Been experimenting with a long-horizon design environment over the past few days. The starting question was simple: Does prescribing the workflow limit what an agent can discover on its own? So the setup is intentionally simple: give the agent a physical product design goal, then expose a set of tools — Taste Machine API for creative exploration, image generation API, web search, terminal/bash, Blender, etc. A typical run is 150+ turns, 30–40 Blender tool calls, and ~1 hour. What surprised me was how much workflow emerged on its own. Beyond research, the agent did structural and lighting analysis, iterated geometry in Blender, locked a direction, then used its renders as references for an image model to push the final visuals further. None of this was explicitly scripted. Another interesting part: if we simply imitate a human design workspace, we naturally think in terms of computer use — Photoshop, Rhino, Illustrator, etc. But the agent mostly used code, search, Blender scripting and generative models to accomplish the same design tasks. It started to form a much more agent-native toolchain. There are still obvious failures. Taste Machine could provide several strong and diverse creative directions, yet the agent would always abandon the more distinctive ones and converge on the safer and more generic direction. That makes judgement feel increasingly important in subjective long-horizon tasks. As agents get better at doing, how do they get better at knowing what’s worth doing?

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!

most of what is considered "taste" (read: design) is in the realm of zero sum signaling games your tasteslop will just be the next AI slop and then your anti-tasteslop will become the next tasteslop taste is defined in terms of slop and therefore can never transcend slop.

Finally got some more time to upgrade taste machine to its 2.0 version, which now runs on a much larger training set than the original version, and it shows some significant difference in its choice of design language. (the fourth image on each is the v2.0 result vs the third being the v1.0 result)

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

Codex grew programmatic policies with no neural nets: max score on Breakout, and SOTA-level scores on MuJoCo. Maybe heuristics were not too weak. Maybe they were just too expensive to maintain. Maybe it's the next paradigm. trinkle23897.github.io/learn…