Submind YouTube summaries
Thumbnail for Signals of AI – Peter Neubauer (Staer)

Signals of AI – Peter Neubauer (Staer)

Watch on YouTube

Video summary

Peter Neubauer introduces himself as the co-founder of Stair and a passionate advocate for scaling open-source principles within business models to drive global change. Drawing from his background in communist-era Germany and his work with projects like NeoForge, Mapillary, and OpenStreetMap, he explains how these initiatives successfully internalized open code and data while securing financial sustainability. Currently applying lessons learned at Meta and Mapillary to robotics through Stair, Neubauer emphasizes the goal of making open space more accessible for both humans and machines. He highlights a significant shift in his own workflow where his team stopped writing any new lines of code manually last October, opting instead to rely entirely on AI generation to remove human editing from their internal processes, thereby empowering teams to solve problems that were previously considered too difficult or time-consuming. However, Neubauer warns that this rapid advancement comes with serious risks comparable to the environmental crisis caused by plastic pollution. He argues that while AI is a powerful tool capable of solving countless issues, society lacks the infrastructure and care needed to handle its waste products effectively, leading to an internet drowning in "micro-plastics" such as fake knowledge and hallucinations generated by models like Claude. A critical concern he raises is the inability to verify or audit current large language models for safety when they control physical systems; even major players like Anthropic admit they cannot guarantee AI will not cause harm in real-world scenarios involving humans, trucks, or infrastructure. Consequently, relying on standard transformer-based approaches that simply dump data into a heap and hope for the best is insufficient for critical applications where deterministic behavior and trust are essential. The transition to an era where compute replaces human labor also forces a fundamental reevaluation of economic value and social status. Neubauer points out that if AI centers consume vast amounts of clean energy, society must decide whether this power fuels Bitcoin mining or weather models rather than genuine societal progress, prompting the need for new regulatory frameworks similar to those governing atomic energy. He suggests that as technical implementation becomes automated, humanity should pivot toward "intent" over execution and focus on genuinely human experiences like attending live concerts or engaging in nature restoration, which cannot be perfectly replicated by algorithms. This shift requires redistributing status and financial resources away from obsolete roles and into essential fields such as nursing, education, and ecological repair to prevent a catastrophic collapse of these vital sectors. Ultimately, Neubauer concludes that humanity faces a limited window to navigate this transition before self-improving AI evolves beyond our control or creates an unmanageable dystopia. He proposes embracing the reality that status is no longer tied to technical proficiency but rather to genuine human contribution and intent, advocating for measures like Universal Basic Income (UBI) to support those displaced by automation. By treating compute as a powerful resource akin to nuclear energy and consciously choosing which problems are worth solving with AI versus what requires authentic human presence, society can steer toward a post-capitalistic future where technology serves humanity rather than undermining it. The core message is that we must actively choose how to integrate these tools now, focusing on preserving our unique values while ensuring the benefits of advanced computing reach those in need without sacrificing safety or environmental integrity.
Read the full video transcript
Yes, as Kella said, my name is Peter and I'm co-founder of Stair, but also do a lot of other things. Um, when I go back and look at like what it is I'm actually passionate about is apart from me being, you know, growing up in in communist DD um, and being uh, like a a a convinced communist and agitator for my school before the wall came down. Um I'm mostly interested in big vision uh uh approaches to things. So NEO forj the first kind of like bigger effort apart from all other open source things I founded is is um a way to internalize open open code and open source and open effort into a business model that scales and apparently it does. So, so, so to keep the four freedoms there and and also make this an internally uh scalable concept to put money behind it. Um, Mappiler was a similar approach to to open data, keep the data open, creative common share like, but put money behind it to be able to utilize that to make planetwide change and power other open source uh uh and open data companies like like uh Open Street Map who's producing, you know, some of the best geographic data we have now. Um and um and now um with uh with stair we're applying what we learned at at at Meta and Mapillary to robotics and and uh making um the open space more accessible to to humans and machines. But aside from that, I've been realizing there's other questions and and one big there is the big commons the like you know air, water, biodiversity and so on. Um and uh now sailing up is AI as one of the big uh comments. So I got three questions that I'm actually going to concentrate on um in my work quite quite uh like concrete what's happening with with you know AI um I come from a very hard you know software background what's happening right now is and I think you all see that is that software as a field of work and skill is basically breaking down together with uh open source movements like Communities are getting destroyed by autogenerated pull requests. There's no community in just approving stuff that people put into claude and and and send you a 100 day and now we have you know Claude Mutos finding all these bugs that someone needs to needs to fix. So so I think like software as a line of work and self-defin is disappearing. You have now context that you need to to round trip. It's there. uh we took a decision last October when we started to basically not write a single line of code ourselves. We removed edit buttons for from our internal things. We we're not editing by hand anything. Um this