[treprinum]

Get a PhD in ML from a top school. If you can't, get a MS CS/DS with ML emphasis from a top school, AI grad cert from Stanford at a minimum so that you can understand the latest arxiv papers. If you can't, YOLO and sift through a lot of low-quality articles on the Internet, find the gold nuggets and learn to apply them rapidly and then hope somebody will notice you and hire you. Competition is brutal right now as AI is the only area that is still hiring like crazy. I still think you are 5-10 years too late to start right now. If you can do DevOps, you can likely learn MLOps quickly but it's the same horrible job as regular DevOps. Also, data engineering is not ML but those jobs are easier to find.

EDIT: For downvoters, that's how I did it. I was a very successful SWEng (some of my work was among top posts on HN under different nicks) but saw the ball rolling towards ML in 2012 so I reskilled.

[Der_Einzige]

Citation needed on "competition is brutal right now".

I'm seeing folks with their first and only workshop paper at an ACL track conference landing 150K offers starting at no-name startups. Some of these folks are not even 20 yet. Workshop papers are considered "easy" to publish, and are held in lower regard compared to main conference publications.

If it's "brutal" to compete against folks like this, I think a lot aren't cut out for this field.

[treprinum]

There aren't that many folks who publish even workshop papers. Most folks are scared of academics and hope to raze their way to ML just with dev skills which is unlikely to work as they won't be able to grasp the concepts they need to implement, especially if they work on anything <2 year old. $150k is also on the low end.

[jampekka]

There's plenty of demand for doing ML just by calling OpenAI or similar APIs as more or less total black boxes. Probably moreso than for designing and training your own models. And even then it's mostly taking a pretrained model from huggingface and doing fine-tuning and prompt churn by trial and error.

E.g. doing or hosting state-of-the-art LLMs is more or less infeasible for many/most use cases. (Applying LLMs succesfully for many/most use cases is probaly fundamentally infeasible, but that doesn't mean you can't get paid doing them anyway.)

[treprinum]

Those jobs are quickly getting commoditized - you can see it e.g. on TopTal where these types of jobs had $150/h last year and $60/h this year. But jobs like "create a framework for interpretable transformers based on some DeepMind research" are still at $250/h.

[drstewart]

So 2014, the year before OpenAI was even founded, was too late to get into the ML space? Very interesting take.

[treprinum]

No, but add ~5 years to master it if you are a decent academic performer, then additional few years to learn how to scale it up. Some folks could master it faster but most would likely fail due to the lack of commitment. There is also a bunch of folks that still live in RNN days and cast evil eyes at anyone who uses transformers (hello Jürgen!), so one has to consistently update their knowledge to be relevant (CS25 could help there).

[padthai]

He does not want to be a AI/ML researcher, but a ML engineer.

[treprinum]

Companies are picky. FAANG don't need puzzle solvers but want them anyway.