This AI Company Is Making The Most Underrated Models (Desert Ant Labs)
BBetter Stack
컴퓨터/소프트웨어창업/스타트업가전제품/카메라
스크립트
00:00:00This might genuinely be the coolest AI company out there.
00:00:03It's called Desert Ant Labs and they are a European AI lab
00:00:07and they've set out on a very specific mission.
00:00:10Instead of chasing building big models,
00:00:13they're building tiny models that run entirely on your phone or your laptop
00:00:17offline with no server and no API keys required and with no token cost.
00:00:23So in this video, we'll take a look at 12 of their little AI models,
00:00:27see how they work and test them out one by one.
00:00:30It's going to be a lot of fun, so let's dive into it.
00:00:37So Desert Ant Labs whole philosophy is built around a certain body part, the cerebellum.
00:00:43It's about 10% of your brain by size, but it holds 80% of your brain's neurons.
00:00:49And that's the part of your brain that handles the fast thinking and reflexes
00:00:54like balance, timing, coordination.
00:00:56So the rest of your brain is free to think.
00:00:59And Desert Ant Labs wants to apply the same layer for software.
00:01:03So what they're doing is building small specialized models that handle the constant background work
00:01:09on device while the big cloud models handle the actual reasoning.
00:01:13And so far, they've shipped 12 of these for various tasks like converting audio, text, and vision.
00:01:19And they have a handful of beta models coming out soon as well.
00:01:23And look at their marketing pages.
00:01:25Look how gorgeous each of them are.
00:01:27I'm just so blown away by how stylish and tasteful their model presentations are.
00:01:32And everything they've shipped is open on Hugging Face.
00:01:35Their SDK is open on GitHub.
00:01:37And it's free to use up to 100,000 monthly active devices per platform.
00:01:42So I'm going to try out every single one of these 12 models and see how they perform.
00:01:47So let's go through the lab.
00:01:52First up, we have Voz.
00:01:54That's their special speech recognition model.
00:01:57The claim on their site is that they can transcribe 10 minutes of audio in about two seconds running entirely on an iPhone.
00:02:04So let's see if that claim actually holds up.
00:02:07So for this purpose, I vibe coded a little app using Opus 5 that uses this Voz model under the hood.
00:02:13And here, I'm going to select the file.
00:02:16And I recently visited Fan Expo where I saw Mr. T in person.
00:02:20And I recorded some clips of him talking.
00:02:23So I'm going to use one of these clips, which is a seven minute long video, and see how well it transcribes it.
00:02:28So, and as you can see by the numbers here, it did actually transcribe it in two seconds.
00:02:34And here's a little playback of what that transcription actually looks like.
00:02:38A black kid in the White House.
00:02:42You know, it wasn't a joke, but the main thing, you know, I tell people, I wasn't an uninvited guest.
00:02:49I didn't jump over the wall.
00:02:51I wasn't out there protesting.
00:02:56Next, we have a model called Clear.
00:02:58And this model promises speech enhancement that gives you warm, closed mic sound of a podcast studio.
00:03:06So let's try to record some awful noisy audio and see how well it processes it.
00:03:11I vibe coded a little iPhone app so I can test out the model directly on my iPhone device.
00:03:18And I recorded a little soundbite in my kitchen, and I turned the fan up so you can hear how that sounds in the original recording right over here.
00:03:28Hello, I'm coming to you live from a very, very loud kitchen in my apartment.
00:03:35Let's see how this all sounds after post-processing.
00:03:40Indeed, let's see how that sounds.
00:03:42So, with Clear applied, it sounds like this.
00:03:47Hello, I'm coming to you live from a very, very loud kitchen in my apartment.
00:03:53Let's see how this all sounds after post-processing.
00:03:58So, I mean, it's okay.
00:04:01I feel like the noise removal was 70% there.
00:04:05But to be honest, I gave this model a really hard task.
00:04:09So I would assume for a slightly less noisy audio, you would have a better output.
00:04:13But I just wanted to test it out on the most extreme example.
00:04:21Okay, next up, we have Clips.
00:04:22And this is a model that turns long recordings into cherry-picked shorts with highlights from the video's transcript.
00:04:30And this one has an insane benchmark because they claim Clips was able to clip 100,000 30-minute videos using 470 times less energy for the same job as Claude Sonnet.
00:04:42And they also shipped this model with a desktop app called Clipper.
00:04:46So for this demo, I'm going to use Clipper and see how it works.
00:04:50So the app is quite intuitive.
00:04:52You just drag in a video and then it processes it, transcribes it, and cuts it into short clips.
00:04:59And these clips are ranked based on what the model decides are the best moments of the whole video.
00:05:04And since I know the content of this video, I have to say the ranking is not the best because in terms of virality, there were better moments for sure that should have been ranked higher.
