Meta introduces Segment Anything Model 2

https://ai.meta.com/sam2/

by bambax • 2 years ago

274 95 2 years ago

ai.meta.com

gnabgib• 2 years ago

Discussion (720 points, 1 day ago, 130 comments) https://news.ycombinator.com/item?id=41104523

cubefox• 2 years ago

Surprising how quickly big news disappear from the front page. Hacker News seems to be optimized for people who check the website multiple times per day.

pjmlp• 2 years ago

Gotta to do something between CI/CD deployments. :)

buran77• 2 years ago

I'm also periodically surprised that topics that seem to be of core interest and very popular on HN just drop off the front page in a matter of hours. Meanwhile discussions like this one [0] (which stayed on the front page with not even 20 upvotes and 4-5 comments for almost 20h before some discussion picked up) seem to artificially stay on the front page for a day or more.

It's probably less algorithmic and more arbitrary.

[0] https://news.ycombinator.com/item?id=41067079

tgv• 2 years ago

Try something like https://news.ycombinator.com/?over=100. It decays less quickly (at the expense of lesser scoring, recent posts).

liamYC• 2 years ago

I built an email notifier for alerting me when there is a popular news article which I defined by X number of comments. I think forums like hacker news are a really cool indicator for an emerging technology, an opportunity, or some exciting tech story. I get a tonne of false positives though, mostly articles that are inflammatory get sent to me. However, crowd strike triggered it as did the release of chat gpt so there are some hits

temp_account_32• 2 years ago

the algorithm that fudges the order of the front page can be so annoying sometimes, i accidentally refresh the page, and a link that was top 10 on the front page a few minutes ago is now buried in page 3 suddenly

greenchair• 2 years ago

It probably has more to do with employees working at the company downvoting the thread.

d3m0t3p• 2 years ago

it says that the released the code, but I couldn't find any except some example code, Did they release the training code ?

Sayrus• 2 years ago

I think that repository is what they are talking when they say "releasing a pretrained Segment Anything 2 model, [...] and code": https://github.com/facebookresearch/segment-anything-2

moffkalast• 2 years ago

Now we just need sam2.cpp ;)

null_investor• 2 years ago

Meta is killing it. Google seems to be lagging behind them in AI research and useful things that is shared within the community.

I'm sure this, LLAMA and the other projects that they have will help drive up new creations, companies and progress.

I'm also sure that this kind of openly sharing code and research will drive up business value for them. It may be hard to say right now what it is, but I'm sure it will.

That's the difference of a founder-led company vs. market led. Google is mostly concerned on short-term goals so they don't report a single bad quarter or has too high CAPEX on a project without profit on its sight (like VR).

Once Meta finds the killer app for VR, all the other companies will be so many years behind that they will need to buy software from Meta or not take any market share in this new space. Similar to what happened about AI chips and Nvidia. Nobody was investing enough on it.

sverhagen2• 2 years ago

I know nothing here, but isn't that also the difference: Meta is just trying out things, and can find that killer application later. Google meanwhile existentially feels that search must be the killer application, and tries to shoehorn everything into that. And in doing so put the bar for success so, so high, ignoring where things are really at.

SXX9• 2 years ago

Problem is that Google core product is not search, but ads. Over years they were slowly degrading search experience because making search better, faster and more efficient means we would spend less time watching ads.

Situation is the same on YouTube where Google have insane market monopoly, but still depreoritize actual creators who made their platform successful to appeal to brands and push more and more ads. Demonetization of everything ever slightly controversial. Somehow Google index whole internet 100 times a day, their OS run on majority of smartphones, their browser is monopoly, they know everything about everyone, but can't even filter spam efficiently.

I never liked or used Facebook / Instagram much, but at least Meta core products are not deteriorating at same insane pace Google products are. I can still use WhatsApp daily without ads being fed down my throat.

maximus-decimus2• 2 years ago

If Meta doesn't care for money, then I really don't understand what Facebook has devolved into. It used to be a chronological feed of what your friends are doing. Now all I see is an infinite steam of anger-bait, thirst traps and ai-generated spam. How can it still be controlled by the same guy?

logicchains• 2 years ago

Now all I see is an infinite steam of anger-bait, thirst traps and ai-generated spam. How can it still be controlled by the same guy?

