yeldarb

I think anyone who says remote is strictly better is fooling themselves. There are definitely down-sides.

We’ve been doing a lot to try to bring back the benefits of being in-person while maintaining the flexibility of being remote.

Wrote about our specific tactics here: https://blog.roboflow.com/remote-not-distant/

It feels like there's been one truly transformational wave every decade or so... but it also feels like there are currently a lot of really important shifts going on.

Are we living in a unique time period where there are truly going to be multiple game-changers? Or will one end up vastly outshining the rest in retrospect? (Just like 3D printing and IoT seemed to fizzle last decade.)

1980s PCs 1990s Internet 2000s Web 2.0 & Social 2010s Mobile 2020s... {AI, Crypto, Metaverse, Synthetic Biology}

(I was going to add quantum computing & nuclear fusion on there as well but IMO their impact is more likely to be further out.)

Congrats on the launch! Really enjoyed using an early version for a hackathon project earlier this year. Hugely valuable to automatically start from something that looks good as an engineer without having to put much thought into it.

A DIY lawn-sized version… but with flames instead of lasers. https://hackaday.com/2021/09/11/dont-sleep-on-the-lawn-there...

It was mostly stuff specific to our specific situation.

We were still on node 12 because, until recently, Google Cloud Functions didn’t support anything newer. But when I installed it, I noticed it was running on x86 via Rosetta. So I upgraded to the latest node (17) which installed as native arm64.

Almost everything “just worked” but there were build errors with installing: webpack, node-sass, gprc, and node-canvas.

Webpack turned out to be an issue with node current (17) so downgrading to lts (which is currently 16) worked.

That seems to have magically fixed the gprc build too.

We weren't actually using node-sass anymore; so removed that dependency.

Upgrading node-canvas to 2.8.0 and building from source (after installing its dependencies via homebrew) seems to have worked.

All in all it took maybe 2 hours to follow all the rabbit holes. And our whole engineering team is going to move over to arm64 as soon as Apple can deliver their machines.

Mine just arrived today and I'm blown away.

After upgrading our app's dependency chain to run native arm64 builds (which took a bit of googling) webpack's incremental build time on our app is down to 116ms from 983ms on my 2016 Macbook Pro. Over 8x faster.

Our Tensorflow.js (webgl) models that previously ran about 8fps in the browser are now running at the camera's native 30fps framerate. (Similar to my desktop 2080ti.)

Curious how it will be priced. Can it be dynamically inserted for specific sessions only? Is there an API to grab a link to the current session’s recording? Can you watch a session “live” like with FullStory?

> What should a Foundation for AI look like? The report largely passed on that question, but we feel it is the heart of the matter. Here are seven elements that we think are indispensable.

> First, a general intelligence needs to...

A model that can serve as a solid foundation for a bunch of disparate tasks doesn't need to be a "general intelligence". This post set up that straw man and then spent most of the rest of the article knocking it down.

I think CLIP is a foundation model for computer vision and I wrote about why here (tl;dr - it's extremely versatile): https://blog.roboflow.com/openai-clip/

Roboflow (YC S20) | Developer Tools | Remote, USA | https://roboflow.com

We build tools that enable developers to give their software the sense of sight.

Over 50,000 developers have used Roboflow to build everything from Blackjack helpers[1] and Duck Hunt aimbots[2] to drywall manufacturing QA and oil pipeline leak monitoring. Over half of the Fortune 100 now has at least one employee with a Roboflow account.

We built an end-to-end computer vision platform (upload, annotate, augment, train, deploy) and recently launched Roboflow Universe[3], which is a public space for users to share their datasets and pre-trained models.

We're growing rapidly and expanding the team to support the ecosystem we're building. There are tons of interesting problems to help solve and you'll be working alongside an early team (of ~10 currently) that is extremely talented and passionate about the problems we're solving.

Roboflow is currently hiring for the following roles:

* Full Stack Engineer (x2) - to work on the core platform & Roboflow Universe.

* Head of Marketing - to help us turn our 1,000 weekly signups into 10,000.

* Community Manager - to amplify our users' stories & grow Roboflow Universe.

* Visual Designer - to work cross-functionally with our teams.

* Account Executive (Enterprise) - to nurture enterprise users into customers.

* Business Development Representative - to focus on mid-market deals.

We'll be opening several other positions soon; so if you're aligned with the vision and think you can help but don't fit into any of the open roles please reach out anyway!

You can learn more at https://roboflow.com/careers

[1] https://github.com/roboflow-ai/blackjack-basic-strategy

[2] https://blog.roboflow.com/computer-vision-win-games-duck-hun...

[3] https://blog.roboflow.com/universe/

A recent 99% Invisible dove into the history of standardized sizing that was really interesting (apparently most of the data on human measurements used in industry comes from US military research): https://99percentinvisible.org/episode/on-average/

Congrats guys!

For those who don’t know, Drapr was a YC co from last summer’s batch: https://www.ycombinator.com/companies/drapr

This is great.

One of our interns built something similar for warding off rabbits in his garden this summer (it triggers a bluetooth speaker to scare them off when it detects them): https://blog.roboflow.com/rabbit-deterrence-system/

He trained a custom model which would be a good next step if you need to detect something that's not in the list of pre-trained classes (or you're getting false positives/negatives with your camera/environment -- or at certain times of day/night).

I recently watched this YouTube about visiting a bunker community in South Dakota that was pretty interesting: https://youtu.be/QW6DIchOpnc

Looks like it’s only 1/10 of the price to buy one of those instead!

Zero-shot means it can generalize to things it's never seen before.

You don't need to train CLIP on any examples of "The stacks from Ready Player One drawn by MC Escher" to have it do a pretty good job of being able to judge which outputs of a GAN more closely resemble that concept[1].

