I’ve been using a mix of models lately. Mostly the various Grok’s, and Sonnet 4.5. I keep going back to Sonnet, because overall, it has thus far presented a much better OVERALL EXPERIENCE than any other model.
I noticed something about it today, well, I noticed it before, but after having used this “cheetah” stealth model for a few days, going back to Sonnet 4.5 this characteristic was so noticeable and so NICE!
Most models, thus far, have been this: Hear It, See it, Fix It. You give a prompt, the model sees the issue, and fixes it, and fixes it, and fixes, fixes, fixes. OR does, does, does, does. No stopping, no interaction. You give it a prompt and the model just MOVES FORWARD, regardless of whether what it is doing is correct.
Sonnet 4.5, is different. It is what I would call a conversational model. It seems to recognize, when it does not have enough information, or enough explicit instruction, in order to proceed on its own. In which case, it will preemptively stop, ask for further instructions and then of course, will not proceed until you prompt again. In other words, working with Sonnet 4.5 is having an ongoing conversation with the model, much like you might have when pairing with a junior dev!
This is such an incredibly POWERFUL improvement to agentic coding models!! I wanted this six months ago, when I first started using cursor. Just the last few days, using this “cheetah” model, it was often so frustrating, as it would just PLOW AHEAD and do whatever it decided it should do, and it often just does things wrong. It would make good changes, then revert them when it came across different code it hadn’t consumed into context yet, sometimes it would just make the incorrect changes (often the opposite of what I prompted it to do). I used that model, for three or four days here.
Switching back to Sonnet 4.5 today, and the experience, and the overall quality of the results as a whole, is SO MUCH BETTER. Sonnet STOPS, on its own, when it realizes it may not be able to proceed CORRECTLY. I am not sure how Anthropic’s models determine that, but so far, it seems to be pretty good at determining that state, and stops with information about what it currently knows, and asks how it should proceed.
IMO, this CONVERSATIONAL approach, is vastly superior to the “Just POWER FORWARD and DO IT” approach that most other models take. I am hoping this catches on, and that moving forward, this is the general trend with agentic…well, agentic anything. These LLMS are not really intelligent, they are super-dense knowledge bases and NLPs. So IMO, they cannot be relied on to MAKE GOOD DECISIONS. They can produce content, be it a plan, code, whatever, quite well. But they are terrible decision makers, and shouldn’t be relied upon to make decisions.
Sonnet 4.5, by stopping and asking ME to make decisions, is just so awesome!