The stream isn't controlled by Zuck, it's controlled by AI, an AI trained to "give the people whatever they want" to maximise engagement. The more deep learning you put into a product, the harder it is to directly control the output.

Imagesegmetanto4• 2 years ago

Facebook has nothing besides their social network. They care very little about a lot of stuff (after all they didn't care about fakenews and other things).

The metaverse was a huge waste of money.

Their AI Character thing is weird.

They cater to a lot of small businesses, thats definitly helping them.

But besides that? They have very little to do what Google, Microsoft, Apple and co are actually doing.

raincole• 2 years ago

That's a lot extrapolation from a few open models.

nolist_policy4• 2 years ago

Google Gemini recently enabled 32k context in their free tier.

SXX9• 2 years ago

Yet they still locking access to reading files from Google Drive to some "advanced" plan or might be it's not implemented at all. Just 6 months ago or something ago "Bard" could do it and it was one of few actually useful features to summarise PDFs.

I really dont want to have any dependencies in my workflow on Google products and features because they'll surely break them.

nabla92• 2 years ago

Google still leads in AI research. They are doing the opposite of short term. The reason why it does not seem like it may be because most of their work is either basic research or related to chemistry, physics or things that are not public like Facebook.

They are behind in productization of the research. So far they seem to do the minimal effort from the trained model to product.

nolist_policy4• 2 years ago

Also don't forget Waymo.

markovs_gun• 2 years ago

Yeah I have noticed all of the actually useful ML papers I have read come from Deep Mind and Google because they seem more interested in stuff that actually matters like industrial control systems and other boring stuff that makes money

ryukoposting2• 2 years ago

What should we take away from the AI Overview snafu? Is it an example of an ill-conceived concept for an AI product? Or, is it an example of a technically inferior model hampering an otherwise good product?

nsonha• 2 years ago

why aren't they a research institution then? We live in a funny world where academia is ridden with paper mills and fail to do its job, and a for profit company like Google is doing academia's job and... fail to do its job on applications and commercialization also.

jrnx• 2 years ago

they are both, they have a different focus though. E.g. AlphaProof and AlphaGeometry 2 are "killing it" also in their fields.

But for consumers Meta seems to have a lead indeed...

michaelt• 2 years ago

That's the difference of a founder-led company vs. market led.

Are they really so different?

Facebook throws things at the wall, even expensive things without a clear path to profitability, such as Llama.

Google throws things at the wall, even expensive things without a clear path to profitability, such as waymo, google glass, google fiber, stadia, and everything on https://killedbygoogle.com

Facebook made a big companywide re-orientation to deliver a vision that flopped - the metaverse.

Google made a big companywide re-orientation to deliver a vision that flopped - google plus.

Facebook renamed themselves to Meta. Google renamed themselves to Alphabet.

Facebook has an AI research division founded by a French-American computer science professor and Turing award winner. Google has an AI research division founded by a British-Canadian computer science professor and Turing award winner.

Facebook released a widely used open source Python ML library with a camelcase name, PyTorch. Google released a widely used open source Python ML library with a camelcase name, TensorFlow.

Perhaps they're following the same playbook, it's just luck that Facebook's gambles paid off recently.

killerstorm2• 2 years ago

Well, Meta is not as arbitrary with closing projects. Google closes things which are getting traction, that's a really disturbing aspect of interna politics and incentives

Also you mixed up tensorflow and pytorch.

killerstorm2• 2 years ago

Yeah, Google Deepmind is doing useless stuff like protein folding and math.

