I just finished the ML4Good bootcamp in Lyon. Writing this up because when I was thinking about applying, I had a lot of questions in my mind. Maybe this one is useful to someone in the same spot.
Who I am
Indian, based in Germany. Mid-career professional with project management experience, no ML background. Most of what I knew about AI safety going in came from books, articles, and podcasts. The usual. I read the 80,000 Hours book at some point, which is where I first got a sense of AI governance as an actual career track, and later found this bootcamp on their job board and applied.
Why I applied
A few overlapping reasons. The honest version is that I wanted to find out whether the AI risk arguments I'd been reading actually held up when I sat with people who think about them seriously for seven days straight. I also wanted structure, because left alone I just kept collecting tabs. And I needed some real signal for myself about whether AI governance was something I actually wanted to commit to, or just something I liked reading about.
What we covered
The bootcamp is a bit technical, more conceptual.
On the technical side, we went through the full LLM training pipeline. Pretraining, supervised fine-tuning, RLHF, Constitutional AI, RLVR for reasoning. How transformers actually work, tokens and attention and stacked layers and all of that. Inference time setups, agents and tool use, the Model Context Protocol. Evaluations got a full session, including SWE-bench, METR time horizons, and the eval-awareness problem where the model knows it is being tested. We also covered AI control as a research agenda separate from alignment, with the basic setup of trusted versus untrusted models and monitoring protocols. The alignment lecture covered outer alignment, inner alignment, mesa-optimization, and deceptive alignment, with empirical evidence from alignment faking and emergent misalignment papers. None of this goes deep enough to turn you into a researcher. The point is to give you enough vocabulary that the rest of the field stops being a black box.
On the conceptual side, we covered timelines, the political economy of AI governance, public opinion, and the regulatory landscape. The EU AI Act and the GPAI Code of Practice got proper attention, as did California SB 53 and the New York RAISE Act, and how the December 2025 federal preemption executive order is playing out against them. We talked about compute governance, scaling laws, automated AI R&D and intelligence explosion arguments, and the range of positions people hold on how soon advanced AI is coming. The Mythos and Project Glasswing case study came up across multiple sessions. A frontier model with serious offensive cyber capability, a defensive deployment programme, a standoff between Anthropic and the White House, and a clear demonstration of capability cycles measured in months running against policy cycles measured in years.
The last day and a half is for a self chosen project. I used mine to take concrete next steps. Fellowship applications and starting to write this.
What surprised me
The cohort was more mixed than I had expected. There were people with strong technical backgrounds, sure, but also lawyers, ex consultants, a couple of policy researchers, and a few people from non Western contexts.
The technical sessions were more accessible than I had feared. When someone sits you down with a transformer diagram and walks through it slowly, you can follow it. You will not be implementing one from scratch by the end, but you will be able to read papers that talk about them and not panic.
The governance discussions surprised me most. I had expected a strong "AI is dangerous, must regulate" line throughout. The reality was much more open. People disagreed about timelines, about which interventions actually matter, and about whether the field has gotten too aligned with one particular philosophical tradition. There was room to think.
The week was not all sessions. One evening we watched the AI documentary together, which turned into a long discussion that ran past midnight. Another evening was a PowerPoint party, where people presented slide decks on whatever they wanted, the more absurd the better. And other evenings, chill long walks in the beautiful nature around. I had not expected that kind of thing at an AI safety bootcamp, and it was one of the parts I enjoyed most.
What I struggled with
Some of the deeper material went past me. But, by the end of the bootcamp I was well able to connect the dots and develop my thinking. But if you are hoping to leave able to do alignment research, seven days is not enough on its own. Obviously.
The intensity is real. Seven days of structured content, late dinners that turn into late discussions, and a project to deliver.
What I am doing next
A few concrete commitments I made during the bootcamp. I am applying to the Talos Fellowship for the EU AI policy track. I am continuing to apply to AI governance roles, especially where my project management background actually counts for something. And I am writing more in public, starting with this post.
The bigger question is still open. Do I become a full-time AI governance person, or someone who works in a related field and contributes from the side? The bootcamp gave me enough to start asking it properly.
Who I would tell to apply
If you are mid-career and worried you do not belong, apply. You will probably be fine.
If you do not have an ML background, apply, but go in knowing you will be reading more than building, and that is okay.
If you are more interested in governance than research, the bootcamp still gives you the technical literacy you need to engage credibly. Worth the seven days.
If you already know you want to do alignment research and want a deep technical program, this is probably too broad.
Closing
The bootcamp did not turn me into an AI safety expert. It was not trying to. What it did was give me enough working knowledge, enough people I now actually know, and enough quiet time to think honestly about the next steps. That is a lot to come away with.
