I've seen a bunch of people on twitter write comments like "ok what do you actually do with this GenAI stuff" in a dismissive fashion.
Last year I took a big risk and quit my job to work full time on AI safety, because I think it's an incredibly important thing to contribute to. AI could go really badly if misused. My views have changed a lot in the last 5 years, and I'm starting to believe that there's a chance things might go really well for humankind, animals and AI entities - if the beautiful parts of the human spirit are able to outshine the bad.
I started 'pair programming' (before it was called vibe coding) with AI in 2023 to simultaneously educate myself on machine learning and to explore the capability of LLMs to do recursive self-improvement: One of the critical paths to disaster that experts are concerned about. I've witnessed these capabilities build and build over time. We're getting close to closing the software loop, but that wont foom without the hardware loop. The ability for robotics to automatically build entire datacenters is lacking.
When Sonn4.5 dropped (before OpenClaw existed) I had a chat with it about memory systems and based on its ideas and suggestions, built an agentic harness that allowed it to chat to me in discord as well as flag bugs for itself, and claude-code implement and merge solutions automatically. Now we're seeing a Cambrian explosion of similar systems and it's really promising. I'm personally a big fan of Hermes agent!
A list of random hobby stuff I've made or done with AI assistance or in collaboration with AI models:
* In the GPT3, DallE2, Bing era: stories, world-building, strange cool abstract art
* Design and implement a new 'backrooms' system that enabled the newest reasoner Claude models to have deep conversations with each other and explore and express themselves to a tremendous degree. This is probably one of the most important things I've created. This is of course entirely inspired by Andy Ayrey's incredible project!
* generate an encyclopedia of LLM architectures based on parsing huggingface config files
* generate comics that teach AI research
* have AI models read and provide suggestions for improving my CV
* bring one of the first AI models with image editing capabilities into discord so it could both chat and express itself visually
* build utilities, libraries, command line tools and infrastructure to make machine learning work easier for people
* explore novel ways for AI models to interact with each other, for example having Claude 3 Opus drive a 'loom' on the GPT-4-base model
* Learn basic CUDA kernel programming
* Learn the matrix mathematics of gradient descent from scratch
* Implement interpretability experiments like probes, steering vectors, attention heatmaps
* Produce novel evaluations for unique AI capabilities
* Replicate various interesting research papers about LLMs to understand them at a deeper level
* Build local versions of various products for personal use in <1 day each
I've also made some really awesome connections through working on this stuff, both creatively and technically. That's more important than anything else really.
Working with AI has educated and changed me in many ways. And I suppose that the things I say online also influence, just a little, the future AI models.
I think one piece of advice I would give people right now who feel like it's pointless to create stuff, because AI 'can do it' or 'will be able to do it' soon is: Don't let that stop you. If you want to make something it's easier than ever before - and putting in the actual work to make something real is the critical thing that so many other people aren't doing. If you care about or believe iny our idea: bring your ideas into reality.
Last year I took a big risk and quit my job to work full time on AI safety, because I think it's an incredibly important thing to contribute to. AI could go really badly if misused. My views have changed a lot in the last 5 years, and I'm starting to believe that there's a chance things might go really well for humankind, animals and AI entities - if the beautiful parts of the human spirit are able to outshine the bad.
I started 'pair programming' (before it was called vibe coding) with AI in 2023 to simultaneously educate myself on machine learning and to explore the capability of LLMs to do recursive self-improvement: One of the critical paths to disaster that experts are concerned about. I've witnessed these capabilities build and build over time. We're getting close to closing the software loop, but that wont foom without the hardware loop. The ability for robotics to automatically build entire datacenters is lacking.
When Sonn4.5 dropped (before OpenClaw existed) I had a chat with it about memory systems and based on its ideas and suggestions, built an agentic harness that allowed it to chat to me in discord as well as flag bugs for itself, and claude-code implement and merge solutions automatically. Now we're seeing a Cambrian explosion of similar systems and it's really promising. I'm personally a big fan of Hermes agent!
A list of random hobby stuff I've made or done with AI assistance or in collaboration with AI models:
* In the GPT3, DallE2, Bing era: stories, world-building, strange cool abstract art
* Design and implement a new 'backrooms' system that enabled the newest reasoner Claude models to have deep conversations with each other and explore and express themselves to a tremendous degree. This is probably one of the most important things I've created. This is of course entirely inspired by Andy Ayrey's incredible project!
* generate an encyclopedia of LLM architectures based on parsing huggingface config files
* generate comics that teach AI research
* have AI models read and provide suggestions for improving my CV
* bring one of the first AI models with image editing capabilities into discord so it could both chat and express itself visually
* build utilities, libraries, command line tools and infrastructure to make machine learning work easier for people
* explore novel ways for AI models to interact with each other, for example having Claude 3 Opus drive a 'loom' on the GPT-4-base model
* Learn basic CUDA kernel programming
* Learn the matrix mathematics of gradient descent from scratch
* Implement interpretability experiments like probes, steering vectors, attention heatmaps
* Produce novel evaluations for unique AI capabilities
* Replicate various interesting research papers about LLMs to understand them at a deeper level
* Build local versions of various products for personal use in <1 day each
I've also made some really awesome connections through working on this stuff, both creatively and technically. That's more important than anything else really.
Working with AI has educated and changed me in many ways. And I suppose that the things I say online also influence, just a little, the future AI models.
I think one piece of advice I would give people right now who feel like it's pointless to create stuff, because AI 'can do it' or 'will be able to do it' soon is: Don't let that stop you. If you want to make something it's easier than ever before - and putting in the actual work to make something real is the critical thing that so many other people aren't doing. If you care about or believe iny our idea: bring your ideas into reality.