Ask HN: Getting into AI?

If you’re excited about Stable Diffusion, why not join the Hugging Face Discord and participate in their community?

If you want to learn A.I. try https://www.fast.ai/ Its probably one of the best places to start. Highly recommended.

> as in reality many of them were just using some basic ML models and calling it ground breaking AI to hype investors and potential hires

Without Phd in ML your chances for any leading AI company to hire you to work on SOTA ground breaking models as researcher are near zero, unless you are fine with devops/support roles. As to normal work of ML engineer on not ground breaking stuff it's actually really boring. You’ll spend a lot of your time just collecting + massaging data, slightly tweaking parameters to see tiny incremental improvement in your model.

A full career switch will likely be hard. For everyone working on stable diffusion or GPT3 there probably 1000s of people working on rather mundane tasks in ML (nothing wrong with that).

If I were you I would leverage your existing knowledge in robotics and see how AI might apply there.

For example, computer vision will very likely be a major part of robotics. I could also image that there will be Speech2Text models where you can just tell a robot what to do. However, the main driver here is not necessarily the AI model but the people understanding how to use it in the context of robotics.

I would argue that plain AI/ML is highly scientific work and unless you have a PhD in ML or a similar field it will be extremely hard to get into it. Applying models, tweaking them for your use case and all the work around that will create the actual value in most applications.

You can do a PhD in the field, depending on your background this wouldn’t necessarily mean that you had to give up working. There are many companies interested in „Industry PhD“ and your experience and background in robotics would probably be a great fit at many places. To give you an example I recently interacted with someone that was doing a PhD and was working for Stabilo (a German pen maker).

Here is a random PhD job opening in Stockholm, that I just got send. I know the professor he would be very happy if someone with industry experience showed up:

We have a PhD opening in my group, addressing “event-based vision in challenging automotive environments”. https://kth.varbi.com/en/what:job/jobID:560761/type:job/wher...

My 2 cents is that don't learn anything now.

Apply to SWE roles in companies that specialize in AI or ML. Tell them you are a veteran SWE and the work they are doing excites you. Find a place where you can keep using your SWE skills while you get paid to learn AI or ML concepts on a practical level.

There is a huge demand for SWEs who specialize in specific industries and are not considered more generalized SWE. So first try to become the SWE with an AI interest then look into AI focused roles.

If you are interested in on cutting edge research oriented work like development of models like GPT-3 and Stable Diffusion, then you should know that such opportunities are rare and limited to large FANG type companies. And you need to be super talented with a PhD from a top tier college.

The next group of ML practitioners are data scientists or ML engineers in smaller technology companies and startups. Their work involves applying advancements in these areas to improve their products or services. Very rarely do they do research oriented work. You could look into applying for such roles. An ML engineer role would be more suitable given your background. My advice would be to get your hands dirty and try to build a product on top of some ML model. It would give you some experience in model deployment and MLops type activities which companies often look for.

>For the past year I've been looking at all the progress happening in ML/AI and each day I'm more convinced that there's a lot of game-changing stuff that will come out of it (what we're seeing with Stable Diffusion and GPT3 are some examples of this).

Wow, I must admit that out of all big branches of computer science AI/ML is the least exciting for me. I don't know, but all that unreliability is just putting me off.

I do agree that it's better to have something automated in 90% instead of 0% or 40%, but the impossibility to getting it to 100% is annoying.

It's been like over a decade of huge hype on AI/ML and yet I feel like biggest applied AI/ML that affects my life directly or indirectly is search & ad industry or things like chat bots, but it's mehhh.

I don't believe in autonomous cars based on computer vision

Check out https://www.synthesia.io/careers#jobs

> basic ML models

Disclaimer: I don't work in the game changing language model stuff, so maybe I'm the wrong person to answer

But I'd kind of assume basic ml models are one of the essential prerequisites, so go read Elements of Statistical Learning if you haven't already?

Not a generalized advice, but once you start learning, please check out the Sci-Kit documentation. It's very concise in its explanation of concepts.

You essentially have two choices:

- Get your PhD and try to become a researcher at one of the labs. Super crowded now, but basically the only way if you want to do research.

- Become a ML engineer, data scientist, etc. While you will be working with ML models and understand them to some extent, you may find that it's not exactly what you wanted. You'll likely end up spending most of your time on the same old engineering stuff: Calling into black-box APIs, data engineering, iterating on hyperparameters, building experiment pipelines, dealing with cloud scaling and setting up GPUs, etc.

For example, if you take a look at GPT, 99% of the work that has gone into it is standard engineering/scaling work, not AI-specific work.

what you’re probably looking for is becoming an AI app developer, leveraging inference to create new types of products. No need to become an expert in ML and training, to start with. That’s similar to what some people did about 12-15 yrs ago pivoting from web development to smartphone apps development.

If you get into it now, you'll probably be on the losing end of a pork cycle [1].

All of the hype is creating overinvestment on the side of "producers" of AI. All that overinvestment will mature at roundabout the same time. When it all hits the market at the same time, they'll have to fiercely compete with each other at the same time as having to deal with "reality" kicking in, i.e. learning the difference between hype and real demand to create real value for real paying customers. There will be massive oversupply.

