Timothy Cardoza

a guy

[Timothy Cardoza @TCCardoza]

@suni_code Workflows

[Timothy Cardoza @TCCardoza]

@himyis_ @georgevibing Sol medium - xhigh work really well. It's exponentially slower and less efficient as you go up to higher reasoning levels and Ultra is actually not a reasoning level. It's a system of using multiple Sol agents on Max reasoning. Pretty much costs 100% more for 2% better outputs

[Timothy Cardoza @TCCardoza]

@JoelCarpenter_ @georgevibing You might give Terra high a shot as well. It's a surprisingly good sweet spot. Just did a deep dive (see my most recent post for the details)

[Timothy Cardoza @TCCardoza]

Did a bit of a deep dive to determine how valid Terra is for coding in cost/quality for coding in codex.

Terra high is a sweet spot. Faster and more well rounded in intelligence than Luna max while sitting just below it in code quality and at nearly the same cost. The significant reduction in reasoning tokens drastically reduces the frequency of compaction and needle-in-hastack issues. The increased embedded knowledge and parameter count give it much better judgement capabilities as well.

[Timothy Cardoza @TCCardoza]

@Da7_Tech It does seem to be pretty darn good. But I never do the same kind of work with Grok as I do with Sol so it's hard to really compare. They FINALLY added /goal to Cursor so that helps

[Timothy Cardoza @TCCardoza]

And Sol is NOT token hungry. It's one of the least token hungry models near frontier quality. It's work heavy, it will check everything and test everything and handle every edge case. So yeah, it'll use tokens cause it does everything when you don't need it to do EVERYTHING. But that can be steered reasonably, albeit not perfectly.

[Timothy Cardoza @TCCardoza]

Few hours??? What the heck are you doing with Sol??

I just had 2 Sol threads (1 high and 1 xhigh) run for 17 hours straight and use 10B tokens on a goal. And it used roughly 35% of my 20x plan's weekly limits.

I will say, auto-approve does take some usage and I had it on full access but... You're talking rediculously different rates

[Timothy Cardoza @TCCardoza]

Well, I use Codex or whatever other coding Desktop app (not much of a terminal guy) and then want a way to handle git for multiple repos, some basic file editing, dev servers etc., And Zed is just awkward for that kind of thing and leans too far into terminal-style UI for me.

I literally built my own desktop app just for multi-repo git/terminal/file editing to go alongside my coding harnesses cause I couldn't find anything that was truly minimal but clean and GUI.

Btw I open sourced it if anyone is interested. Wasn't trying to promote but now I feel like I have to 😂
https://t.co/RS5D3Wstdn

[Timothy Cardoza @TCCardoza]

You realize this is just 2 (non-coding) benchmarks right? Sol hasn't had a downfall. I go back and forth between Grok 4.6 and Sol regularly and it's very apparent that Sol is the "more correct" one and Grok is the friendly "get the job done for the most part" one. Always have to have Sol check after Grok and it always finds really important things

[Timothy Cardoza @TCCardoza]

@kayaravenya Yes but sadly 😂

I thought about trying to build one but idk if I'm ready for the scale that would be required to make it happen effectively

[Timothy Cardoza @TCCardoza]

@doja_dan That's actually great

[Timothy Cardoza @TCCardoza]

@validatedev @amorriscode @theo I couldn't get into Zed... Feels to overly complicated for my needs with agentic development

[Timothy Cardoza @TCCardoza]

@theo @amorriscode It's free and open source but still kinda beta. Built it for myself more than anything and only just made it public and added some documentation.

If anyone is curious: https://t.co/3ZItd9n7Sw

[Timothy Cardoza @TCCardoza]

@theo @amorriscode I built a multi-repo git/terminal/file editing Tauri app just so I could avoid opening vs code. Honestly, I'm biased but think mine is way better for the agentic era. Just columns of repos with perfect visibility of all the terminals and which ones are running dev servers etc

[Timothy Cardoza @TCCardoza]

@pawar_sahi63500 Ye

[Timothy Cardoza @TCCardoza]

Fable is much better at understanding the situation and what needs to happen, but it's lazy so it doesn't follow through like it knows it should. Grok isn't lazy and is great at keeping track of things. It goes further but isn't as good at deciding where to go. (Also Sol is basically Grok that checks it's taste at the door and is 10x more thorough. I always have Sol check Grok's work and it always finds important things)

[Timothy Cardoza @TCCardoza]

It really is amazing. I feel better using Grok 4.6 but then I use Sol and realize that Grok skipped critical things... Again.

Also quick tip for handling Sol: Don't tell it to stop adding tests and handling edge cases. Guide it to do it with higher quality. Use it's strengths, don't shoot them

[Timothy Cardoza @TCCardoza]

@doja_dan Very interesting, I've been thinking about agentic discovery for a while now. Wondering how deep that well is

[Timothy Cardoza @TCCardoza]

@LexnLin I wonder if their gonna make that the Luna limits and their just updating that