Gemini 3.7 Flash Is HERE – Testing Google’s BEST Model Yet!

0:00 Rowdy Rod, Johnny Piper. Okay. Four days ago, Gemini 3.75. What the heck did I just say? Four days ago, Gemini 3.7 Flash was released. Maybe I had a Freudian slip thinking about 3.5

0:15 Pro. I don't know if that's coming out or not. I genuinely have no pertinent information. But this model, from what I've seen so far, being that it's been a few days since this came out, seems to

0:24 be getting some pretty rave reviews from folks who I trust in terms of what they say about some models and their performance, which is pretty exciting. Now, additionally to this, beyond just

0:34 being the new flash model that supposedly outperforms the previous ones, this also comes with a pretty significant price cut in terms of using it. So, Gemini 3.7 Flash seems to be a

0:45 pretty big increase in capability over 3.6 Flash. not only just reflected in the deep s benchmark here, but basically all the other ones that are included as well. Additionally to this, I have also

0:56 seen some folks speak pretty highly of this model's performance, which is not something that I saw with 3.6 Flash or I think 3.5 Flash. So, that is good as well to see community sentiment seeming

1:07 to be fairly positive in its capability. I do have to say this is a 3-week gap and I feel that more recently we've been seeing more examples of this where in a smaller time scale we're seeing

1:18 iterative improvements come from a bunch of different labs that are then like smoking the model's predecessors just like this a year ago it was not the case where like every few weeks a new model

1:28 was dropping to like beat the old one from the same lab. So it does seem like there is a speed up at least in terms of some releases or maybe I'm wrong. I don't know. there's been so many this

1:38 week. Now, additionally to that, they don't really say too much about it right here. It has the traditional like examples, a fully playable 3D game, a fantastic benchmark idea in my opinion.

1:48 However, they do also show that this has decreased in price massively. So, I believe one of the biggest criticisms from when maybe 3.5 Flash was released was that it was too darn expensive. And

1:59 now the caveat here is that there is some introductory pricing for this that is good for the now till the end of the year which is like four months. With that though like there probably be like

2:11 Gemini 3.9 flash by the time that pricing expires. But still it is significantly cheaper now than 3.5 flash was when it came out or 3.6 flash at 75 per million in and $3.75 per million

2:24 out. Now, the little star here is basically because that is introductory pricing expiring on December 31st. Starting January 1st, 2027 with whatever flash model exists. Then, it'll be $1.50

2:37 per million in and 750 per million out. So, it will double in price then. Keep that in mind. And I think this benchmark chart right here pretty accurately depicts why 35 flash may have not had

2:47 the best community sentiment around it because the performance and then the pricing these are both just like in the wrong direction of where you want to be on this chart. But we see 37 Flash is in

2:57 a very happy space of the top right corner of this chart. So that's also good to see as well. And it does reflect that deepsw SWE benchmark increase over its predecessor that we see up here

3:07 which is pretty exciting. Now, I have to say that just based off of like community feedback, most folks I don't think really look at the existing Gemini models as coding powerhouses. I will say

3:18 I have probably sent more tokens through Gemini 31 Pro than any model that currently exists and none of it's been coding, but they do have a lot of strength in both image and like general

3:31 knowledge capability. They have good long context performance. And I find that even if they're not specifically great at some of the tasks we like to run in these benchmarks, I do find that

3:42 the vision capabilities and just some of the general knowledge capabilities of the Gemini models have always really impressed me. And truthfully, I'm not like BSing right now. I have 31 Pro open

3:52 right there on my main computer that I use for a lot of my work. I use that thing a lot. So for a bit of technical information about this model, we can see right here that it is a multimodal

4:02 model. So it takes text, image, video, audio, and PDF input. And it has text output. The token context length is a little over a million. And the output token limit is 65536.