Saying this isn't "zero-shot" because CLIP has been trained on a corpus (the Internet) that had some info about the setting of the movie Ready Player One and the artist MC Escher isn't a particularly useful interpretation. By the same reasoning, humans aren't capable of "zero-shot" classification either because they have prior knowledge of real-world concepts.

The model does object tracking well whether you feed it fish, or playing cards, or sperm because the generalized feature vector it extracts represent all the possible properties. It "knows" (and uses) properties like "facing right", "profile view", "blue sheen", "foggy", etc. to make its determination of which instances are which.

[1] https://twitter.com/braddwyer/status/1430576141572657162

It uses an object detection model (in our example code[1], we used one from Roboflow Universe[2] but you should be able to use any object detection model) to get the bounding boxes and then sends a crop of each detected box to CLIP to get the feature vector that Deep SORT uses to differentiate between and track instances across frames.

This is in comparison to the original Deep SORT[3] which requires you to train a second custom "deep appearance descriptor" model for the tracker to use.

[1] https://github.com/roboflow-ai/zero-shot-object-tracking

[2] https://universe.roboflow.com

[3] https://github.com/nwojke/deep_sort

We could have done a better job of explaining CLIP in this post, but we've written about it quite extensively before[1][2][3][4][5] if you're interested. I'm convinced CLIP is the most important advancement in machine learning in the past year; it's incredibly versatile.

In a nutshell, CLIP is a multimodal model that combines knowledge of English-language concepts with semantic knowledge of images. It can just as easily distinguish between an image of a "cat" and a "dog" as it can between "an illustration of Deadpool pretending to be a bunny rabbit"[6] and "an underwater scene in the style of Vincent Van Gogh"[7] and any other concept you can come up with[8] (even though it has definitely never seen those things in its training data[9]).

This is how these CLIP+VQGAN notebooks can create such a symphony of artistic renderings[10] (CLIP steers the GANs towards its interpretation of any English string that the artist can imagine).

[1] ELI5 CLIP: https://blog.roboflow.com/clip-model-eli5-beginner-guide/

[2] How to Try CLIP: https://blog.roboflow.com/how-to-use-openai-clip/

[3] Content Moderation with CLIP: https://blog.roboflow.com/zero-shot-content-moderation-opena...

[4] CLIP Prompt Engineering: https://blog.roboflow.com/openai-clip-prompt-engineering/

[5] CLIP for Semantic Image Similarity: https://blog.roboflow.com/apples-csam-neuralhash-collision/

[6] Deadpool Pretending to be a Bunny Rabbit: https://paint.wtf/ranking/erFA7/uCzEnPBtKkgoBAhBesQY

[7] Underwater Van Gogh: https://paint.wtf/ranking/xboNk/0fjxp0kEi0WeDgkP3LZX

[8] Other human drawings as judged by CLIP: https://paint.wtf/leaderboard

[9] CLIP-judged Pictionary game: https://blog.roboflow.com/how-we-built-paint-wtf-an-ai-that-...

[10] AI Generated Art with CLIP+VQGAN: https://blog.roboflow.com/ai-generated-art/

We haven’t had a chance to run it through eval on standard datasets yet to see how it compares quantitatively. I’d like to compare it to some of these: https://paperswithcode.com/task/multi-object-tracking

The code is available here if anyone wants to give it a go before we can get to it: https://github.com/roboflow-ai/zero-shot-object-tracking

Yeah, not ideal. We typically re-host images we share on our blog but had apparently hyperlinked these from something the user had shared.

Hey thanks, someone just let me know that this post from a while back ended up on HN; let me look and see if we have those broken images saved anywhere!

Edit: fixed!

By the way, Maxwell is still in high school! It's been humbling to see the amount he's been able to do/learn in such a short period of time at such a young age.

Apple does, and I created a proof of concept for how it might work to guard against adversarially perturbed images here: https://blog.roboflow.com/apples-csam-neuralhash-collision/

How is the human reviewer expected to make that judgement call? They’re not allowed to see the original to compare with for obvious reasons.

Wonder if this was all a publicity stunt to put pressure on their payment processors.

That is a good point; has Apple stated how many bits two images’ NeuralHashes can differ by and still be considered a “match” by their system?

> allows the producers and distributors of CSAM material to ensure that nearly all of the next generation of CSAM will suffer from hash collisions with perfectly innocent images

That’s a really interesting attack vector I hadn’t seen mentioned previously.

Most people are talking about the potential for adversarial images to be sent to users. If they were instead injected into the database itself (either by poisoning real CSAM or social engineering) that would have far wider ramifications.

I wonder what the most widely-saved pornographic images are across iCloud users.

If actual CSAM were perturbed to match the hash of, say, images from the celebrity nude leak a few years back and added to the database then thousands of users could be sent to “human review”. Since the images are actually explicit how would the human reviewers know not to flag them to authorities?

PRs welcome!

We had pretty good success on https://paint.wtf (AI-judged Pictionary game) using CLIP for content moderation. Feel free to probe it a bit by drawing a dick or something on one of the prompts; it seems to do a pretty good job (if it doesn’t work please hit the “Report” button and let us know).

Wrote about it here: https://blog.roboflow.com/zero-shot-content-moderation-opena...

Hey, OP here - while that is also true, I think the average observer would think this number _should_ be zero. The fact that any exist at all is surprising (but the fact that only two exist is also surprising; I expected to find either zero or lots when I started the script).

It doesn't seem like it; check out the adversarially constructed images here. They don't look anything like the original despite perfectly matching the NeuralHash: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issue...

The instructions are here: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX

This was discussed on HN yesterday here (and a few other front-page stories): https://news.ycombinator.com/item?id=28218391