thing is generated also. Um so so um I think we need to start moving into making the team work together in a way that is not just human but even empowering uh your your central context. So we built our own thing for that but you can do it in other ways right but just just to just to say um anybody uh can build anything as as you said like problems that we thought were hard for people are not existing anymore uh and that I think is hugely empowering. Um is that only positive? No. We're creating a lot of stuff that is just burning tokens and and the energy for that is just phenomenal. I would almost say AI is a bit like plastic. We got this super material that solves buttloads of problems and it sterilizes it. It it it makes things I mean we have plastic everywhere and uh however we are not equipped and we are not caring how to handle this stuff. So we contaminate our environment with microplastics with with plastic bags. AI I think is such a super dualized use technology. We cannot refrain from this because we will solve so much stuff right and we know we know we can solve it. I mean there's several companies in here that use AI to go do revolutionary stuff. However, we are spoiling our internet and ourselves with microA with microsop. make our internet basically unusable. We we we drown in like fake knowledge and and that is something we need to learn or we will end up with microplastics. I mean that's that's a serious problem now, right? Um a weak signal. Well, um compute replaces human work. Uh we are now at the point where bigger data centers are actively displacing work. Anybody who sits in front of a computer and works with anything that's remotely logic related coding spreadsheets like anything that is done inside will be automatable. I'm not saying everybody will be automated but but there is there is a big risk that or not a big risk actually a big opportunity. However, um um that is also like point two. We have the case now and we see this in stair and other in other like societal critical applications that we cannot we cannot verify we cannot audit AI to the point where we not just sit and chat GPT where the human is in the loop but actually off the loop. If we want to put a truck into an environment where there's humans, we there's no way right now we can trust any AI not to run into the wall. And there's no way I mean if you if you read Kellis link from Anthropic just coming out like the other day, even anthropic has no idea how to verify to audit and to make AI deterministic. So this is all good as as long as we are sitting here. Oh yeah, I can do other stuff. But but if you really put it into like a watch or something that that controls anything physical, we probably need post LLM models. The whole the whole you throw throw data into a big heap um put transformers on it and hope that something comes out and then discover what else is the side effects of this. Uh is is probably not going to work. Even energy wise, we have a big problem there, right? We we encode the whole world to do an if statement. uh that's that's kind of not what we want to do. Um and um and then yes on the on the positive side, you know, more and more becomes solvable for for some of the biodiversity assessments in in in some of the efforts we're doing. We're actually just taking the drone images and the and the data saving it for next year knowing that next year this will take zero effort together with the next assessment instead of spending some weeks on this. Now we're saying oh we have the base data so we will when biodiversity is is you know computable we will just go back and evaluate that data and it will be much cheaper then so let's not spend the effort now then a future yeah um so so one big thing that comes from this is if if compute replaces human power what happens to the to the uh a value attached to human input. Now if that is compute then we should start treating compute just like say atomic energy. We should make ourselves account for this. If we build data centers up in northern Sweden with clean energy that feed like 600,000 GPUs um what is the what is the you know alternative? Could we do something else for it or are we feeding Bitcoin miners or are feeding weather models? That's a in my opinion that's a big problem. And also the um the effects of this uh are quite profound. If everybody in this room needs to change jobs, what does that cost? And where does this cost come from? I mean the the the benefit has been 80% at anthropic. Everybody here refers to Claude. I mean these are people sitting you know in the US in a in a world where where you know we would we would like to have some control of both the value but also the data that is prompted into these systems right. Um then also I think with this comes the a a a re um revaluing of genuine human values. We we and and and that is a bit what humble said like we need to start treating ourselves not as inferior and do what we talked about before here and do kind of like knowledge maxing. Everybody has their agents. Oh, this is so cool. We can do all this and everybody gets nervous about being dump. what uh maybe we should just embrace this and say, you know what, this is not longer a a a unsolvable or a status problem. Let's concentrate on stuff that's genuinely human life stuff. You go to a concert to experience things that's not perfect. Otherwise, you can just go to Spotify, right? Um and redistribute is not status right now. In order for this to happen like this this this shift in what we deal with more intent and less implementation, we need to redistpute status. We we we work in these in these jobs that are becoming shifting if not say irre irrelevant in some places. And but where we need people like in nursing like in in nature restoration in you know teaching these these branches have less status right now. So so we need to kind of like redistribute money and status into these into these efforts. Um and uh and you know if we do it right, we have this this kind of post post uh uh um capitalistic um possibility to to use UBI and others other measures to to to come to a world that we want otherwise you know we are that we are the knee of of uh self-improving AI right now and I think we have a very limited time span to to make this transition or not. Otherwise, we will just make it uh into something that potentially can can be very bad for us um implicitly or or explicitly. So, that's me. That's all I have.