4:13 So not super long there, but you know, the context length is pretty decent at least in terms of that. All right. So, here's the browser OS v2.7, which is basically the same traditional browser

4:23 OS, but it needs to have procedurally generated wallpapers with movement and an email client. All right, we have our implementation plan. And for some of these simpler results, I'm just going to

4:33 allow it to just continue. We don't really need to look through it too much. The speed here is quite something as we can see just by how rapidly it's putting out paragraphs of text. So, that is

4:42 something I've also been seeing just in terms of community feedback in this model is that the speed is quite good. It is a flash model, so it does live up to its name. But also in that point, it

4:52 is also like the high speed is just available in this fast mode for a limited time. So keep that in mind as well. There we go. All right, let's take a peek at this browser OS

5:03 Nexus Pilot. I Okay, we have to enter. I'm going to just make sure the speaker's on. Okay, I'm assuming the pin is probably 1 2 3 4. Excellent. Okay, this is quite the synth

5:19 wave aesthetic. And these models I've like in previous flash. Okay, there is a right click. In previous Flash model results, they've seemingly very very much like the neon cyber aesthetic.

5:31 Okay, we do have the date in the bottom right corner and we can actually click on that and it opens a notification thing. Interesting. We have sound. Master volume. Wallpaper speed is

5:42 adjustable from there. Let's slow that down. I like that better. Wi-Fi 6E Bluetooth theme changed. Oh, it's got a very very feature-p packed

5:54 just like bottom right section. Make of that what you will. Here's our start menu. Okay, we have search. No. Okay, let's just start out with Grand City 3D. Good.

6:10 I I literally clicked in the viewport and we got one star. Oh my goodness. Well, it seems like we'll be able to outrun the police on foot. Oh, you can actually like All right,

6:22 I do have the sound on. Good. We now have like five stars, but there don't seem to be any police in this game. Oh, cool. What's uh E to enter exit? Oh, okay.

6:51 All right. So, troubled, but um yeah. All right. What's next? Aeros Strike 3D. Is this an asteroid? Yep, of

7:03 course it is. They always are. Am I supposed to like blow this planet up? Okay. I don't think these games are like 100% there. They're like 80% there. Omni

7:25 mail. Okay. New sports card drop in District 4. Oh, we're getting like spam mail from our games, which is a real thing. Tactical alert.

7:35 Greetings. Sound designer starred. Send trash. Oh, okay. So, we only have our inbox compose. Send message. Good. And it tells us, hey, you didn't

7:47 enter a message, right? Quantum sent. I've seen one. It looks so simil This is not There's no prior art on the system for this to go off, but I This looks pretty similar to something I've seen

8:00 before. All right. and this thing actually like gets bigger or smaller. Kronos City Matrix Time Machine. Okay. I don't know why that would be shoved in the quantum

8:27 synth. Well, maybe because it's like quant I don't know. square snapshot OS. Okay, let's close that. Rewind OS. Okay, it's it looks cool. Uh settings,

8:52 procedural wallpaper, hyperspace, neural web. Okay, Matrix Rain theme and color schemes as well. And they I kind of like the of course I like the synth wave flow. You know, if you watch the

9:06 channel, thank you. And you know, I'm a kind of a sucker for whenever they put in synth wave stuff. Okay, nothing else here is working. Then finally, we have our Cyber Deck, which is just the

9:15 terminal application. I don't know why it named it that. Okay, that's a pretty that's a light list of potential commands. There's no Neoetch or anything. So, this was competent. It

9:25 wasn't like the best I've seen, but I will say the amount of like interactive H. Look at all these notifications. Can we actually clear them? Yeah, we can. Okay. The amount of interactive things

9:35 down here was nice to see. It's not often we see those then the fact that it had a pin to log in, which it told us. So, not very good OPSAC, but kind of interesting. Some interesting features

9:51 that are not as often seen. And the synth thing was cool with the UI. All right, we are going to do the skate game, but to be honest with you, I really want to try the new 3D wrestling

9:59 game prompt with us first. This was one that was introduced in the GLM 5.3 video. So, yesterday's video and it seems fun. I noticed some folks seemed to also share the same level of

