Claude Can Already Control a Computer, While Apple Intelligence Still Gets in the User's Way
Why can Claude already control a computer while Apple Intelligence can still get in the way of a simple user task?
Determine which work can safely be entrusted to Apple and Claude before granting real permissions; the assessment must set permissions, boundaries, stop conditions, and ownership of the outcome before automation begins.
What to watch for
Key takeaways
The “ToTheMoon - a sub-categist on the world of modern technologies” topic becomes clearer once this point is included: the risk depends on the scope of access, the scale of the consequences, and whether the system can be stopped and its actions reconstructed.
The discussion of “AI and Machine Training: practical context and limitations” yields a practical test: the important signal is not one funding number: the next round, available runway, and closure rate show whether a company can survive the new cost of capital.
The discussion of “AI in the defence industry” yields a practical test: the important signal is not one funding number: the next round, available runway, and closure rate show whether a company can survive the new cost of capital.
The “Claude from Anthropic learned to manage PK” scene leads to a working conclusion: it does not have to wait for a dedicated API from every service—it repeats the user's path. That makes the agent more universal, but also slower and more dangerous: an error now ends not in incorrect text, but in an incorrect action.
The boundary of the “Why? What scenarios of use of IIs, the manager of the PK?” case is defined by this point: the system can gather data from websites, process email, transfer information into a spreadsheet, or perform a routine operation in old enterprise software. But reliability has to be far higher than for an ordinary answer. If the agent clicks the wrong place once, sends an email to the wrong person, or confirms a purchase, the time savings disappear quickly.
The decision in “Apple started to introduce AI” depends on one criterion: this section clarifies the mechanism behind the topic and preserves a constraint that would otherwise be lost in an overly simple conclusion.
For the “Voice models: what is the main problem of voice recognition and voice communication with the AI” scene, the decisive point is this: the conflict reveals which rights, money, and control points the parties consider strategic.
The “Microsoft is putting every model into Copilot: who wins the product race?” scene leads to a working conclusion: for developers and content creators, that may be stronger than a standalone chatbot: the system chooses the appropriate tool and remains inside the workflow. The race will be won not by the company that first taught AI to click a button, but by the one that makes the action predictable, understandable, and safe.
What this episode is about
Anthropic demonstrated a model that opens a browser and clicks buttons like a person. Apple is adding AI directly to email and the operating system, but the functions appear at the wrong time and for unclear reasons. The contrast shows that the future of assistants will be decided not by model power, but by the quality of their actions and interfaces.
Claude Computer Use looks more important than another chatbot because the model gains the ability to act inside an interface: open a browser, find information, fill out a form, and move between windows. It does not have to wait for a dedicated API from every service—it repeats the user's path.
That makes the agent more universal, but also slower and more dangerous: an error now ends not in incorrect text, but in an incorrect action.
The real use cases are obvious. The system can gather data from websites, process email, transfer information into a spreadsheet, or perform a routine operation in old enterprise software.
But reliability has to be far higher than for an ordinary answer. If the agent clicks the wrong place once, sends an email to the wrong person, or confirms a purchase, the time savings disappear quickly.
Apple Intelligence demonstrates the opposite problem. The company has already integrated email-editing functions and suggestions into the system, but they can appear without being requested and fail to explain what they are offering. Users should not have to guess what a new button means. A good AI interface either solves a task at the right moment or stays out of the way.
Voice models still run into the same boundary. High-quality speech recognition is useful, but a conversation with an assistant often collapses into clarifications, mistakes, and the need to look at the screen again. Search inside ChatGPT also has no value by itself when email, cars, or office applications are already receiving their own embedded models.
Microsoft is therefore betting not on one winner, but on Copilot as a layer over different models. For developers and content creators, that may be stronger than a standalone chatbot: the system chooses the appropriate tool and remains inside the workflow.
The race will be won not by the company that first taught AI to click a button, but by the one that makes the action predictable, understandable, and safe.
An agent’s usefulness grows with the radius of possible harm, so access control is part of the product rather than a separate security setting.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 64 segments: 31 identified, 1 mixed, 13 marked with ✓, and 19 unresolved.
Loading…