Lucas Petralli (@PetralliLucas) header
Lucas Petralli (@PetralliLucas) profile picture

having fun with LLMs building https://teapawt.net/t.co/UpJp6B5byk 🇦🇷 LangChain Ambassador

Joined July 2021
Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)Photo by Lucas Petralli (@PetralliLucas)
If frontier models can edit Qwen 3.5 0.8B weights to be better at math, can it give the same model taste? Probably not, but it can make it a better design critic. Here's what we found:
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
7
11
79
8,056
playing a little bit with nano banana 2.1 + the photography extraction endpoint of the taste engine 🍌 goal: generate images that actually follow a brand's photography style. tested it with patagonia!
1
5
13
660
Replying to @tastelabs
@tastelabs 🤝 @trymelius two teams obsessed with making AI a better creative tool making on brand assets just got a whole lot easier ;)
Introducing Brand Kits. Whether it be an indie content creator or a creative director at a Fortune 500, staying on brand is crucial to creating visual content. Type in your website or upload guidelines, and the agent will handle the rest. In partnership with @tastelabs
1
6
1
57
2,732
Over the last few days we shipped some new cool stuff to the Taste Engine. One of my new fav features is that now brand extractions includes extracted assets from ads, social channels, and more! Really useful to give your agents more visual references 🧙‍♂️
5
2
1
27
2,003
Amazing use case of the brand API to verify brand adherence on @vercel previews ▲ Really really loved this! Thanks for sharing @mfbzme 🫶
made a @github action that uses the @tastelabs API to score your PRs' @vercel previews against your brand's design guidelines and give your agent the fixes. here's the demo 🤖
2
5
18
765
Lucas Petralli retweeted
check out the full demo on how to implement the API or MCP docs: docs.tastelabs.com/
🤖 Made with AI
13
11
9
76
8,762
Comparing Jev and Opus to see how they classify the visual style of a webpage.
4
7
4
52
6,327
Lucas Petralli retweeted
Introducing our first agent tool: the Brand API This helps gen AI agents solve 3 problems: - Extract: produce on-brand assets when users already have a brand - Search: find and retrieve brands for inspiration, from our curated index - Verify: check their own work against, see what's off brand and how to fix it Anyone can try it for free today at tastelabs.com/api Check out the full blog, and case study with @intelligenceco @ndrewpignanelli below.
22
13
5
217
40,900
Lately i find code review to be an especially valuable human skill. But not the traditional code review; agents that write and review code already do a pretty solid job in terms of functionality and optimization. Where the skill really matters is in guiding agents to write 'tasteful' code, not just working but ugly code. Even with the latest models, i believe agents optimize toward functional and performant, but ugly and hard to read code (maybe due to reward hacking in post-training RL environments? or maybe is just an skill issue?). Whether we'll still need this in the future is up for debate, but for now, human reviewers with this skill seem like the best way to build maintainable (and agent-friendly) codebases at a fast pace. Optimize for faster and better code reviews.
2
3
1
22
945
We've been working with the top frontier labs to improve their design model capabilities. Today we're bringing that to your agents. Introducing our first agent tool: the Brand API. As the cost of creation goes to zero, and slop floods around us, the challenge is creating things worth existing. Agents need the right tools to craft great things. This helps gen AI agents solve 3 problems: - Extract: produce on-brand assets when users already have a brand - Search: find and retrieve brands for inspiration, from our curated index - Verify: check their own work against, see what's off brand and how to fix it Anyone can try it for free today at tastelabs.com/api Check out the full blog, and case study with @intelligenceco below.
113
58
37
975
193,531
Replying to @tastelabs
@tastelabs offsite! feel so lucky to be building with these 30 incredible humans
18
6
3
188
10,257
introducing the @tastelabs research fellowship! for those excited to work on evaluating model creativity, distilling human preference, and ending slop. you'll get compute, custom data, time with top designers and researchers, and a competitive stipend. in person in SF or remote, full time or part time. check out some of the research questions you could explore and apply here: tastelabs.com/blog/requests-…
23
23
15
387
63,928
Otro evento de LLM Native, buenísimo una vez más 💪🏻 Gracias @PetralliLucas @lgesuelli_p @fpingham por haber creado esta locura hermosa 🙌🏻
3
3
25
1,334
Lucas Petralli retweeted
We always dreamed of having a space like this, what a night, what a community. LLMnative.dev 💙
5
4
1
33
2,022
Ya estan abiertas las inscripciones para el segundo @LangChain Community meetup en Buenos Aires, de la mano de LLM Native ;) Sera en las oficinas de Deloitte el 24/7. Muy pronto anunciamos speakers, no duden en inscribirse!
11
10
3
56
6,391
Who's gonna be at @aiDotEngineer in SF next week? Join @tastelabs across two tracks to hear about data quality in subjective domains and how to measure AI slop!
3
10
70
5,758
Lucas Petralli retweeted
and what better than doing it with people you admire? @tastelabs Argentina team 🔥🇦🇷 @tadeodonegana @lgesuelli_p @PetralliLucas
RL works best when you can clearly define and verify success. Taste, creativity, and design quality are challenging precisely because you often can’t. That’s why I joined @tastelabs , working with top frontier labs on post-training and RL to give AI some taste. Thanks @thaiscbranco_ for this opportunity!
6
2
1
29
3,865
wow, inbound over the last 48h has been insane. pumped to help more companies raise the aesthetic bar of their models and agents. bear with us as we work through requests :) and, we're hiring (even more now)! research, design, eng, ml, product, ops, and more! apply below.
24
11
4
322
20,524
Lucas Petralli retweeted
I am thrilled to share that Taste Labs is live. Grateful to be building this next to @thaiscbranco_, on a mission to make AI feel right, and to end AI slop.
We’re excited to introduce Taste Labs. Our mission is to end AI slop. We’re building the data and infrastructure layer to give AI models and agents taste. And today we’re coming out of stealth, announcing our $18.5M seed funding, co-led by @CRV and @AmplifyPartners AI has nailed objective domains and made it easy to generate anything. But it still feels off. Now, the challenge is judgement. What fits, what feels like you, what’s GREAT. This requires turning a fuzzy, subjective domain into something we can measure and codify. We’re starting with design. There are two sides to cracking this, the foundation model layer and the agent layer: - We’ve already been working with the top frontier labs to evaluate and improve their models, crafting the right post-training data and RL environments. - We’ve also been working with app-layer companies to build the context and verification tools for their agents to produce better, more on-brand, more creative outputs. We want a future where AI feels right. If you’re passionate about this mission, join us!
4
1
14
1,600