00:05:16And it's a shame that there's no way to tweak the way the model ranks the clips.
00:05:21But anyway, you also get a little title for each clip and a summary paragraph.
00:05:27And this comes from their little title model, which we will check in a moment.
00:05:35Next up, we have Redact, which is a model that strips personal data as you type.
00:05:40So sensitive information never reaches servers or logs.
00:05:43And it supports 27 languages.
00:05:45And this one has a ready-to-use demo on their marketing page.
00:05:49So I'm going to use that for this test.
00:05:51All right, as you can see, we already have some kind of an example here.
00:05:55It's highlighted and we can see how it looks redacted.
00:05:59So let's start typing something in.
00:06:00My name is Andrus.
00:06:03Let's try another one.
00:06:04My password is hello something something.
00:06:09And look at that.
00:06:10It doesn't recognize passwords.
00:06:12So that's a bummer.
00:06:13I think that should be included in the model.
00:06:15But nonetheless, what if I say my SIN number is...
00:06:21And then let's write something in.
00:06:23So as you can see, it doesn't even recognize a Canadian SIN number.
00:06:27So that's a bit disappointing.
00:06:29Let's try my native Latvian language.
00:06:32And as you can see, it also understands my native language, which is pretty cool.
00:06:38All right, let's move on.
00:06:42Moving on to the next one, which is UM.
00:06:45And this model claims to remove all filler words within 20 milliseconds.
00:06:49And it can process an hour of audio in just 12 seconds.
00:06:53So once again, I have installed this in a vibe-coded app, just so I can test it out on my iPhone.
00:07:00So I am working on a video today.
00:07:06And it's...
00:07:08I guess it's going pretty well.
00:07:11I am now recording an audio for a certain portion of the video.
00:07:18So it found five fillers.
00:07:20And it did this in under a second.
00:07:23So I am...
00:07:25I am working on a video today.
00:07:28And it's...
00:07:29I guess it's going pretty well.
00:07:31I am now recording an audio for a...
00:07:36Hmm...
00:07:36Certain portion of the video.
00:07:38Let's see how that looks.
00:07:40So it did not clip my hmm portion.
00:07:42But it did remove most of the ums.
00:07:45So I would say it's a nice model.
00:07:47I wouldn't say it's perfect.
00:07:49But it does its job.
00:07:51And I'm going to be honest.
00:07:52I wasn't a fan of the cut as well.
00:07:54Because it didn't slice the audio neatly.
00:07:58Nonetheless, it's a cool little model as well.
00:08:04Next one is called Emo.
00:08:06And this is probably one of my favorite ones.
00:08:08It basically suggests an emoji that fits the text you've just written.
00:08:12And they claim it can do it faster than you type.
00:08:15So let's check how this works.
00:08:17So we have a bunch of examples here.
00:08:19Book a flight to Tokyo.
00:08:20And it shows a plane.
00:08:22Let's do a country.
00:08:24Canada.
00:08:24It gives me a maple leaf.
00:08:26So pretty close.
00:08:27I'm surprised it doesn't give you a flag.
00:08:30But okay.
00:08:31I'm running late.
00:08:33Running late.
00:08:34So we have a running man.
00:08:35Clock.
00:08:36Another clock.
00:08:37I guess for late it suggests a moon emoji.
00:08:40So I think this is a pretty fun little model.
00:08:42And for most of the part it's quite accurate.
00:08:45And it's only 5 megabytes on device.
00:08:47And it gives you a suggestion in under 2 milliseconds.
00:08:51So that is a pretty low memory footprint to have on your apps.
00:08:59Next up we have Gist.
00:09:01And this is a model that basically predicts topics and tags based on the titles of your post.
00:09:07And it claims to be able to do this in 101 languages.
00:09:11And they also have an on-page demo.
00:09:13So as you can see we have an example here of an English title.
00:09:17And it recognizes that full body workout means health and fitness with 99% accuracy.
00:09:25So here's a news article I just pasted from a news site.
00:09:30And it talks about dogs.
00:09:32Whether they're sad or scared.
00:09:34Study suggests.
00:09:35And it suggests arts and culture.
00:09:38Which is not actually accurate.
00:09:40Because this is a scientific study.
00:09:42But okay.
00:09:43Here's another one that talks about space travel.
00:09:46And this gives me careers and productivity.
00:09:49Not the most precise result I was hoping for.
00:09:52I think this model needs some improvement.
00:09:54Because I think this model mainly focuses on keywords and not on context.
00:09:59So I wouldn't say this is production ready to be honest.
00:10:06Moving on.
00:10:07The next one on our list is Align.
00:10:10And this is a small model that generates word timestamps on device from any transcriber.
00:10:15And they claim that it is more accurate than Apple's own transcriber.
00:10:19Because it is able to detect the actual beginning and the end of a word more precisely.