While Meta excels at chatbot of the day

joelanman• 2 years ago

I still find Gemma 2 9b to be better than Llama 3.1 8b often

dontwatchthis• 2 years ago

Totally agree with you

petters• 2 years ago

I don’t think this is true at all. Alphabet has Waymo and Isomorphic Labs, trying to commercialise solutions to problems with very, high impact. Time scale is at least a decade

andrepd2• 2 years ago

Once Meta finds the killer app for VR

I like that it's taken as a premise that billionaires do have some kind of superhuman vision, that it's not even a matter of "if" but "when" Zuck's VR gamble pays off.

bradleyjg• 2 years ago

* That's the difference of a founder-led company vs. market led. Google is mostly concerned on short-term goals so they don't report a single bad quarter or has too high CAPEX on a project without profit on its sight (like VR).*

It really depends on the founder. Some are very unhappy to see the stock price drop even if they don’t really need the money this second.

Also, it’s a mixed bag. I personally think Zuck is wrong on VR but right on AI.

apwell232• 2 years ago

Not sure how meta is killing it. Their AI integrations into their apps like whatsapp, instagram are beyond useless. They were stuffed in there just to fool markets into beliving meta is some sort of ai player.

That's the difference of a founder-led company vs. market led.

Zuckerberg is the most unimaginative CEOs of the them all. Meta never had a single original product ( save for portal device). All their products are acquisitions. Meta is phenomenally bad at innovation.

Zuckerberg seems to have pulled some sort PR campaign to cleanup his image. But we all know FB is still a shady company run by a shady man, nothing has changed in its rotten core. It was fined billions of dollars just this week by Texas.

Meta is far from a "founder led company" . All the founders of the apps he buys leave immediately and are in turn run by managment consultant types like Adam Mosseri .

Once Meta finds the killer app for VR

Sad to see that ppl are still buying into metaverse hailmary zuckerburg scammed to convince markets that meta is still an innovative company amid slowing user growth. I still don't understand how metaverse fraud pulled by zuckerberg wasn't some sort of SEC violation.

1d22a• 2 years ago

Meta can still very much be an innovative company regardless of user growth though, right? They are very much innovating in the AI research space, and are absolutely leading in open* source research/models, even if their AI integrations are flopping.

Sure, none of this research might be obviously contributing to their products in a share market sense, but it's still innovation which is contributing to the (entirely separate to their "product") AI research community.

Adrig• 2 years ago

Meta has insane distribution. Even if they don't match the best AI integration in a couple of months, they can push it to 1 billion people in days when it hits.

Same thing for AR/VR, they can afford to invest massively in different plans, use cases and form factors. Fitness, work, social, metaverse, gaming, hardware... It's a collection of bets and in this day an age, one great product-market fit can scale into billions of revenue.

In a recent interview Zuck stated that he was surprised by how well the RayBans worked. It's also a great synergy with this new wave of AIs. Because they invested so much they might well get ahead in this race

latentsea2• 2 years ago

Once Meta finds the killer app for VR

The killer app for VR is porn.

indigo0086• 2 years ago

Google is notorious for in creating and sunsetting projects who's only purpose is yo showcase talent without regard for any other business model outside of their primary money makers. No support, no open sourcing, just something to show off and show that they are in the innovative app space.

SXX9• 2 years ago

If someone would tell me just a decade ago that Facebook will become one of the most open innovating companies and Mark Zuckerberg one of sane billionares I would really laugh in their face. But now...

Regardless of how actually successful their attempts at VR and AI end up they already going to have some place in history for that.

sverhagen2• 2 years ago

Grading on a curve...

cube2222• 2 years ago

To be fair, they have a very long history of open-sourcing internal software, and it becoming industry standards, this is not at all new, as some have been saying.

Especially so in database technology: rocksdb, zstd compression, presto, Cassandra, Hive, Velox (this one’s new) are all Meta’s doing.

And those are just the popular ones, there’s way more (database-related) projects they open sourced that didn’t get too popular!

So all in all, as much as I’m happy to complain about them as a company, they’ve always been huge contributors to the open-source ecosystem.

SXX9• 2 years ago

While this is true they wasn't so different before compared to Google and even Microsoft in later years. Other companies also released a lot of open source software.

AI research is different though. Everyone else decided against open source and started to dig their own moats justifying it with different kind of BS.

