Comment by ozim
10 hours ago
In terms of IDE yeah it is not that great.
I do have Copilot in VSCode and Cursor.
I thought both should be equal in solving problems - turns out Cursor with the same model selected somehow was able to solve tasks that Copilot would get stuck or run in loops.
They have some tricks on managing file access that others don’t.
Cynics on HN easily dismiss AI service wrappers (and many of them are in fact overblown and not worth their own code). But writing a genuinely good harness with lots of context engineering and solid tool integration is in fact not that easy. The biggest issue is that model providers also see what the community likes and often move on with their own offerings that are tailored to their own models, potentially at the training stage. So even if you have the best harness for something today, unless you are also a frontier LLM provider, there's zero guarantee you will still be relevant in the future. More like the opposite.
> But writing a genuinely good harness with lots of context engineering and solid tool integration is in fact not that easy.
true, but its not worth $60 billion fucking quid.
It's not like someone paid $60 billion for a product the way you pay for bananas at the store. They invested a much smaller amount and essentially bought an option to acquire. And even if you don't believe the company's assets are worth the current valuation, an acquisition can still make sense if you believe that valuation will go up further. And if they actually do acquire, it will probably still not be in cash. They'll just be swapping stocks. That is essentially how all startup funding works. There is nothing strange about this. It merely reached new dimensions thanks to AI.
it's insanity.
the whole thing is driven by irrational stock market investers who NEED ai to be the thing that saves the world.
they're betting everything on it.
1 reply →
> (...) writing a genuinely good harness with lots of context engineering and solid tool integration is in fact not that easy.
This. They are after the harness engineering experience of the Cursor people, I'd assume the they want to absorb all that into Grok's offerings.
The value and the room for innovation on the harness side seems to be underestimated.
Oddly the harness also affects model training, since even GLM/Z.ai for example train (I suspect) their model on the actual Claude Code harness. So the choises made by harness engineers affects the model. For Kimi/Moonshot and OpenAI the company makes their own harness. Alibaba uses Gemini.
Very interesting dynamics.
There are plenty of harder things in the world and very few are worth 60B.
Something being harder and attributing value to that makes no sense. Sure a big moat is important for value but "difficult to do" is just a unidimensional angle.
Isn't Codex TUI available for free though? Besides others like Pi and OpenCode of course.
It can use local/oss models, but it doesn't make it simple to do (easiest with ollama) and it's not clear what else you 'lose' by making that choice.
If you had a really good (big) local model, maybe it's an option, but on the more common smaller (<32b) models, it will have similar problems in looping, losing context, etc. in my experience.
It's a nice TUI, but the ecosystem is what makes it good.
I actually now think ai prompt writing in the IDE is completely overkill nowadays.
IDEs are made for just a human to interact with code. I think the paradigm of forcing these tools that weren’t built for this to do this, is us trying to fit a square peg in a round hole.
Call me old, but don’t put ai in my ide. My ide was made for a human, not an ai. For the established players for sure it makes sense since they already have space on our machines. But for the new ones imo terminal, or dedicated llm interfaces are where it’s at.
If I’m writing code sure suggest the next line. If the machine is writing code, let it, and just supervise properly. and have the proper interface that allows the strength of each
>They have some tricks on managing file access that others don’t.
I thought it was a Windows thing. My Windows work computer is so heavily managed and monitored I assumed that was why Copilot stops being able to get terminal output or find the file I'm looking at. It's the same problem in IntelliJ and VSCode, with different models trying to find things in different ways.
Now that I think of it though, I've only used Copilot at work. At home I use Debian but I've never tried using Copilot. Claude, OpenCode, Gemini, and IntelliJ's AI Chat pointed at local Ollama models never have issues finding files or reading files and terminal output.
Sure, but is it worth 60 billion?
Their annualized revenue run rate is on track to surpass $6 billion by the end of 2026 so it's not ridiculous for them to be valued at $60 billion at some point. Also worth noting that if they do get access to SpaceX compute, they could start pretraining their own model. Composer is good but its built on top of Kimi 2.5.
Definitely not if someone frames it "shitty IDE with some plugins".
But if someone frames it "engineering talent that knows how to make LLMs even better at software development than competition" it might.
I see with my own work it works so it is not like Devin that was basically a scam that was valued at 10 billion.
In this kind of context yeah feels like it is quite possible to be worth 60 billion.
SpaceX thinks so.
SpaceX the space rocket and internet satellite company? Or SpaceX the Elon Musk piggy bank used to buy up all his financial misadventures?
You mean Musk thinks xAI need to be shown making AI investments to keep getting outside funding.
They're using the code intelligence from the IDE to run the AI, while Claude Code only does greps.
AI coding is much more than just the model - all the tools that human use in IDE are also useful for AI. Claude Code on the other hand just works with grep.