00:10:25And this one is a bit tricky to test.
00:10:27Because in order for me to run a nice visual demo.
00:10:30I would need to generate two separate transcripts of an audio and video file.
00:10:34And then try to see if it matches precisely with the millisecond precision.
00:10:38So for this one I'll just trust their claim here that it's pretty performant.
00:10:43And I'll let you folks try it out for ourselves.
00:10:45But let's move on to the next model.
00:10:50So next up we have Shapes.
00:10:53And this is an on-device shape recognition model that turns your hand-drawn strokes into precise vector graphics.
00:11:00And I gotta say it's nice.
00:11:02I mean it does recognize a line, a rectangle, a triangle, and a star.
00:11:08But I couldn't get it to recognize an ellipse properly if I draw it with two separate lines.
00:11:14Only if it's like a circle line.
00:11:16So this model is cute and all.
00:11:18But I think it still needs some improvement.
00:11:24Let's move on to the next one.
00:11:26Which is called Title.
00:11:27And this one basically generates a short factual title and a short description based on a text block.
00:11:34Or in this case with Clipper as we saw before.
00:11:37It can suggest titles for your short clips based on the text content.
00:11:41We already saw it how it works in Clipper and it's pretty cool.
00:11:45So I'll just leave it at that and let's move on to the next model.
00:11:52Next one is called Tongue.
00:11:54And this model is able to identify the written language from just a few words.
00:11:58And it does it in tens of microseconds.
00:12:01And it supports up to 84 languages.
00:12:04And they also have an on-page demo here.
00:12:06So let's test it out.
00:12:07Okay, we have some examples here.
00:12:10It's French.
00:12:13Mitsubishi, it seems to think this is Polish.
00:12:17Not quite, but solid effort.
00:12:20Let's see if it recognizes Latvian.
00:12:22It does recognize Latvian in this case, which I'm super happy about.
00:12:26I have no complaints.
00:12:27Tongue works pretty good.
00:12:29And this model is only two megabytes in size.
00:12:32So I think the memory footprint is so small that this is probably worth putting in your application if you have a production need for it.
00:12:45And last but not least, we have ear, which is a similar concept as the Tongue model.
00:12:50But this one can identify 99 languages from a short stretch of audio.
00:12:55So for this one, I'm going to be using my Vibe Coded app as well.
00:12:58And let's start talking and see if it can recognize this is indeed English.
00:13:04And yes, indeed, it is English with 100% accuracy.
00:13:09And it identified it in 10 seconds.
00:13:12So I would say it is a bit slow, but it does the job.
00:13:16Let's try different languages.
00:13:21And let's see how well it does with my native Latvian language.
00:13:26Okay, this is not good.
00:13:27The model took five seconds to process the audio.
00:13:31And with 29% confidence, it thinks it's Turkish.
00:13:36When in reality, it is not Turkish.
00:13:40I don't think Latvian is even similar to Turkish.
00:13:43So I'm quite disappointed in that, to be honest.
00:13:46But okay, let's move on.
00:13:47Let's try French.
00:13:52And again, five seconds for processing.
00:13:54And in this case, it did identify that it's French with 96% accuracy.
00:13:59So at least it works for some languages.
00:14:03So there you have it.
00:14:04We've gone through 12 of their recently published models.
00:14:07And Desert Ant Labs also have five models still in beta.
00:14:11Who is for speaker identification.
00:14:14Shemmer is for pooling structured JSON out of raw text.
00:14:18Toxic is for catching hate speech before posting.
00:14:21Moderator for flagging nudity before upload.
00:14:24And eye and face for different vision tasks.
00:14:27These aren't fully released yet.
00:14:29But seeing how fast this little AI lab is pumping out models.
00:14:33I'm not surprised if they're going to be available very, very soon.
00:14:37So there you have it, folks.
00:14:38That is a little introduction into Desert Ant Labs.
00:14:42I really think they're going for the right approach here.
00:14:44Because in the world with so many frontier AI models that require a ton of disk space.
00:14:50And a ton of RAM to run.
00:14:52I think it's so cool that there are companies like this one.
00:14:55That are focusing on models that are super good at one specific task.
00:15:00And I think that is the sensible future where on-device AI is heading.
00:15:04I'm sure we will see a lot more examples of these micro AI models.
00:15:08That are super efficient at one single task.
00:15:11So we can build interesting apps around them.
00:15:14But what do you think about Desert Ant Labs and their models?
00:15:17Have you tried any of these?
00:15:19What are your first impressions?
00:15:21Let us know in the comment section down below.
00:15:23And folks, if you like these types of technical breakdowns.
00:15:26Please let me know by smashing that like button underneath the video.
00:15:30And also don't forget to subscribe to our channel.
00:15:33This has been Andrus from BetterStack and I will see you in the next videos.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기