
Can you RL design taste into a model? I explored training Qwen3.8-27B over some open-source design environments. The answer: you can, but that taste probably won't be yours or mine. And definitely not a real designer's. jasperlu.com/blog/rl-design-…

continually learning. formerly taught models to design @figma, to see @nuro

Can you RL design taste into a model? I explored training Qwen3.8-27B over some open-source design environments. The answer: you can, but that taste probably won't be yours or mine. And definitely not a real designer's. jasperlu.com/blog/rl-design-…

idk I just think it’s funny that every new “scaling law” in llms (reasoning, swarms) just happens to perfectly correlate with burning more tokens

post-training still very much not dead
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…

High-Flyer, the hedge fund behind DeepSeek, has seen 6 of its 9 publicly reported funds post negative ytd returns in 2026 If AI was the difference maker we’d know it by now now trust me How do these guys keep convincing top VCs to invest in them? Top 10 worse punt of the cycle
Today, @bayesiankalki and I are launching Mnemonic Labs (YC F26). We're a neolab training a foundation model for quant trading that learns directly from market activity. In <6 weeks we’ve built our training pipeline and model architecture. Our current training run uses over 25TB of tokens. We met 28 years ago shortly after I was born (we’re brothers!). We’re Olympiad finalists (5x USACO combined) and MIT graduates. Kalki made $20M in annual PnL at 22 at Tower Research and I was the youngest quant at Citadel at 19. Over the past ten years, we’ve been part of quant trading’s evolution from linear models to trees and now neural networks. Learn more about us at ycombinator.com/launches/XRK… or reach out at founders@mnemoniclabs.ai!

It’s been funny playing through the Ocarina of Time remake while at the same time seeing all the Claude slop games on my feed. It took revisiting OoT as an adult for me to really appreciate how much goes into designing a great video game. I’m now convinced it will be a while before AI can one-shot the kind of games millions of people actually want to play. In a way, game development is the ultimate exercise in empathy and pedagogy -- two things llms still struggle with today. At every moment, the developer has to put themselves in the player's shoes: What do they already know? What will they find fun? What are they going to try? Like this post calls out, every dungeon in OoT (except maybe the water temple) teaches you what to do as you go, and it does so in a subtle way, just difficult enough to not feel like a tutorial. Now add onto this that every design choice can clash with every other one in exponentially many ways. And unlike with ui, you can't just spew all the states onto a canvas and inspect them all at once. You really have to play it through to judge if it works…which takes taste, another thing llms are still bad at.
I beat Ocarina of Time as a kid without a strategy guide. Here are a couple of examples of how based on your examples: Phantom Ganon throws a giant ball at you. It hits you. Ow. He throws said giant ball at you again and you try hitting it with your sword, like you would a baseball with a bat, because why not? That works. Now you know. You figure out the Poes in the paintings by realizing they run when you get close and you've just conveniently been given a long-distance weapon (the bow). Its actually all very logical. There. Thats Zelda for you. If you get an item in a dungeon, you'll need to use it in that dungeon. Everything just takes a little thought. The problem nowadays is people don't have the patience. Handholding in games has gotten out of hand, IMO. Zelda is not random experimentation. You just need to use context clues the vast majority of the time.

Jack did indeed Cook
Inkling 2 Small is our first model trained partly in NVFP4, with a mixed-precision recipe that enables a great tradeoff between speed and accuracy. Check out our new post to see how we did it, and the lessons we learned along the way!

On-policy dislillation in papers: - Outperforms direct RL - Dense signal - Solves credit assignment - Sample efficient On-policy distillation in real life:

This is a good example of something I’ve noticed with “ai products”: consumers care more about the appearance of something working than they do about that thing working well. Ai calorie tracking, ai clothes tryon, ai therapy, color analysis… None of the consumer apps for these work to the extent that they’d pass adversarial benchmarks today, but that doesn’t stop them from getting popular.
Putting aside the fact that AI clearly cannot predict this, it’s very scary to me how much some people how little some people notice AI changing their features for the “better” Her figure is not visible in her reference photo! It took her outline and drew another person in it!

We need more public research on pre-RL midtraining of already post-trained models

Wildwood is the best animated movie of 2026
In a day and age where people are pushing more fake art than ever, Wildwood should be screamed about from the mountaintops. It's art at the top of its craft and my favorite movie of the year. Please go see it! You won't regret it.

AI companies should focus more on making tools for humans rather than human-shaped tools
It could be vastly underestimated how much human expertise is needed to diffuse the benefits of AI to the long-tail of knowledge work. I think more data companies should figure out this branding, where experts help AI companies understand how to improve the tools that make them more effective, rather than replaced. Some jobs will be replaced for sure, but I think a lot of knowledge work is deeply interpersonal, messy, and context driven in a way where AI adoption will be a productivity boon to existing employees. We are seeing this with the demand for SWEs today. This rebrand will serve the AI industry writ large very well, but goes against the RSI narratives of the labs as "ASI will replace all knowledge work." In some ways, the difference is more philosophical about "what is the nature of work." I believe knowledge work is much more human than the evals give credit for, and this will make diffusion take a long time. Seems like Mercor can be at the center of this diffusion puzzle. Disclosure: Yes, I advise Mercor on trying to realize this puzzle, as I think it's super interesting.

> no 3D software, no render engine > 1. A path tracer written from scratch in C++. Hmm
🚨 This is made by Fable 5.2. No image model made this. No diffusion, no photos, no textures, no 3D software, no render engine. Every pixel was computed by code Fable 5.2 wrote for this one picture.

1/ Pretty soon we’ll see a great proliferation of slop games and disposable, ai-generated experiences outside of just the bubble here. 2/ This proliferation will, ironically, end up giving existing institutions even more power in dictating what has artistic merit (yes, including videogames as art here).
Second "journey through generated worlds" claude demo I've seen today, both incredible This is beautiful and yet it's slop, beautiful slop Removing the barriers to entry and the effort of implementation will result in a glut of product and an attention economy spread too thin

I think what this shows most of all is that Opus 5.5 and Gemini Carbon likely bought their svg envs non-exclusively from the same vendor. Which is good news for Carbon i guess because Opus5.5 is pretty good. (Or ⚗️)

Orgs are notoriously sticky. Many companies that end up an industry incumbent simply have too much inertia to transform into an ai native company. Add onto this competing egos, pms and managers, and the fact that the ai landscape changes more quickly than perf timelines support, and most of the time you'll end up with a half-baked agent product no one really wants to use.
One of my biggest conclusions from this AI era is that it is easier to take over an entire industry from zero than to convince the incumbents to transform themselves by selling them AI tools. Don’t sell tools, just build the AI native competitor.

> there will be out-of-distribution data in perpetuity as models refactor the economy and create new operating conditions/environments to study. That is, diffusion itself will keep moving the goalposts. Until a few years ago there was no such thing as an AI-native firm. In the operation of an AI-native neofirm, you are not just relying on data from existing firms but creating an entirely new corpus of interactions that has never existed before, and is therefore definitionally out of distribution.
yoni rechtman @yrechtman Nov 6 Article The corpus is the economy Over the last month I’ve done 60+ calls with people up and down the data supply chain: brokers, wholesalers, researchers, frontier lab buyers, RL and data labs leaders, etc. For those of you who are
