Illustration: LittleDot concept art, AI-generated.
When I hear that the next generation of models may be on the way, a small wish comes to mind first: could I hand it one piece of work and get back something genuinely finished?
Take a website update. The model has to understand the existing code, find the right place to change, preserve the parts that already work, and check both desktop and mobile. At the end, it should tell me what it changed and what still needs attention. That sounds simple, but it involves a long chain of details.
Change too little and the page may look fixed while the interaction still fails. Miss one check and the desktop view can look polished while mobile buttons overlap. If the delivered result does not match the actual page, I have to retrace the work myself. After a few rounds, the time saved has disappeared.
Fireside Alpha's notes from a conversation with Boris Power mention GPT-7, GPT-8, and later research. That is a signal worth following. But we do not yet have a GPT-7 hands-on test that can answer whether it will solve these everyday problems.
I would test it on a real project, not start with a scorecard
If I could try GPT-7, my first assignment probably would not be a puzzle. I would choose a real project with legacy code, decisions made over time, and a few requirements that were not written down clearly, then ask it to take the work from the beginning through delivery.
I would watch whether it can separate what has already been decided from what it still needs to ask me. It should handle small choices on its own and pause when a decision could affect the whole project. Asking me about every step turns me into its operator; never asking and discovering the direction was wrong at the end is just as frustrating.
Continuity matters too. If we stop today and continue tomorrow, I want it to pick up the existing work. It should retain what changed, what did not, and which questions still need answers. When switching projects, it also has to keep the two sets of requirements separate.
Then there is recovery from errors. When an approach fails, the model should investigate why, try another method, and check again. Repeating the same step or declaring success with different wording wears down trust. I would rather hear: "This step is blocked for this reason, and these are the approaches I have already tried."
Speed still counts. I can wait a little longer for one answer; it is much harder to stay focused when one task takes a dozen slow rounds. For me, avoiding one rework-causing mistake can be more valuable than shaving a few seconds off a response.
That is my simple expectation for GPT-7: give it a task, open the finished page, run through the workflow, and see whether the result matches the requirements. If it can do that, I will be glad to keep using it and recommend it.
When it arrives, we should test it on work we actually have. The next article I want to write is about how far a project got, not just which chart a new model topped.