Hacker News new | ask | show | jobs
by kamikazechaser 17 days ago
It's a shame that they exfiled private data. The model is actually good (better than opus 4.8 imo) and the harness itself is butter smooth with the potential of being the best out there.
7 comments

It definitely doesn't feel like opus. I constantly switch to opus to fix up or finish what grok generates, it feels like sonnet 3!
Grok 4.5 is somewhere between between Opus 4.8 and Sonnet 5.

Source : https://artificialanalysis.ai/models/capabilities/coding

I know this is what the leaderboards say, but empirically, I'm also seeing it make silly mistakes that opus and fable never do. So it really feels more like sonnet to me as well.
It's definitely around Opus level. It's definitely a lot smarter when it comes to review or asking if there's gaps or things missing.
This has been my experience as well. In fact, Grok 4.5 is better at visual design than Fable from what I've seen.

And being (based on vibes) 2-3x faster? It's an easy sell to me.

I agree. I subscribe to SuperGrok but never used the grok models a lot for coding. Now with 4.5, I’m gonna hit my weekly limit tomorrow and even considering trying SuperGrok Heavy

I really like the feel of Grok 4.5

I had a very weird experience two days ago where Cursor-Grok-4.5 was either stuck in a loop (it would keep attempting to answer the prompt over and over), or else it would just quit halfway through a reasoning loop. Might have been that I was using omp, but it's still not the most stable thing out there.

Nonetheless when it's working, it's pretty good, and for the price ($10 a month) is an absolute bargain.

What’s this $10 deal?
Find a referral code for Cursor. Sign up with first month half off.
Its amazing the speed of build with grok 4.5 its a taste of whats to come.
It’s a shame that their leader exfiltrated government data.
More specifically, DOGE exfiltrated private and precious information about US citizens which the federal government had collected.
Now this is contrarian!
Or a spaceTwtterAi stock holder…
I’m an anti-Musk zealot and I now pay for a subscription, it’s pretty good stuff
You have a weird definition of "zealot".
Imagine have principles this worthless. Must be pretty nice actually.
Everybody has a price - Kim Jung un
> I'm

Wrong tense.

That was a mistake and they deleted all the data.
How do you accidentally upload your user's repos to a storage bucket?
That's easy, you ask a LLM to do it
A naive implementation would just upload everything in the current directory.
That's what they did to upload a repo, the question is why they would upload the user's entire repo in the first place. The idea itself is egregious, and the implementation was horrifying on top of that.
'on startup, grok upload current directory for faster cache access'

Was it intentional for data exfil? Only internal staff could answer.

The entire directory (not repo) and all repo history regardless of secrets or instructions otherwise is not needed for faster cache access. I'm not going to say that this was intentional exfiltration. Sneaky motivations are not needed for this to be a bad idea and a horrific implementation.