MLOps leads/lags research depending on your application patterns so it’s an extremely dynamic place to be to see what’s happening
I’d argue based on what I’m seeing with implementations, and importantly how FLEXIBLE transformers seem to be, this is the most true part of this article:
“we’re going to get way further with the Transformer architecture than most ideas in the past”
It's statistics. If you can cast your data to a common format (relatively trivial in cases where they are stored as binary representing numerical values) you can learn from patterns. It is not surprising.
I haven’t been keeping up with the LLM papers because it’s not really my academic interest but I do find them impressive, so maybe this has been figured out but there are two reasons they could accommodate new data modalities really easily: either the sequential data we generate in the real world is more similar than we would have guessed or the hard part isn’t in handling the domain-specific data, but learning to process and predict future signals really well. The former case would be more surprising to me but it is certainly a possibility — most domains have a “language” of sorts, e.g., visual motifs or licks that get passed between musicians. The network could be picking up on those “linguistic” features born out in data it is fed and just needs to alter its vocabulary from words to pixels or whatever. The second case would be my guess. If you have an algorithm that is good at predicting the future based on the recent past, the hard part is done. The rest is just optimizing it for the task (language, sound, video) at hand.
It's extremely impressive that the same algorithm can handle many different types of data with no changes.
If they're ints you can trivially make them floating-point using routines that have existed since those data formats were invented. If they're discrete values you can encode them in ways that make them legible to the models. One of the more recent interesting developments was token embeddings, admittedly, but again this is just an example of taking slightly more abstract representations and turning them into bits representing floating-point values, which has been an established paradigm known as one of the pillars of "feature engineering" since the beginning of the ML field. One-hot encoding is just a special case of token embeddings.
It's amazing that humans figured out how to store data in useful ways, not that models can "figure out" what to do with things they are already capable of ingesting and processing.
How did you get into this? Seems like a lot of places are stuck on the idea if you didn't do it in the past you cant do it now
Take some courses and get some certifications. And also make some serious projects where you demonstrate your capabilities with cutting edge tools.
This is more focused on tools and use of said tools.
Take some trained models, and demonstrate how well you can use them.
Some ideas:
1. Take a cats vs. dogs model, deploy it online. Design an API around it. Document the API well. Create a mechanism to show confidence score, and store low confidence score examples in a database that you can later manually label and retrain the model with.
2. Take a smallish LLM, design a VS code extension that documents your functions based on docstring.
Just demonstrate your basic knowledge in ML, and really good software engineering skills, learn the vocabulary well, and then start applying for jobs. It's much better if you have a CS/EE degree.
Certifications will do nothing for you. The harsh reality is only real world experience doing this stuff at scale will help you understand all the complexity involved. There are tons of people trying to hop onto this train after taking a few online courses and it's making it hard to filter down candidate pools.
I do think they can be valuable if they help you learn the basics and get started on a bigger personal project, but not as something to put on your resume.
The problem is a lot of tutorials just show you how to make a Flask/Gradio website (maybe FastAPI) and call it a day. A lot of the experience here is the sort of in the trenches practical stuff that you can't cover in a MOOC (and it's expensive to experiment with GPU clusters). I suspect there are better non-ML courses people could take though.
A sibling commenter mentions that certifications will do nothing for you. They're not exactly wrong, because what ultimately matters is that you can demonstrate your skills. Certifications and to a large extent even degrees mean very little; what matters is that you convince them you know how to do stuff. The best way to convince people you know how to do stuff is to be able to show a list of cool things you actually did. These courses and their certifications may not mean much on their own, but in the course of completing the courses you will develop skills and capabilities you can demonstrate and talk about in your resume and cover letter.
For me, my work in software and AI specifically predates 2012 - blood sweat and tears of going from non-big data statistical forecasting programs (Bayes nets) to big data forecasting (R, Python stat packages) to geometric vision (SURF, HOG etc) to big data CNN & MDP image processing for CNNs (tensorflow) etc…
Like I said, blood sweat and tears
The trick for breaking into something like this is to produce a portfolio of one or more projects you did where you demonstrate experience with it. This means actually doing it yourself, have a repo with notebooks and text explaining how everything works.
AI is definitely not the first field that is like this, where it at first appears only the people already doing it are qualified to do it. I have had to do this quite a few times over the last 30 years to stay relevant. It takes a lot of work to do this, but it's easier than ever to do today. Today the tools you need to break into almost any technical field can be freely downloaded. A couple of decades ago if you wanted to create a portfolio for something the tools were not freely available. For example, vxWorks for embedded systems programming, or Oracle for demonstrating you can administer large databases, or 3D Studio Max or Maya for 3d modelling: all of these tools were expensive enough to be inaccessible to an individual.
But today, you can go do independent work, take courses and get certifications, and create your own body of work that demonstrates you have an understanding of the field.
If you want to start making your own body of work in the field of AI, I suggest starting with these resources:
1. FastAI
2. Deeplearning.ai. Get a certification and put it on your resume.
3. Karpathy Zero to Hero
4. Re-create the technique in the ReAct paper (Reasoning and Acting).
If you want to demonstrate capability in AI in general, proceed in sequence 1-4. If you want to demonstrate capability with LLM's in particular, proceed in sequence from from 4-1.