10:09 enthusiasm for this as I do. So, I was happy about that. And it can definitely be funny, if not good. I mean, because it's an 80s theme, this model seems to love that stuff. So, I am expecting

10:19 something pretty good for this. All right. I actually want to see Slam Mania 85. Interesting because GLM53 named it like Mega Slam 86. So 3D high-end indie wrestling game. Oh

10:35 wow, look at this rowdy rod. Johnny Piper. Okay. Irish whip and rope rebound launcher. I think I know who's going to use that

10:44 one. All right. This is a pretty decent implementation plan just after a quick scroll through. I'll be excited to see what we get. Quite excited. I honestly think this

10:55 took like less than five minutes. So, I mean, we still haven't seen it yet. So, that may be like, oh, yeah. Okay. I just going to go out on a limb here and say I do see some issues maybe with the

11:07 movement of the ring. Quick three-way match. Uh, enter the ring. It's just >> three WAW championship match begin. Why is

11:21 >> one 2 P I N F A L attempt. One 2 P I N F A L attempt. >> One 2 P I N F A L attempt. One 2 P I N F A L attempt. One two P I N F A L attempt.

11:40 >> One two P I N F A L attempt. >> One two P I N A L attempt. >> CRT on or off. >> Two P I N A L attempt. one up. Okay. So,

11:54 uh oh, we're we noticed this with GLM where the sidewaysness. Okay. This this is just not I need All right. I think we found a new prompt that just like totally totally

12:10 Oh, there's the referee right there in some Okay. Uh All right. So, I've given it some feedback and we'll see what we get. All right. Apparently, the updated game is

12:22 live. So, let's go back there real quick and refresh it before having the speaker on infuriates me again. Okay, good. The mat isn't moving. Let's just Hey, not bad. It seems to have All

12:33 right. Oh, super slam. I'm just So, I think where this one Okay. Yeah, we are. All right. It's Oh, that's a lawsuit against the crowd.

12:54 Can I pick this chair up? Oh, look. Did you see the rebound in the in the um I know what the term is for these. The ropes. This is

13:16 Okay. I genuinely don't know what to say. I think it was okay. It's just tough because this is a test that's this is the second time we've done it now ever.

13:30 I will say GLM looking back retrospectively. The way it actually drew like some weird wrestling masks on some of the folks was pretty unique. And I think overall that one actually may

13:40 have had a more immersive or impressive like scene to it. This was definitely more playable in terms of just spam clicking and running around and you could actually see some

13:51 tricks being performed or moves I should say. It was interesting and I don't actually really have a great deal to say about it. So make of that what you will. So, next up, we're going to be doing the

14:02 self-contained C++ skateboard game. This is the one that is the early 2000's New York City map, and it is also told just not to use Rayb. All right, that was very quick to make this skate game. And

14:14 this is, of course, one. It's just completely unnecessary. All right. It's It's It's troubled. Um, okay. Here's what we're going to do.

14:36 How does this look? All right. That was a very, very quick fix. Now, I didn't get to explain to it that also like the Ollie logic didn't work. Okay, it seems to have fixed that.

14:51 Just Oh, nope. Okay, it's still messed up. It did put like a skyline in it now and it added some like heads up display things but it's not quite for some reason they always work intermittently.

15:04 Okay, cool. We do have All right, so this is going to be almost somewhat of like a multimodal improvement test where I've given it screenshots from the greatest of all

15:14 time result for this specific prompt which was Quen 3.8 Max, a 2.4 4 trillion parameter open weight model. And I'm saying basically the OIE works like half the time. The rider model is bad. It's

15:26 bad. Make it better and then make it look like these. So we'll see how it does. All right. Again, extremely extremely fast. And this is the last time that I will take a peek at this

15:36 just because normally we don't give them multiple times. But all right, it definitely made it look more like the Quen result. It's still just like the ollie is fixed, but now I notice that

15:48 none of the none of the tricks are working or it either like remapped the keys to properly reflect that of the Quen game. Yeah. Eh, overall, I'm not blown away, but it did interestingly