Yet Facebook / Meta has decided to be on side of open innovations.

iddan• 2 years ago

Also in web development technologies: React, React Native, Jest, GraphQL, & Docusaurus

space_oddity2• 2 years ago

I love my Oculus but VR has yet to achieve the cultural ubiquity (mb off topic).

latentsea2• 2 years ago

I first had a headset back in 2017 for the fun factor of playing around with the hot new thing. It has to be tethered to a powerful gaming PC, was heavy, and the resolution / FOV was relatively poor, and of course no passthrough.

I got a quest 3 about months ago and it's a world of difference from where the form factor was at in the past. It's to the point now where it's actually better for certain use cases than other alternatives and hence sees some daily use from me.

If the trend continues to get better and hardware that's smaller, lighter, more comfortable, and generally more useful, then I do see it hitting mass adoption at some point.

I think the demo Zuck did with Lex on his podcast where they did the podcast in VR using their avatars was pretty neat and the point where that is commercially available and as easy to use as a smartphone, it's a no brainer that it'll sell well.

SXX9• 2 years ago

I totally understand that VR is loss leader for Meta and a lot of Metaverse stuff they do is cringey, but at least they do have some vision. They actually build some new tech and trying to make it work for mass market.

What Google trying to do other than maintain their ad revenue? IDK.

01945503192• 2 years ago

https://www.facebook.com/profile.php?id=100083868796961&mibe...

jonatron• 2 years ago

Thanks the thousands of African workers doing the boring repetitive dataset work.

anentropic• 2 years ago

the real heroes

lakfha• 2 years ago

So you can track "objects" like people in a shopping mall and they are releasing it as a commoditized complement.

What are possible scenarios?

1) Advertisers build tracking systems in supermarkets and malls and Facebook sells them the profiles of the victims.

2) Facebook is getting people hooked on the "free" models but will later release a better and fully closed model with tracking software so advertisers have to purchase both profiles and the software.

3) Facebook releases the model in order to create goodwill among (a subset of) the developer population and normalize tracking. The subset of developers will also act as useful idiots in the EU and promote tracking and corporate power.

artificialprint• 2 years ago

Tracking objects is a different task from segmentation, whatever you are describing doesn't need what Facebook released today. In AI a lot of things are overlapping and some stuff that they have been releasing has no direct use for them.

Segmentation is helpful in creative applications, think of video editors that can segment the objects in the video and replace the background etc. Green screen type of thing.

IshKebab• 2 years ago

It's a closely related task. This seems capable of at least some tracking.

SXX9• 2 years ago

Wouldn't it be easier to achieve these nefarious goals if it's was a closed model that only few actors know about and have access to? Do you think such models not yet exist or deployed?

If you want to look for something behind what Facebook doing now then it's pretty clear they damaging a moat of their competitors like Google and Microsoft (AI.com).

Meta competitors dump shitloads of money on GPUs and running their cloud AIs. When Meta releasing open models in wild where everyone and their cat can run them it's only make this arm race more expensive for proprietary AI offerings.

thinkingemote• 2 years ago

It's (still) for the metaverse even when people say it has flopped, this is directly for that aim.

Tracking of everything via live camera vision

npiano• 2 years ago

I always think of the holographic orbit mapping UI from The Expanse when I see something like this. It's the paper of the future that will be hooked into everything we think about. Such a powerful tool for exploring the world.

asah• 2 years ago

Impressive results! Here's a test video from inside Mercer Labs: https://youtu.be/W7kM0ISXkpQ?feature=shared

afro88• 2 years ago

What am I looking at, and how is it related to SAM2?

switchstance• 2 years ago

I would have killed for this back in my editing and motion graphics days.

The Roto Brush in After Effects is similar, but its quality has always been lacking and it takes forever to process.

space_oddity2• 2 years ago

Roto Brush in After Effects is a lifesavers but with limitations. SAM is definitely a game-changer!

jdhzzz• 2 years ago

No love for Firefox.

Meta introduces Segment Anything Model 2

Related Stories