jasonhong

https://avalovelace1.github.io/BrickGPT/

We introduce BrickGPT, the first approach for generating physically stable toy brick models from text prompts. To achieve this, we construct a large-scale, physically stable dataset of brick designs, along with their associated captions, and train an autoregressive large language model to predict the next brick to add via next-token prediction. To improve the stability of the resulting designs, we employ an efficient validity check and physics-aware rollback during autoregressive inference, which prunes infeasible token predictions using physics laws and assembly constraints. Our experiments show that BrickGPT produces stable, diverse, and aesthetically pleasing brick designs that align closely with the input text prompts. We also develop a text-based brick texturing method to generate colored and textured designs. We show that our designs can be assembled manually by humans and automatically by robotic arms. We also release our new dataset, StableText2Brick, containing over 47,000 brick structures of over 28,000 unique 3D objects accompanied by detailed captions, along with our code and models.

Here's an excerpt of the video where he explains his idea: https://www.youtube.com/watch?v=svl_1upFpQo

My summary: Science is like a hike. You have a general sense of what direction you want to go. You explore things along the way. You might make wrong turns but might also discover something new and interesting. While on your journey, you might spot new mountains or waterfalls or other vistas that look like they might be good to explore.

In contrast, AI is like taking a helicopter to your destination. Yes you got there quickly, but you missed a lot on the way.

Related, Tao later talks about the pipeline. It used to be that seminal proofs were rare, so there was lots of time for the community to review them, process them, share them, and summarize them in textbooks. Nowadays, AI helps a lot with generating more proofs, but not so much with the other parts of the pipeline.

Tao also explains the math pipeline issue in his talk at the International Congress of Mathematicians, also shared on HN last month: https://news.ycombinator.com/item?id=49056620

My overall take: Tao is very thoughtful and insightful about progress in AI and math, and about what the tradeoffs are. We're seeing similar problems in computer science (my field), where conferences are now getting over 10x the number of submitted papers over just a few years ago (not an exaggeration). It's straining the research community in ability to review, understand, and present the papers.

https://www.kompas.id/artikel/when-nobel-winner-in-chemistry... https://www.yahoo.com/news/science/articles/why-nobel-laurea...

Nobel Laureate in Chemistry Omar M. Yaghi joins Tsinghua University full-time https://www.tsinghua.edu.cn/en/info/1244/14984.htm

For example, take a look at Albert Einstein's Google Scholar profile. He's not the top cited physicist. Not even close. It's because other researchers don't explicitly cite his papers. https://scholar.google.com/citations?user=qc6CJjYAAAAJ&hl=en...

Same with Tim Berners-Lee and the World Wide Web. Imagine if his original paper were cited every time someone deployed a web site.

If you want to try it out, we offer some free credits at https://fuguux.com

Any feedback you have would be incredibly helpful! We're considering more kinds of reporting, support for QA testing, better integration with CI/CD, and more.

Note: we don't want to replace real user testing, but rather complement it. With AI user testing, you can get quick feedback on potential usability problems in hours for a fraction of the cost, making it so you can iterate much faster. We advocate doing user tests with real people to understand problems that require domain knowledge or nuance.

Diffusion of Innovations is a widely cited theory explaining why people do or don't adopt any kind of innovation, from boiling water to eating limes on British ships to installing telephones. The concept of innovators, early adopters, and late adopters comes from this theory. More relevant to this post is that this theory posits five factors contributing to adoption, one of which is Observability: you can easily see other people gaining benefit from an innovation. The more Observable an innovation, the more likely it is to be adopted. https://en.wikipedia.org/wiki/Diffusion_of_innovations

The other is Social Proof. Seeing what other people are doing, especially those that are similar to you in some way, can help steer your behavior, often in subtle and unconscious ways. There are studies about how simple signs like "people who stayed in this hotel room re-used their towels" or "most of your neighbors are reducing their electricity usage too" can shift people's behaviors, even without people explicitly realizing it. https://en.wikipedia.org/wiki/Social_proof

My colleagues and I used these concepts in several pieces of research on what we called Social Cybersecurity (joking that the term "Social Security" was already taken). The insight we had was that cybersecurity has very low observability, making it hard for innovations to diffuse through one's social network. That is, I don't know what your cybersecurity practices are, and vice versa, making it hard for best practices to be adopted.

One intervention we did was a large-scale intervention on Facebook to improve observability, showing that simple messages like "108 of your friends use extra security settings" did increase clickthru and adoption rates of those settings. https://dl.acm.org/doi/10.1145/2660267.2660271

We also have many other studies along similar lines, e.g. many triggers for talking about and adopting security are social in nature (https://www.usenix.org/system/files/conference/soups2014/sou...), that security settings that are more social in nature are more likely to be adopted (https://dl.acm.org/doi/10.1145/2675133.2675225), and more.