16:04 very much change the visuals. And these are pretty low res images. So it focused more on like the overall scene and maybe not like the niche detail like the sidewalk textures, the trash bags there,

16:14 the rider. So just interesting. All right. Next up, I'm giving it the 3D CAD model creation test. It needs to create the 3D printable model of the inline 6 RB26 motor, which inside must fit the

16:26 tiny N20 RC motor and then give me a 3D printable model. What did that just say? I'm thrilled with I'm thrilled with the user's approval. Okay, good. All right, so this is by far the longest we've had

16:37 to wait for a result. I think first and foremost, I want to just take a look at the preview. Okay, on first glance, this actually looks pretty good. I did like Is that like a half cutaway? It is very

16:49 interesting. Okay, and we have our separated pieces right here. I do see alignment pegs and things like that just to make sure they fit together. Okay, I don't know what that was. Maybe

17:00 like a plate. It probably tells us print bed layout. Okay, I've seen enough here that it is worth actually looking at some of these models individually. Let's just start by looking at the complete

17:11 assembled one. That's fairly big, we'll say. Okay, there's definitely some weird geometry here with the turbos, but this is one of the first times they've

17:25 been oriented in the correct way and not like sideways. So, that's good to see as well. And keep in mind that this is the entire thing put together. I don't interestingly see any actual text on

17:37 here. A lot of times the models will say like RB26D or something. This didn't. Okay, we have the block cradle. And unfortunately, this is not printable without supports

17:51 because of this oddness across here. This is troubled. Oh, yep. Okay. That right there is like the the killer. These pegs just basically killed us from having any hope of being printable

18:03 properly. So overall, it looked pretty, but it unfortunately did not satisfy the printable with little to no supports portion of the test. But overall, as a model, when judged, it was actually

18:15 competent, although not super detailed. All right, this thing is so fast. Let's just give it something that's like very difficult. Very, very difficult. All right, here's

18:24 an Android TV box. It's called a T95 Super. There is a variant of this that has some ability to actually have a flavor of Linux installed on it. Now, that is not what I'm going to have this

18:36 do. All right. So, I'm just going to be connecting it via the USB cable. Now, these do also have SD card inputs. I'm not going to opt to use one. I'm going to try to just give it like

18:48 a little help. That is the wrong thing. All right. So, the computer might not be seeing it. I was just going to give it the specific device ID cuz it's connected from USB, but let's see if it

18:59 can figure out. I'm changing this a bit. I'm saying I want you to just wipe everything on this and then get it to boot some old open- source textbased operating system,

19:10 something like DOSs. Now, I was not able to see the Okay, so it's going to try to use ADB to this to do this, which is fine as it does run Android. I don't think this will work. I just want to

19:23 clearly state that prior to running this, but it's sometimes still interesting to see to what degree of completion these models actually get. Tasks like this that are significantly

19:33 more difficult than like even a C++ game. All right, so it's searching for the specific specs for this boxes like hardware. All right, so I don't think it can find this just based

19:46 on USB because it wouldn't be looking for it in this regard. I would imagine if it was able to find it. So maybe I'll plug it in via Ethernet. All right. I'm just telling it if you can't find it on

19:57 the USB hub, which I don't think it's going to be able to. I just told it I plugged it in to this host through Ethernet. All right. Good. It does seem like it found it because now it's just

20:07 asking if it can back up the bootloadader and boot partitions and basically all the software stuff on this box, which does bode well for like it then wiping it completely. I will say

20:16 the speed is is cool. It's just nuts to have it. Like I like doing more ridiculous tests like this that probably won't even work when the model's this fast cuz you're going to know real fast

20:28 like whether or not it's going to work or not. All right. It's just trying to figure out more about like how this device actually works from like the lower level. The goal is now to wipe the

20:37 Android partitions to create a clean DOS/ext environment. Oh wow. So we're just now onto our implementation plan. I actually would like to take a second and just look through this a bit because it

20:47 will show us some of what it's done so far. This plan details the procedure to completely wipe the existing Android TV OS on this T95 Super with the AllWinner H3 and these other like bits of hardware

