I find it hard to believe that Anthropic couldn't have similarly impressive results if they tried. I'm honestly not sure it's worth it though, it's mostly marketing, right? But given how easy it is to switch from one provider to another, why would you choose the model that solved a Millenium problem, instead of the one that solves *your* problems?
I understand that it's still useful marketing, say for IPO or hiring, but it's not clear to me that it really matters in the long-term that they fund this out of their pocket.
Anthropic did have some great results with the Jacobian Conjecture [0] and the Riemann hypothesis [1], so I think Anthropic was aiming to solve one of the Millennium Prize Problems, but OpenAI did it first.
If OpenAI did it purely for marketing, I still think it was well worth it. It marked an important milestone in the history of computing (and also human history). It is on the same level as when IBM's Deep Blue defeated Garry Kasparov in chess, or when DeepMind's AlphaGo defeated Lee Sedol in Go. The search for winning moves in those games is not dissimilar to the search for paths towards mathematical proofs, and AI has finally proven that it is capable of carrying out deep, complex, and intricate research that exceeds human capabilities.
Besides, a Super Bowl ad costs about $8M, so it is not absurd if they spent a lot of money on this and became the first company to definitively demonstrate AI can surpass humans in mathematics.
Not all merges are equal, even if it's a monorepo, it's probably rare that you'd have to look at the whole codebase to implement a new feature or find the root cause of a bug. Once you start looking at the history of a specific folder, it becomes more tractable.
Of course that does require some folder structure, and agents are not necessarily great at that, but they're still decent enough that you do end up with many folders not being impacted hour or even daily.
Maybe it's all about how often we encounter the quirks of specific model? Or simply that one's way to prompt triggers more annoyances with a model than another?
I'd be curious to see if people have managed to just add custom instructions to make the outputs more standardized and less annoying. I never have such annoying cases, whatever the model, it's just too verbose, but in this case I ask it to be more concise and it's usually fine.
I didn't know about this limitation from openrouter, and I thought plenty of users were using it to quickly switch as needed without incurring such a penalty. Why does there seem to be so many users then? Are some folks just fine paying multiples of what they could?
You’re right, but the keyword in my post was “should”. If you have to look up “short selling” in the dictionary (as in OP’s case), naked options aren’t on the menu.
Regarding the multi-tasking induced by this new way of working (agentic coding), I've moved from quick and short iterations to a longer ideally one-shot prompt, optimizing for less reviewing and less context switching.
I used to send a prompt as early as possible, meaning that'd really just give the rough idea, barely a sentence, and see what the model would come up with, thinking that I'd save time if it got it right for the first time, and maybe I'd just need to correct what it got wrong (it's faster at typing than me). That means a lot of multitasking at a rather high pace, which takes its toll if you do it for too long.
I've recently switched to writing longer prompts (doing part of the planning myself), hoping to get better results and to have less to review: I'm realizing I'd rather write more and read less, past a certain threshold. I don't really want to iterate with an agent; I want to tell it, maybe in somewhat verbose terms, what I want, I want a quick confirmation that it'll indeed work, and I want the final output (e.g. a commit), that's it.
That's probably what some users have been doing since the beginning, but with all the hype around this tech, maybe I got caught by this idea that with just a few words, I can get stuff done. In my experience, it's indeed just a few words, but for useless units of work (iterations with a model) rather than what I really want (commit) which itself will often require non-trivial amount of input. In this case, I'd rather give it all in one go, in a somewhat focused state.
I would try to learn some new tech. Definitely not something you can do in a vacuum with no goal for months in a corporate setting, but e.g. learning more about a programming language you already use, or some libraries, some tooling, you can easily spend a few weeks.
After that, yes it'd make sense to find something else.
I understand that it's still useful marketing, say for IPO or hiring, but it's not clear to me that it really matters in the long-term that they fund this out of their pocket.
reply