Interesting answers to "What advice do you have for mid-career AI researchers?"
Some great comments on pushing deeper into the tech stack, get closer to GPUs, and keep your eyes open, the next thing can come from anywhere.
Gives me half a thought to how we’ll evolve tech and if NPUs also has the compositional multiplication issues and potential for rounding errors when the GPUs get hot…
Here’s the ones I like.
“What advice do you have for mid-career AI researchers?”
@jeremyphoward @answerdotai @fastdotai Mid-career AI researchers (in fact all levels!): focus on becoming really good coders. Learn to replicate interesting research papers from scratch. Code is the medium we use to experiment, so if you’re better at it, you can run more complex and creative experiments more quickly.
@jacobmbuckman , Manifest AI Get as close to as many gpus as early as possible. Almost nothing else you could do is higher value.
@rronak_ , Google DeepMind Stop studying, build. Go one layer deeper into the infra than feels comfortable, since that’s where the value is. If you’re on Langchain, write the agent loop yourself; if you’re on Verl, write the pytorch yourself; If you’re on Megatron, write the cuda kernels yourself.
@finbarrtimbers , Ai2 People need to know you exist to give you opportunities. Write about interesting ideas you have or things you are thinking about. There is an extreme hunger for “interesting lunch conversation at DeepMind” level content (not hype boi threads, not paper level technical).
@arohan , Anthropic
- Perspective matters more than novelty, many so-called ‘solved’ problems still hide unsolved challenges in the details. Don’t dismiss anything as trivial; breakthroughs are hidden in plain sight.
- Don’t worry about pedigree or fitting in. Hinton bet on neural networks when the field dismissed them. The real breakthroughs come from researchers who ignore the consensus, think for themselves, and tackle hard problems—and that researcher can be you
@danielhanchen , Unsloth AI I would definitely watch MIT, Stanford videos much much earlier - CS231N, do FastAI courses, MIT’s AI course, Gilbert Strang’s courses + CS229
@BlackHC , Google DeepMind Something I tell people: Just try and do things, and if you’re in a place where you can’t innovate or learn, figure out whether to switch. Everything compounds, so it pays off long-term to be picky. Compromise compensation for role fit if necessary. Something I wish I had internalized earlier: When in doubt, double down and work hard. Nothing builds confidence and motivation like working harder and getting results.
@angli_ai , Simular AI The opportunities often come from places you couldn’t have imagined - stay curious, adapt fast, and be willing to reinvent.