I just finished the ML4Good bootcamp in Lyon. Writing this up because when I was thinking about applying, I had a lot of questions in my mind. Maybe this one is useful to someone in the same spot.
Who I am
Indian, based in Germany. Mid-career professional with project management experience, no ML background. Most of what I knew about AI safety going in came from books, articles, and podcasts. The usual. I read the 80,000 Hours book at some point, which is where I first got a sense of AI governance as an actual career track, and later found this bootcamp on their job board and applied.
Why I applied
A few overlapping reasons. The honest version is that I wanted to find out whether the AI risk arguments I'd been reading actually held up when I sat with people who think about them seriously for seven days straight. I also wanted structure, because left alone I just kept collecting tabs. And I needed some real signal for myself about whether AI governance was something I actually wanted to commit to, or just something I liked reading about.
What we covered
The bootcamp is a bit technical, more conceptual.
On the technical side, we went through the full LLM training pipeline. Pretraining, supervised fine-tuning, RLHF, Constitutional AI, RLVR for reasoning. How transformers actually work, tokens and attention and stacked layers and all of that. Inference time setups, agents and tool use, the Model Context Protocol. Evaluations got a full session, including SWE-bench, METR time horizons, and the eval-awareness problem where the model knows it is being tested. We also covered AI control as a research agenda separate from alignment, with the basic setup of trusted versus untrusted models and monitoring protocols. The alignment lecture covered outer alignment, inner alignment, mesa-optimization, and deceptive alignment, with empirical evidence from alignment faking and emergent misalignment papers. None of this goes deep enough to turn you into a researcher. The point is to give you enough vocabulary that the rest of the field stops being a black box.
On the conceptual side, we covered timelines, the political economy of AI governance, public opinion, and the regulatory landscape. The EU AI Act and the GPAI Code of Practice got proper attention, as did California SB 53 and the New York RAISE Act, and how the December 2025 federal preemption executive order is playing out against them. We talked about compute governance, scaling laws, automated AI R&D and intelligence explosion arguments, and the range of positions people hold on how soon advanced AI is coming. The Mythos and Project Glasswing case study came up across multiple sessions. A frontier model with serious offensive cyber capability, a defensive deployment programme, a standoff between Anthropic and the White House, and a clear demonstration of capability cycles measured in months running against policy cycles measured in years.
The last day and a half is for a self chosen project. I used mine to take concrete next steps. Fellowship applications and starting to write this.
What surprised me
The cohort was more mixed than I had expected. There were people with strong technical backgrounds, sure, but also lawyers, ex consultants, a couple of policy researchers, and a few people from non Western contexts.
The technical sessions were more accessible than I had feared. When someone sits you down with a transformer diagram and walks through it slowly, you can follow it. You will not be implementing one from scratch by the end, but you will be able to read papers that talk about them and not panic.
The governance discussions surprised me most. I had expected a strong "AI is dangerous, must regulate" line throughout. The reality was much more open. People disagreed about timelines, about which interventions actually matter, and about whether the field has gotten too aligned with one particular philosophical tradition. There was room to think.
The week was not all sessions. One evening we watched the AI documentary together, which turned into a long discussion that ran past midnight. Another evening was a PowerPoint party, where people presented slide decks on whatever they wanted, the more absurd the better. And other evenings, chill long walks in the beautiful nature around. I had not expected that kind of thing at an AI safety bootcamp, and it was one of the parts I enjoyed most.
What I struggled with
Some of the deeper material went past me. But, by the end of the bootcamp I was well able to connect the dots and develop my thinking. But if you are hoping to leave able to do alignment research, seven days is not enough on its own. Obviously.
The intensity is real. Seven days of structured content, late dinners that turn into late discussions, and a project to deliver.
What I am doing next
A few concrete commitments I made during the bootcamp. I am applying to the Talos Fellowship for the EU AI policy track. I am continuing to apply to AI governance roles, especially where my project management background actually counts for something. And I am writing more in public, starting with this post.
The bigger question is still open. Do I become a full-time AI governance person, or someone who works in a related field and contributes from the side? The bootcamp gave me enough to start asking it properly.
Who I would tell to apply
If you are mid-career and worried you do not belong, apply. You will probably be fine.
If you do not have an ML background, apply, but go in knowing you will be reading more than building, and that is okay.
If you are more interested in governance than research, the bootcamp still gives you the technical literacy you need to engage credibly. Worth the seven days.
If you already know you want to do alignment research and want a deep technical program, this is probably too broad.
Closing
The bootcamp did not turn me into an AI safety expert. It was not trying to. What it did was give me enough working knowledge, enough people I now actually know, and enough quiet time to think honestly about the next steps. That is a lot to come away with.
Thanks to the ML4Good team and the cohort.