Hacker News Reader

Comment by vorticalbox | Hacker News Reader

its not just the models, their auto complete is actually really good. when you make a change it will give you "tab to next" which makes refactors super easy.

composer 2.5 is also a very decent model, it go 90% of my AI tasks using it now.

View on HN

So 60B dollars and your primary reason it’s worth that is “tab to next” autocomplete?

Maybe the real Mythos secret is that it found out it could fit on a GPT2 size model

Scaled to every developer in the world? Yeah, that's productivity.

I'm not sure it's worth the price tag for developers though, I mean resharper promised similar, delivered half, bought by millions and still don't think Jetbrains, even with its other good tool suit is valued 60B.

I just laughed out loud reading this

Hyperfocused edit assist is awesome. I’m the magician the assist is the magic wand.

Investing in companies is not about what they are delivering right now, it is about what you think they can deliver in the future.

It is not like purchasing soap in the supermarket.

Not to mention VSCode has that too

Yes, that's why Cursor was very popular when actually reading model output paragraph by paragraph was still the way you used them. That's no longer the case, their use has cratered, and in fact they have been disintermediated by their model vendors, leaving an empty shell.

I guess that fits into the Musk empire because "empty shell" increasingly describes his companies.

Tesla is a car company that doesn't want to make cars. And xAI is an AI foundation model company that actually is a data center REIT...

I tried it a few times and was not convinced by the autocomplete.

I found it less effective than free copilot autocomplete on vanilla VSCode.

composer 2.5 is literally just a fine-tuned kimi model, and the autocomplete is exteremly meh.

The only kind of AI I want in my editor is an autocomplete, but this isn't very magical to non-programmers (their TAM) or all that valuable (you can't charge thousands), they bought Supermaven and basically killed it, I'm not sure how you think tab is really good, I've not been impressed when I played around with it.

What tab is better?

I'm honestly not up to date on all the FIM autocomplete at the moment. However, autocomplete getting good is something I look forward too. I believe its mostly a problem of latency. You needs <200ms (and accurate) for it to not feel annoying, which is hard to do and use a larger more capable model.

Unfortunately it doesn't seem like this is a problem companies are looking to tackle because it doesn't really make you any money or impress investors. Inceptionlabs.ai seems to be making attempts, but I haven't tried. They are apparently in the Zed editor, something else I haven't used.

https://www.inceptionlabs.ai/blog/introducing-mercury-edit-2

Yeah, is that a bad thing? They did some really interesting stuff when fine tuning.

https://cursor.com/blog/composer-2-5#targeted-rl-with-textua...

Reply on news.ycombinator.com