hadsed

Well, like most engineering in the real world, people will just test the crap out of it

Having a model that understands physics helps us certify safety. But how much physics is enough? There's a lot to knowing about gravity. You probably don't need orbital dynamics, but you do need to not jump out a second story window.

(Video generation models are an interesting case study: https://www.alphaxiv.org/overview/2501.09038)

This is just way too cynical

There's an exceptional amount of work to get a transformer to learn such a thing from scratch even if you had data from all the different cameras and basements and lighting conditions out there. Mentioning opencv is just silly

The entire point is that robotics lives in the full complexity of the real world and therefore cannot afford to do all that for every little corner of every factory, engine room, warehouse, aircraft, etc.

Robots are finally using ML because ML is finally useful!

The gauge is an excellent example partly because of its mundanity. There's a huge number of these little tasks everywhere in real environments. It does an incredible job at adapting to the specifics of its camera resolution, lighting, and the gauge labels to do the task.

Wow, i had no idea i could use cursor this way

JUST CONVERT THE PDF INTO A STRUCTURED SCHEMA AND THEN LAWYERS CAN BECOME PROGRAMMERS LIKE GOD INTENDED

So what makes this better than Streamlit for example? Either today or in the future

Architecturally this is great. Introducing more infra in an ML stack is a huge pain in the short and long run and I really love that this doesn't do that (or at least, the new infra are things I understand).

In the past I've always opted for a feature store as a library that operates over an existing database/data warehouse/data lake in the offline case, and computing features on the fly in the online case. The internal feature cache for scaling an online service is nicely implemented here using Redis. Bravo, that's probably how I would do it too.

My one bit of feedback is the API. The code just doesn't look nice, out the gate there's a bunch of objects and methods I don't immediately understand the need for. I'm sure they're useful, but for starting out I'd expect a lot more from that interface. I'd suggest something higher level that looks pretty and is easy to understand. That would be my one hesitation.

this catches my eye because it highlights an industry mindset shift to come:

"[...] the concept of integrating power ML methods into the real-time experimental data processing loop to accelerate scientific discovery"

data scientists study the science of what the business does (laundry delivery, manufacturing TVs, tracking patient health), and the point of science is insight and understanding from data to build a theory of how it all works

what this article highlights is that ML can be an exceptional tool for discovery. this is in stark contrast to how ML is usually deployed, which is some big analytics or product effort. the obvious big reason for that is the infra is expensive, the know-how is lacking, and the data sucks. well, that all is quickly changing and we're gonna see folks weaving in ML to bolster their workflows in a much bigger way

great to see academic scientists leading the charge here too. they stand to gain a lot from that perspective

i agree. the overhead of teams and external dependencies is often just too high--you're far better off if you can keep things small but mighty

Beautiful. So many annotation tools focus on "text classification" which assumes you've already got segmented samples. In the real world of documents that's a whole challenge in itself.

Another challenge is that sometimes you're working with PDFs and that means not only ingesting but also displaying. The difficulty is in keeping track of annotations and predictions across the PDF<->text string boundary, both ways.

There are understandably even fewer solutions to that problem because it's a harder UI to build.

LOL. let the car drive itself, but give up music control? over my dead body!

it's funny to me because you'd expect music to be lower stakes and such, but it just highlights that actually driving a car is a much more well defined problem than picking the music i like

i think i believe that. what is the easiest thought abstraction that can be captured by our sensors? well the abstraction is largely defined by the UI. i like to think of it like language. UI components (words) come together to enable complex actions (sentences or thoughts). it evokes questions, like what language does the brain speak in certain contexts for certain outputs? that's gonna be interesting to follow. what if we all think super differently and that makes it hard? i can't imagine why, but i don't have a background in real brains

it may be that our current "AI" tools might be helpful--they're really good at composing "languages" for tying together different types of data. seems that tying noisy brain sensors data to our English alphabet might be an example of that.

would love to see an experiment on this

Seems it's the yearly cycle not matching up to the longer cycle of certain needs that's a problem.

i think focusing entirely on one type of evidence is a little unimaginative. if these guys have data on doctors performing some task with and without their tool, they're in a good place to measure the difference. they can take that all the way to the bank, and to me that would contribute to what id call evidence.