20:58 and install an open-source retro textbased OS. The installed system boots instantly on bare metal hardware into an authentic free DOS style text operating environment displayed over HDMI with

21:09 full USB keyboard support. Classic DOSs utilities, full screen text editor, basic interpreter, disc utilities, retro aski games, and seamless Unix shell access. Important user review. Yeah,

21:20 it's going to wipe it. That's fine. I have like too many of these. Okay. Host connectivity system on chip architecture. All winter H3 storage proposed operating system features

21:31 direct frame buffer text rendering engine. It's going to build the retro text operating system engine. So, it's going to write it itself and see of the free DOSs text operating system or it'll

21:41 use something that exists. I'm not very in the know about like some of these things. So, keep that in mind. Build the minimal minimal bootable file system. Package the new boot image. Wipe the

21:54 Android partitions and flash the new boot image. Reboot and verification. All right, let's let it rip. Now, again, I don't think this is easy. I may be mistaken, but essentially I have a bunch

22:04 of these because I used these to build my social robots in 2022 when I was designing them. It was very difficult to get Raspberry Pies and things like that. There were still a lot of supply chain

22:15 backups. So getting even like a bunch of pies or any other like traditional single board computer like an orange pie or whatever was cost prohibitive one and like difficult to find too. I basically

22:26 would have had to pay multiple prices to people who were just trying to flip them for money, which I was not cool with. So, I found these Android TV boxes instead and then made the robot software

22:36 in Unity as an Android app. And therefore, these were basically like getting a cheap $20 single board computer, which is exactly what I needed to build the robots. So, that's why I

22:45 have a bunch of these and I don't sell those anymore. They're open source and they're on GitHub now. And they're not designed to work with these. They're designed to just work with like a Jetson

22:54 now or some other single board computer. But at the time, these were like a pretty cool workaround. And some folks were also installing Linux on some versions of these because they were a

23:04 decent alternative to a Raspberry Pi when you couldn't even get a Raspberry Pi. So, for some backstory on why like I have these. So, it's made the image now, and it's

23:18 going to now try to get it running and stuff on this TV box, which is the difficult part. It says it did it. All right, I have a monitor, so I'm going to plug this into

23:32 HDMI. I hope I didn't like massively overestimate the difficulty of this task, but Oh, monitors generally need power cables plugged into them, too.

23:45 Okay. Okay. It did something. I can tell you that it did not 100% extrapolate the proper way to display things here. However, the fact remains we have something to give it. I need

24:00 another camera. So, this is what it's showing on the monitor and I'm going to just give it that and show it to it. Now, in this specific scenario, and something I've ended up doing before

24:11 with other things is you just give the model a webcam attached to the computer and you tell it like, "Hey, go wild." Like, point it at whatever is showing the actual like real output. And then

24:24 you give it the webcam and say like, "Now you can check your own work without me needing to just send you images each time and things like that." So, that's something else that is pretty useful

24:34 depending on like if you're troubleshooting more of like a hardware software thing. The fact remains though, it it did do something with this TV box. Now, I may be kind of out of date in

24:44 terms of the level of difficulty here. And for all I know, maybe it leaned more on some open- source stuff than I expected, but I don't think there's a lot of info on this all winter H3 still.

24:54 So, okay. So, you want me to plug a keyboard into this and then type the entirety of this without being able to see the text? Yeah, I think not. How about you do it for me? That was

25:04 actually kind of cool, though. It basically made us like a software update command to run on this little DOSs box by itself, which I think is cool as well. I just I'm not about like trying

25:14 to type that. If it needs me to, I will, but I think there are some ways that it could probably do this itself. Okay, it just wants me to do the keyboard thing, so I'll do that.

