Tweet fixado
Luke
6.8K posts
Luke
@ldo_dev
Building @LogoLattice to make the best in browser logo creator 🎨 Building @FlowInvoice so you can get paid faster and stress less 💰
Entrou em Aralık 2010
459 Seguindo2.1K Seguidores
@LailaAiClub Yeah I'll definitely be automating which model is recommended. I'll also give the user the option on changing their model too
English
LLM testing update for Logo Lattice:
Opus 4.8 is struggling the most. Inconsistent coordinates, breaks schema rules, can't hold geometric precision across attempts
Complex shapes: decent. Simple shapes: inconsistent.
Maybe I need model routing based on logo complexity 🤯
Luke@ldo_dev
I'm integrating LLMs into Logo Lattice, but not for image generation It gets a brief ("minimal fintech mark, teal + charcoal") & our schema It return JSON: shapes, points, grid snaps, fills Testing GPT/Claude/Gemini What would you try for structured spatial JSON?
English
@vaademola I'm writing it up at the moment :) I'll post about it in article form when it's finished. Should hopefully be out today
English
@TechWithMatteo That's really interesting to note. Did you find that any other models were any better? From my limited testing across limited models, I've found that gpt5.5 and Opus 4.7 are leading at the moment
English
@ldo_dev model routing on complexity is smart, i had similar issues with Claude on structured outputs where simpler tasks were somehow less reliable than complex ones, feels counterintuitive but its a real pattern
English
@aaliya_va Yeah definitely. I think I'll try and handle it semi-automatically with a script that rates the complexity of the prompt, but I'll also give the user the ability to choose their own model
English
@SaniAiTech Yeah definitely. I'm going to write a script to determine the complexity of a prompt, and then have a popup with a different model recommendation. Lots to work out with it
English
@ldo_dev Interesting finding—seems like Opus 4.8 struggles with strict geometry but does better on complex shapes. Model routing based on logo complexity might be the real solution here 🤯
English
@farrux_hewson I have created an internal benchmark. I’ll release my findings when I’ve finished writing them up. Open sourcing it is a good idea too. It’s very niche but interesting
English
@ldo_dev You should create internal benchmark and make it open source, research community will like it!
English
@ldo_dev Nice.
It's better to do real testing than relying on assumption
English
@bas_fijneman GPT5.5 and Opus 4.7 are by far the best two so far. I’m mid write up of my tests. I also will run some more benchmarks on text based logos too. This is just the start of the can of worms I’ve found 👀
English
@ldo_dev Hmm I would advise staying with 4.7 for this for now! I see to many inconsistencies atm. Also a lot of work done that I didnt ask for
English
@MythThrazz 55k views a day on average is so hard to do without consistently posting engagement farming or striking gold on a super viral post 😂
English
I think its not a problem to get 5M impressions overall.
It’s a challange to do so without:
- spamming
- cheating
- crypto
- sports
- ragebaiting
- politics
What’s left is unique trivia and wholesome content.
English
The Sunday routine to avoid founder burnout 🔋
✅ 5k morning run
✅ Cricket session in the sun
✅ Sunday brunch
⬜️ Planning the week's marketing sprint
Have a great Sunday! ✌️
Suhas@zuess05
2 weeks ago, I quit my $100k job to indie hack full-time. Not gonna lie, I was super terrified. Since then, the momentum has been unreal: ✅ Got monetized on X 🐦 ✅ Made $197 so far from it 💰 ✅ Hit $161 in revenue for apps (300x)📈 ✅ First major YouTube video goes live next week 🎬 I’m not saying this happened because I quit. But I quit because I knew it was about to happen. We’ve got a mountain to climb, but the view from base camp is pretty damn good. 🤌
English
It is possible that Opus 4.8 is a much better model than I give it credit for, and it is held back terribly by Claude Code.
Sadly I cannot confirm this myself without spending thousands of dollars because their API prices are absurd and you can't use your sub anywhere good.
Theo - t3.gg@theo
I haven't used Claude Code much recently. Are these write failures just, like, expected? I haven't seen this many random tool call errors since I last used a Gemini model
English
@ldo_dev Looking forward to that, especially the logo/text cases.
English