A colleague of mine has written up a quick explainer on the key features (https://encord.com/blog/segment-anything-model-2-sam-2/). The memory attention module for keeping track of objects throughout a video is very clever - one of the trickiest problems to solve, alongside occlusion. We've spent so much time trying to fix these issues in our CV projects, now it looks like Meta has done the work for us :-)
Those are both great tools. However, there are a number of differences, but the two most prominent are that: 1) Encord Active automatically analyses internal metrics to find the most relevant data and labels to focus on to improve model performance; and 2) it is optimised for the full 'continuous' training data workflow including the human-in-the-loop model validation and annotation.
We run usage-based and tiered pricing, but we haven't gotten around to building out a self-serve "sign-up-with-credit-card" product yet. For all the advances in Stripe and automated billing, these things still take some time to implement for a short-staffed engineering team :-)
We base our pricing on your user and consumption scale and would be happy to discuss this with you directly. Please feel free to explore the OS version of Active at https://github.com/encord-team/encord-active. Note that some features, such as natural language search using GPU accelerated APIs, are not included in the OS version.
Yes - the tool can definitely help with that. We combine the newest embedding models with various other heuristics to help identify performance outliers in your unseen data.
It takes me 2 seconds to switch to a different design tool. Adobe has 30K employees and been around for 30 years. Figma is a small scrappy startup. It seems pretty preposterous to me that the deal will have any meaningful impact on consumers.
They also have an insane amount of traffic - very envious, they'll be able to capitalise on other things as soon as they start to build commercial products people are willing to pay big money for.
How do you plan to expand the functionality? I've been looking for a more modern alternative to OpenCV for a while - especially replacements some of the aged object tracking functions, there's been a wad of new stuff coming out that OpenCV has been slow to implement.
Very interesting -- had not seen that. Indeed, I think they could do with a bit of a makeover. Their ad platform is also terrible, way worse than Twitter's, so it is both good and bad that they're feeling a bit of a pinch I suppose.
Their traffic is rumoured to be down 25-30% after ChatGPT was released, so good to see that they are doing something to keep the community alive in the new era of GenAI.
The docs are still terrible, coupled with AWS / GCP docs around these things it makes it near impossible to get this stuff to work without investing a significant amount of time.
This is by far the best piece I've ever written on the impact AI will have on society & the most articulate response to the hysteria gripping the discourse.
More powerful chips are great, but NVIDIA really ought to focus some of their best folks on ironing out some of the quirks of using their CUDA software and actually getting stuff to run on their hardware in a simpler manner. Anyone who's ever fiddled with various CUDA device drivers and lining up PyTorch & Python versions will understand the pain.
New architectures and models will inevitably replace LLMs, and it'll probably happen pretty quickly. The models we're using today will be outdated in 12 months time. The next big thing will a fusion of LLMs, LVMs (large vision models) and LAMs (large audio models) into modal-modal models.
There are several stories of companies going from $0 to >$5M ARR in just a few months with the boom in LLM-driven applications. The downside for these types of apps is that churn rates are through the roof, and will probs die when the LLM fad inevitably fizzles.
I'd download and run a few open-source LLMs that can be run without internet access and a laptop - can be an infinite source of math problems, quizzes, programming challenges, etc., that should keep you occupied for a very long time.
'Unsupervised reinforcement learning' is how these large models and systems ultimately will end up becoming sentient. We recently tried a similar approach on a toy problem in the computer vision sphere (https://encord.com/blog/we-employed-chatgpt-as-an-ml-enginee...) with pretty decent results.
Segment Anything is orders of magnitude more powerful than anything we've ever seen before within the real of image segmentation and the first 'foundation model' in that regard, so this is a real step-change in background removal/editing, etc. -- hell, we might not even need green screens anymore!
https://paperswithcode.com/ is arguably the best source and overview of all the research. Its also (somewhat) unbiased (owned by Meta), not being an SEO-optimised company blog.
It's not that there's only 10% left to improve. It's that the data needed, compute requirements, and model size are as intensive, getting from 0 to 80 as they are getting from 80 to ~85 or ~90. See https://paperswithcode.com/sota/image-classification-on-imag...