It helps to be able to deliver value at a low bar. This stuff takes a lot of time and energy to improve upon, and it won't "just work". Search is usually a good place to start. Just make one little process a tiny bit better. Make sure you're tracking data very well, because if you're not already at a place to be making data-driven decisions you can't possibly take advantage of machine learning. After all that, just iterate until you've found a better problem to tackle with your newfound capabilities :)

And even if you could, you're getting more value by getting that thing sooner. Or an even better argument: in the end, maybe it took just enough resources at the quality you got to even be born in the current market.

Yeah most things are frivolous... you can survive without a better camera in your phone for another year. But that ham-fisted implicit cultural aspect is what also brings people medical devices 5-10 years sooner than it could have. We all know these gains compound over time. So maybe it's ok if our programs are a little slow. We buy the truly mission critical technology faster with that frivolity.

But I'll also say, I still get mad at my phone and throw it pretty often. Just the way it is.

There's an important thing missing from the HPC community, and that is a focus on software usability.

Maybe things have changed since I worked as a scientist in training at a government lab with one of these crazy machines. But I doubt it.

Trying to use such a computer efficiently was one of the things that took me to my current career as an engineer (and the scientific background landed me in ML), and that was a great thing for me. But for a scientist who is focused on science and not computer architecture, distributed algorithms, programming quirks of working with accelerators (thank you deep learning, because writing CUDA code was painful), they are extremely unproductive when they don't get the help they need.

So I'm fairly skeptical that we're really getting the best return on our investment with Cray and IBM at the helm of supercomputing.

But I have to confess I don't know what I would do to fix it. One thing comes to mind: hire a software tools team at these labs to build a nice software layer to make the rest of the scientific teams more productive. But this is very much what I would do as a company executive, not director of a government science lab. I'm not sure what constraints exist on the finance or politics side. But a change is definitely going to require some leadership.

Great looking app. I tend to setup a table in Notion, and then use the chrome web clipper to add pages to it. Cmd+K and enter is all it takes. Hitting enter again takes me to the page where i can add tags or notes which make them searchable later.

How involved is it to get the training for doing some of these technical diver jobs?

I would say often times this does and should lie with the senior folks on a team. They're the ones who must be able to communicate well since their impact should be broader, perhaps even outside the team. And they should be scaling the team with their own knowledge and experience. I know it's a little different than what you're getting at but i think one person can do both.

One of the most important and underserved aspects of actually doing machine learning is data collection and error analysis (and in software they are essentially the same thing).

Prodigy by Explosion AI (creators of spacy) is very good, great UX focused on making you extremely efficient. It's a paid product but I'm happy to help fund such a talented and impactful team.

That said we don't use any tool out there nearly as much as we could. One fundamental reason is that they don't cover all of our use cases. My team works with legal contracts, and oftentimes the flow has to be: scan through the entire document to find the region you're looking for and drag your cursor over it to highlight. I haven't seen any annotation tool that works for that, so we built our own.

In a similar vein, doing error analysis for those predicted highlights on large documents is also painful. Scrolling here is a chore. If anyone has seen LiquidText in action then you know the solution here.

There's so much left to do in this world. The exciting part is that the better the models start to work, the more interesting UX challenges you have to push efficiency even further. All ML projects largely have to look at the costs of building and analyzing their datasets, and making that cheaper with models in the loop and exceptional UX is super critical and super fun to think about.

so what's the gap here with docker? what incorrect assumptions would i make from assuming this as a model for containers, if anyone knows

The cold hard reality of machine learning is that most useful data isn't readily available to just be collected. Semi-supervised and weakly supervised learning, data augmentation, multi-task learning, these are the things that will enable machine learning for the majority of companies out there who need to build datasets and potentially leverage domain expertise somehow to bootstrap intelligent features in their apps. This is great work in that direction for computer vision.

Even the giants are recognizing this fact and are leveraging it to great effect. Some keywords to search for good papers and projects: Overton, Snorkel, Snorkel Metal

Isn't choosing to not put resources (working hours, investment money) a bet against something?

You should take a look at startups, especially keep your eyes peeled for the growing ones. You’ll have to read the culture to see how they choose managers (promote from within or hire from outside) but opportunities are usually abound and companies just need to get things done. Try to hold out for people above you that you really like and are good, kind folks that want to see you grow. That’s also typically a good approximation of company culture in general, in addition to a good litmus test of whether your bosses can help you in your career progression.

Vectors may be the proper way to represent symbols as they are the link with perception. I would call that a pretty core part of AI.