I agree, it was written some time ago, and the language is hyperbolic, but many of the key points still hold true. Building a skillset to reach $100/hour consulting gigs was key to work 2-3 days per week, have time for research, and saving money for an ML rig, while living in Paris. Around $5K MRR is sufficient to live off, $10K MRR is more comfortable and allows renting a few extra A6000 Ada and start outsourcing, and $20K MRR to afford full outsourcing to focus on research and renting 8xH100s. That's a good goal to balance research freedom with other responsibilities. After that, it's optional to trade research freedom for higher MRR and more responsibilities.
Close to Nsawam, in the village Darmang, I was the chief for 2-3 months, and when I left they wrote on my door, "Never forget your king Nana Darmang", I don't know if I'm still considered the chief, although last time I spoke to them it seem that way. Yeah, during my time electricity was on and off, and there was no internet or cell connection. These days, many of them have facebook, which was a pleasant surprise.
Thanks! I saw your earlier question related to working with AI at FAANG, from the context of being 30ish, and self-taught. I was around 27 when I started learning software engineering and AI.
I made a free e-book here (https://emilwallner.gumroad.com/l/no-ml-degree), and I believe most of the key points are still valid, however, when I learned software engineering and ML, ML was a rather small field and tools like chatGPT and claude didn’t exist.
Imo, asking about a curriculum is the wrong framing, for me, it was more about how to find resources to focus full-time and being in an environment that increases my motivation.
I started learning software engineering at home taking courses, but I procrastinated too much to be effective, maybe I did around 10 hours of effective learning per week. For me, studying C at 42 (https://www.42network.org/42-schools/), a free peer-to-peer school was crucial, and I recommend something similar. It enabled me to focus 70-90 hours a week, and after 6 months I was good enough to get competitive startup job offers.
During my time, the FastAI course (https://www.fast.ai/) was the best practical AI course. I'd probably spend a week looking for ambitious projects made by recent autodidacts, and ask them which course they think is best now. And spend max 1-2 months taking the course.
As for picking projects and building a portfolio, the advice in my e-book is still valid. An ambitious but realistic timeframe for landing a FAANG job is 3-5 years. Once you have a solid portfolio, I’d recommend joining say a YC-startup or similar with ex-FAANG employees to get up to speed and references. My first gig was at the YC-startup FloydHub with ex-FAANG employees.
If you are self-taught it’s often easier to get on the FAANG radar by making highly domain specific portfolio projects that are core to their business, or making open-source contributions to their projects. The other route is applying for jobs, however, most people without an ivy-level degree don’t pass the screening stage. If you choose this path, plan for at least 6 month to learn the first part of Ian Goodfellow’s book (https://www.deeplearningbook.org/) using say ChatGPT as your tutor, also grasp the key content in Chip Huyen’s books (https://huyenchip.com/), learn cracking the coding interview, and get good at solving leetcode hard problems.
tldr, most of technology in rural Ghana was considered magic, like phones and tvs, and the only explanation was that it came from "the white man". The chief had recently died and for most of the time I was the only caucasian person in the region, and i met many that had never seen a caucasian person before. The locals started joking by greeting my as the king, the joke picked up and soon most were greeting me as the king. I always though it was a joke, and went a long with it. However, one day I was summed by the neighbouring regional kings and the elders of the village I lived in. After chatting for 30 minutes in the local language, they were convinced that the dead chief's spirit had entered me and they asked in english, "So, do you want to be our king?", I asked what it meant, one person said, I would be given 4 wives and they would slaughter a goat and pour the blood over me, another said I should ask the king of Sweden what it meant, every person said a different thing. I sat there utterly confused, but thought to myself, yolo. So I agreed. I still feel like it was a dream, a surreal experience. The anointing ceremony was a lot of fun, since many of the people in the village were christians, the dropped many of the accent rituals since they were considered unethical under Christianity. There were a hundred or so people, many of the regional kings attended, they offered me a new outfit, a tunic, carved a chair out of wood that only chiefs can sit on, lots of music, dancing, and regional ceremonial aspects. Here's a pic from the ceremony: https://imgur.com/fNd55hB
I end-up doing part-time work for Google at the interaction of Art/Culture and ML doing project like this (https://artsandculture.google.com/story/the-klimt-color-enig...), I saved up enough to build an ML rig (https://www.emilwallner.com/p/ml-rig), since I worked 2-3 days a week, I could spend the rest of my time doing research. I spent 1-2 years working on reasoning, trying different adaptive compute mechanisms and RL on code and mathematics (similar to R1/o1), however, I realised it was hard to compete with the established labs, and if I published my work it was hard to monetize it to have enough time to stop doing consulting work and fund my compute needs.
Instead, I started researching AI colorization, and launched it as a side-project (https://www.reddit.com/r/InternetIsBeautiful/comments/xe6avh...), I ended up having a few hundred thousand users in a few weeks and realized it had enough legs to bootstrap into a company. So I left my consulting gig at Google to go full-time on the colorization project (Palette: https://palette.fm/).
Fast forward to today, Palette is still running with a healthy margin, I’ve outsourced most of the things and I can spend most of my time doing AI research. I’d love to publish and open-source more, but since it becomes too easy to copy, it makes it hard to fund myself and my compute needs.
Indeed, I realized it's hard to compete with PhD students for grants, and subsidising my work with content marketing does not fit my style, and I prefer owning my work and choosing my own research direction. I also want people to use my work, and create solutions that are cost-effective.
So the most logical way was to bootstrap an AI start-up in the area I'm interested, so that's what I'm doing. Unfortunately, it's hard to publish or contribute to open-source, since it becomes too easy to copy, which cuts my margins and ability to fund my research and compute.
Now I spend most of my days doing AI research, and outsource most other parts, really enjoying it :)
not yet, these are great suggestions. it's always a dilemma to add features to mitigate the performance of a weak model, instead of making a better model. most of the problems go away with a better language and colorization model, and many model-specific features are made in vain
ty! i'm glad you enjoy the tool. As for whether or not this tool could replace a professional colorist, I think it depends on the specific project. For some projects, Palette could do a great job of automatically colorizing photos, and for others, a professional colorist would still be necessary to get the best results. Especially when the projects require historical accuracy or a high aesthetic standard. it also makes colorization more accessible, which leads to more opportunities to refine results manually or say print the results.
lol, thanks! onnx, docker, and fastAPI on CPUs with AWS fargate. although i'm switching to GPUs in a month or two, so if you have any suggestions let me know.
thanks for the feedback! it's made out of two models, one model creates a caption and the second model takes the caption and the black and white image and colorizes it. if you click on the edit button you can see the text that generated that colorization. if the text is incorrect, you can edit the text and recolorize it. this often leads to a better result, however, some cases are still hard, especially damaged photos.
cheers! i'm working on a proper privacy policy. I don't store any images that are uploaded. i use google analytics and mixpanel to store user interactions.