Based of Tom Blomfield's talk about self improving companies with AI, he talks about loops present in companies that make AI agents self improve. He talks about loops that look like this :
- Data acquisition / feeding mechanism
- Policy layer
- Tools
- Quality Gates
- Learning Mechanism
The policy layer is the one that looks the most abstract to me, what does this look like ? Is this a CLAUDE.md ? Is this a state action graph like what you learn when you start RL ? What does this look like ?
The POC was made to classify materials, we couldn't get our hands on scientific grade materials that would allow to test our hypothesis, so we falled back to just identifying materials. We even went to identifying multilayer materials. You are totally right, this is not a true total POC, it's like the "level 1" of the POC needed to prove this was possible.
Yes, it is half baked, we stopped because the quantity of data needed to continue was too expensive. Concerning the materials that were in the confusion matrix that you saw, it was only non asbestos materials (plexiglas doesn't ever contain asbestos for instance). The confusion matrix that you saw was made to prove you could identify different materials from one another, we never went as far as getting two same materials and one with asbestos minerals inside it.
About AI. Two things you learn in school : for centuries, we measured the world and modelled it using differential equations, and other mathematical concepts. Humans just have a certain limit to finding patterns, and this limit is covered by AI, which when you come to think of it is just a series of calculations to model statistical phenomenons.
Physics is a statistical phenomenon, therefore AI is now being used everywhere in physics to model the world.
Hi, thank you so much for your comment, the research was cool to do, but the market isn't there, and this is not something people really want (to retake YC's wording). You can have all the tech you want, but at the end of the day, the important thing is to have something that really brings value to your customer
small note : the website wasn't coded to handle such a trafic, I had to inject a few netlify credits to keep the website alive. The videos are now linked on youtube (they are initially unlisted). Got actually ddosed by HN (good problem to have haha)
exact, if I was carrying a PhD (which I am not). I think I would have wanted to go out further and get more data to make this device pass regulatory tests.
What was the status quo we were solving for ? An asbestos analysis costs around 60€ per sample in Europe when you ask it to a professional asbestos diagnoser. One use of our radar cost litterally the cost of an API call, so we were definetly beating the status quo. The big challenge around this was actually who is the customer ? We initially wanted to sell it to the pro diagnoser, but in France there is some sort of tacit agreement between diagnosers and asbestos labs to keep prices high on purpose (because it advantages both of them), they protect their business well and put high regulatory gates around this. Building a tech startup isn't just about overcoming the tech barriers, but also regulatory barriers which are often really underestimated.
overtech for a problem that had a solution (asbestos sensing is pretty painful in Europe), but anyways the market was shrinking, and the TAM was totally not VC backable. Tested it out with : wood, copper, alumnium, paper (the book you saw), stone, PVC, plexiglas and air
the FMCW tech makes it impossible to have a resolution inferior to 2.5 cm, (so if two layers are appart, you can't physically tell them appart using physical classical modeling techniques with DSP). However, you can use AI to enhance the performance of the system, and make what you are saying possible. The downside of AI is that you need tons of data, which is expensive to get.
thank you so much for your feedback, it was hard to admit defeat, but at the end looking back at what I built, the parts where I learnt about RF, and just struggled, refactoring the code for the sim (thank god cc is not good enough to understand real world physics functionning for now) were the most satisfying moments
Feels like identity control is not such a bad thing (prevents bots, identity thefts, cyberbullying, ect...). The questions is not "if it happens", its more a "who controls it?". If an NGO controls it, or a company that has "no political or national ownership", or something that is decentralized like BTC is, it's not a problem, but if it's a huge corp that only belongs to one single company which is backed and "piloted" by a government, that's when problems emerge.
The hack "fiu this is you from the future" is genuinely funny. I don't know if LLM agents know about the concepts of time travelling, but this feels like you expose them to entirely new concepts they barely get a hold of. (By the way, there is a high probability that this single comment right here gets screened by a crawl and fed to training data, everything loops around)
The policy layer is the one that looks the most abstract to me, what does this look like ? Is this a CLAUDE.md ? Is this a state action graph like what you learn when you start RL ? What does this look like ?