You'd have to find some way to be short that thing, i.e. to somehow take the other side of that trade.

You want to be on the receiving end of that investment with no exposure to the crash that will follow (if any). For example, if you had an AI background now, you could start an AI school. Your customers would be people taking the hype at face value. You'd take their money now, but when it later turns out that the skill isn't worth in the job market what they thought it would be, you're not exposed to that. ...that's what acting school does for wannabe Hollywood superstars. Running an acting school for wannabe stars is definitely a better business than trying to actually be a star.

[1] https://en.wikipedia.org/wiki/Pork_cycle

SWE ~20 years, curious I have spend like 4 days to learn and do some hacking around that Stable Diffusion hype :) Well, there are three types of "boring" work I discovered there:

- researchers - math and stuff, optimizing training, inference, reading tons of research papers, writing papers, implementing low level algos, a lot of trial and error work - that is like 0.0001% of the entire community,

- integrators - building workflow front-ends, wrappers, PS plugins, optimizing the stack, doing some Dreambooth training, tinkering with high-level Python, 0.1% guys,

- users - running one-click tools or just clicking thru some AI-art generator webapp, participating in reddit d* size contest, spreading the news, writing sensational articles for mass audience, hype preachers on YT, designers augmenting Photoshop skills, the best are trying to monetize the tutorials and AI-artwork, 99%.

All that reminds me a bit of "blockchain" career hype cycle which is now fading. For now it is all "vitamins" (vs painkiller).

I will definitely play with more of it on the side. But for now I am back to my boring and stable devops/backend web job, I see a lot of K8s pains to kill :)

Even ten years ago this would have been a difficult career switch without a serious (graduate school level) investment in the very specialized mathematical background of deep learning models. Nowadays its even harder. The mathematics haven't changed much, but the easy pickings are long gone. Now you'd have in addition to target specifically a sub-domain, with all its associated context and heuristics, whether that is text, image, audio or more exotic stuff like protein folding.

Nevertheless the above caveat concerns the overhyped "AI/ML" space. Digitization, quantification and automation of information flows is much more general phenomenon and with a more modest investment in statistics / data science you can be part of this general trend of productionizing "analytics". Just don't expect bubble era FAANG salaries. These were the product of very specific conditions.

If a general data science transition works out for you and you are still interested in the AI/DL/ML bandwagon after you are in better position to understand why and how it works you could easier drift into that space later as it is just an extremely specialized subset of that world.

The other job is data processing for AI.

Collect data from field, cleanse, and then feed to ML. Lots of work. Especially the organizing, cleansing side, which can be pseudo automated upto a point and then manual process the rest.

This data collection for ai already has a large market. And if you want to get involved in ML/AI, either you need to buy data from the market, or do the heavy lifting of collecting processing by yourself.

I found the Fast AI course to be a great way to learn the foundations. It was recently updated after some years too:

https://course.fast.ai/

> In the past I've been wary of companies throwing AI words around, as in reality many of them were just using some basic ML models and calling it ground breaking AI

What made you change opinion? Afaik, nothing has changed in the field except the amount of hype surrounding it.

Learn what ML and AI "really" is (start with neural nets, simple classifiers, get into GANs etc.) and just experiment with PyTorch or whatever else you like. Pick up Ray at some point and just go on from there. Rebuild models from papers and understand them. I wouldn't recommend "toying around" with ready to use stuff like premodeled and/or pretrained models (especially stable diffusion, the hype is really big right now) right at the beginning because that won't really teach an understanding of what is going on exactly. Compare with "script kiddies" as people called them back the days: They just use ready-to-use tools but have no idea what is happening and hence never surpass those tools or the level they are on, unless they bother to actually start at the beginning some day after all ... Like just using Linux won't make you an OS developer. What's all that generated prompts and content worth if you can't even tell at first glance if 0.05 or 0.0003 is a better learning rate for Adam optimizer?

For many problems large companies have, decision trees or SVMs work just fine :) The rest is a huge hype, were a hyped solution searches for problems, like blockchain.

This is an excellent and not-often-suggested course: https://bloomberg.github.io/foml/#home

I would like to create one that helps me write. I tried a few and they were very superficial.

Something like I supply an article and the AI optimizes it with well known techniques that help learning the conten, etc.

You can make the transition, there's demand for all kinds of roles along the spectrum, likely in different companies though. Start working closely with the ML team as a SWE would be a start, then take over some of the pipelining stuff, and so on. Just be realistic with it being precisely that, a transition, and that you will have to apply yourself and take initiative. No, probably you won't become a Google AI researcher, but probably that's not your genuine desire anyway.

I switched from being a SWE of ~10 years to a Research Engineer at DeepMind, so while it is difficult (as other comments have noted), it is absolutely doable. I wrote a blog post about it here: [EDIT: fixed] https://medium.com/@kfedvanilla/switching-from-software-engi.... Happy to answer any specific questions you may have beyond what's in the blog!