bulk, the real gap moves from "does it look real" to "does the story hold up." Authorship, serialization, and audience response decide how far this kind of content can go.
Once the barrier to creation has fallen so low, how a work actually reaches its audience becomes the new bottleneck. Distribution does not happen automatically just because something is well made; it needs to be designed and validated in its own right.
Models keep getting faster and code keeps getting cheaper to write, but the speed at which software reaches users has not automatically caught up. Between individual output and organizational delivery sits a whole set of steps that have yet to be redesigned.
A promo video can be composed from a musical score, vector animation, and a handful of generated images. Code handles the orchestration, image models supply the key assets, and whether it looks good still has to be judged by eye.
Remembering a fact, tracking a goal, and completing a task are three different states. When an assistant cannot tell them apart, it keeps nagging about matters that ended long ago.
When an agent needs to connect to a new service, should the user tap once to authorize, or go apply for an API key themselves? The difference in that one small step separates two generations of product design.
Buying an always-on machine does not mean you have to run models on it. This article pulls apart three needs — local inference, an always-on device, and privacy control — and judges when each is worth adding to your own setup.
