AI should work for you.
I want to give a model a whole assignment and let it choose the tools, find the information, and bring back work I can use. I want fewer interruptions and less time managing every step.
I also want to recognize when it starts making decisions I should own. The purpose of the work, the tradeoffs I will accept, and the point at which the result is good enough still need my judgment. Handing over more work should leave me with more control over my day.
That’s why Lenny Rachitsky episode with Tara Seshan is worth a listen.
Tara is giving us a roadmap on what’s coming for GPT. Namely that we are looking at a world where models do more of the work we want them to handle, and at the same time, start making the uncomfortable jump into having us work for the machine, instead of the machine work for us.
Remember, Tuesday’s issues are for learning. Articles, podcasts, content that I find helpful, distilled for you. Friday’s issues help you learn to build with practical workflows and tooling exemplars. Everyday, you use AI to improve your work.
What’s Next
OpenAI describes Astra as a model for complete workflows across code, apps, and research. Anthropic describes Fable 5.1 as a model for complex knowledge work with less supervision.
That is the kind of AI I want working for me. Give it the purpose, the material, and a clear finish line, then let it make the routine choices. Fewer interruptions should mean more room for your own work.
Staying in control means understanding which model fits the assignment and why, and when separate agents can help. You also need to recognize when the AI has misunderstood the goal or made a decision that belongs to you, so you can step in before that decision impacts the rest of the work.
You learn that by trying, using, building, and failing on work you can check. Compare two models on the same report. Notice which one needs repeated correction. Try dividing the research into separate questions. Keep the change that improves the result. A failed attempt gives you something specific to fix before you trust the workflow again.
OpenAI’s Astra guidance notes that it can pause for clarification more often than a user expects. A useful brief tells it which routine choices it may make and which decisions need your input. The aim is to make AI work for you without turning supervision into another full-time job.
Are Astra and Fable the beginning of that promise or the end?
Useful now
Choose the model for the assignment. OpenAI’s current guide places Astra on demanding work across several steps and tools, Sol on difficult work needing depth or polish, Terra on everyday work, and Luna on clear, repeatable tasks. Treat that as a starting point for comparison. Judge the finished work, the corrections you had to make, and the usage it took. A model that handles a routine summary well may be the wrong choice for a recommendation built from conflicting sources.
Fable 5.1 targets longer assignments. Anthropic’s latest release focuses on coding, knowledge work, and problems that take sustained effort. One useful detail is that Fable 5.1 defaults to Medium effort in Cowork and Claude.ai. Try a familiar assignment at that setting before increasing it, and compare the finished work with your previous result. Treat the launch benchmarks as the vendor’s evidence; your own report or briefing still needs a check.
How much work should you inspect before trusting an AI reviewer? Hamel Husain and Shreya Shankar are the best at teaching evals. Why does it matter? Think of an eval as the exemplars of your daily work. A way to show the AI agent what good looks like so it can do similar work. How do you build that? Well they recommend starting error discovery with 100 varied session records and reviewing at least the first 30 yourself before asking an agent to suggest failures. Their advice applies to evaluating an AI application; you do not need 100 examples to try one report. The useful habit is to establish your own judgment before letting AI help apply it.
Email could become an interface for more of your work. Matthew Guay’s essay makes the case for interacting with software through the inbox you already use. Its a great reminder that as we speed ahead with new technology, the old, but still useful tech, sometimes becomes even more important.
Thanks for reading AI for Humans! Subscribe for free to receive new posts and support my work.
What’s possible
Google is testing Gemini as a desktop superapp. TestingCatalog’s report describes Ask and Assign modes, folder-based tasks, and Obsidian connections in a closed test. The possibility is one place to ask a question or hand over a job using your own files.
Claude can carry context into the next assignment. Anthropic’s update brings shared memory to chat and cloud Cowork tasks, with saved topics you can inspect, edit, or delete. The possibility is less time re-explaining your work while retaining a way to correct what the system remembers.
People using more AI still want to own the thinking. Microsoft found that 87% of surveyed AI users say they remain responsible for the thinking and treat AI output as a starting point. This is self-reported behavior from a vendor’s research, rather than proof of better decisions. Teams can use the finding to discuss which judgments they will keep practicing as they delegate more work.
Harvey is training a model around professional work. Harvey’s Tenet preview describes post-training Kimi K3 with Fireworks for legal assignments. The company reports gains in task performance and cost efficiency. I think the effort is sound. Whether those gains give users better work with less supervision remains an open question. More broadly, I think open models and small language models still need more guidance than the average professional can manage directly. That is a concern about usability, but there is real promise if the provider absorbs that complexity and returns work the user can trust and check.
The loop this week
Pick one familiar report or briefing and run a small comparison using fictional or approved material. Give two available, approved models the same sources, audience, and definition of a useful result. Allow routine drafting choices. Require a pause if a missing fact would change the recommendation, and keep sending the result under your control.
Review both outputs. Record the errors, the interruptions, and the corrections each needed. If the work includes independent research questions, try that part with subagents and compare the evidence they return.
Build what you learned into the next assignment. Keep the brief, the correction, and the version that worked so you can repeat the method and check whether it holds up.
Get the next one on Friday.
Applied AI for professionals, by Sal Carranza. One workflow a week, free.