I have to admit, for a "dumb" machine that ultimat...
# random
s
I have to admit, for a "dumb" machine that ultimately generates text by predicting statistical patterns, Ox Alpha is surprisingly "smart". 😄 I'm using it again to review & improve my libraries and it just came up with the idea that a made-up hostile TIFF file could send my image metadata library into an endless loop. To best honest I didn't even think of such a scenario. I attached the prompt I'm using to let it compile a TODO list. ChatGPT crafted that for me. The prompt says "application", but it works for libraries as well. I run it for all my apps and libs through different models and 90% of the findings DeepSeek V4 & Ox Alpha had were valid points.
👍 2
Mind the 10% false-positives where perfectly correct code is flagged as wrong for unknown reasons. Even Ox Alpha made mistakes today and wasn't able to explain why it did so.
f
Time to burn some tokens 🔥 😄
s
Yes, Ox Alpha is free on OpenCode & OpenRouter this week. Perfect opportunity for another bugfix round. 😄 I'll release fresh lib versions later this day when AI is done fixing my mistakes. 😉
It has high demand right now, so you'd need to write some "Continue" statements from time to time. 😄 AI automated everything we do, but not yet this. 😁
f
I’ve just typed “Continue” for the 6th time 😂 Now time for lunch
👌 1
s
I also recommend multiple passes of the review after all issues have been fixed. I had an CI cache key misconfiguration that for some reason needed three passes to surface.
👍 1
f
Did 12 passes; Almost 3000 lines of TODO.md have been generated 😅 Most not very shocking.
🎉 1
s
That's a good sign. 😄
I found 0x Alpha to be great on security & vulnerability stuff and pretty weak on generating UI code. Very interesting. I assume it was trained with another focus.
Too many people 😁
a
I am wondering what Ox Alpha really is
s
Everybody does.
Some suspect a new GLM version. Reddit is full of speculation. I guess that's what they wanted. 😄
a
So basically US stocks are gonna tank because it is the next open weights model which outperforms the Claude’s of this world.
s
I don't know if it will be open weights or who made it.
US stocks tanked first after DeepSeek V1 release and recovered well.
a
I am really for open weights models. So fingers crossed. Ox Alpha worked near perfect in my projects.
s
I had two really stupid things so far, but overall it's a really competent model
This is where it's really strong.
@Arjan van Wieringen z.ai/blog/glm-5.3-flash
a
Is that the ox alpha one?
s
Yes.
You can use it now on OpenCode Go/Zen and OpenRouter.
a
Interesting. In benchmarks it scores way lower than opus and glm5.3. But that was not my “feeling”
➕ 1
s
I found it pretty competent, too.
It's also discounted on OpenRouter right now. So far costing only a few cents per task for me.
I loved DeepSeek V4 Flash pricing. 😄 Converting my old dots-and-boxes Android game to CMP took 3 hours and they billed me $0.20 on OpenRouter. So awesome. 😁