25:27 Which key is the pipe? since I can't see anything I'm doing. Okay, so let's shift. Yeah, I just see a bunch of red text that appeared now. So,

25:46 all right. So now I'm going to It's literally cheaper to just go get another not cheaper. It's literally easier to just go get another one of these and then have it try to flash it without the

25:57 goggle text than it is to try to type this. All right, I plugged in a new one. So it's booting like the traditional Android TV now on the monitor. So it's a

26:10 clean one and it's running the automated flash script though. So, say this was for example like a say I was building for some reason like an old DOSs emulator and I had a bunch of these. It

26:21 did just make a flash script for them. So, I could basically just hot swap them and then flash the software to them. So, that's cool. Go ahead and plug it in whenever you Yeah, it is plugged in. So,

26:32 it's plugged in. Oh, starting flasher. Good. So, it it ran a listener and it automatically by itself detected that this was in fact plugged in. Good. It did just wipe it.

26:44 Okay. No, it's it's rebooting into Android TV. So, we'll see what we get. Oh. Oh my god, it did. I really don't think that I underestimated the complexity of this,

26:59 but I may in fact have because it literally Look at this. Welcome to Free DOS 1.3 T95 super retro console. All the text is working now. I mean, and then we can basically just

27:13 start using it. Let's do snake. All right. All right. This is a little more difficult than I'm used to, but

27:40 something's not 100% right with the keyboard. Like key reading speed, but the snake molted. All right.

28:00 Yeah, I don't know how to use DOSs. So, how do you go home? Like Matrix. Edit is a full screen editor. Oh, look at that. Hello. There's a Unix shell.

28:19 Oh, okay. For some reason, it's not happy there. All right. I think the point is that this actually worked and I don't know how difficult or not difficult this task

28:34 was. So, uh, I need a a second. I'm going to have it like give us the culmination of this task in an attractive single file website. I

28:49 wonder if I reboot it. Like if I just kill power and then plug it back in. Let me unplug everything except the HDMI cable and we'll see if it still boots as like a self-contained little

29:01 like old school DOS thing. Nope. It just booted back into Android TV. Okay. Oh, no it didn't. That was just a splash screen. It did. It completely booted back into this DOS little T95 super

29:15 retro console. Fascinating. Nice. And we have a good little render. DOS PC. A fully wiped high performance TV box engineered into a dedicated open- source free DOSS 1.3 computing appliance.

29:28 Instant cold boot zero bloat. Direct 60 fps HDMI VGA frame buffer rendering and USB keyboard native input. I can verify these do work. Launch live emulator. Nice. And that is kind of what it looks

29:43 like except Oh, this actually I think put this in the browser. It's It says T95 super retro console. So, okay. How the transformation was built. Storage wipe and partition repurposing.

29:56 Stock Android TV boxes use a complex super partition with DM linear mounts. Okay. All Android partitions were zeroed out in the 12 gig udis partition was formatted as clean high-speed drive C.

30:08 Stage two direct HDMI frame buffer in CP437 engine. Instead of heavy display servers, the DOSS kernel communicates directly with the hardware frame buffer memory.

30:19 Okay. An authentic 8 by6 IBM font table was mapped to render and that fixed the garbled font that we saw at first. USB event subsystem and hot plugging. The input layer actively pulls Linux device

30:31 nodes with hot plug detection. Any USB keyboard plugged into the unit is immediately bound to standard DOSs key codes. And it did work cuz we played snake and we were using just like a a

30:41 free keyboard. I got from MicroEnter. Okay. The entire OS fits into a compact 14.2 megaby Android boot image V2. It cold boots the TV box straight into free do shell in under 2.5

30:55 seconds. Built for retro lovers and hackers. Now, this is the big question because I don't believe there are open- source existing solutions for like playing with

31:06 these specific chip on this. We reverse engineered the AllWinner H3 Android boot chain and display pipeline wiped a proprietary Android layers and built a custom direct

31:15 frame buffer Fred DOS kernel tailored specifically to the box's hardware. Here's what it did use. Linux 4.9 Sun 8i kernel. Okay. Retain the vendor's kernel binary to guarantee 100% hardware driver

31:26 compatibility with the HDMI. Okay, so that made things a lot easier because that would have been probably a lot of back and forth. FredOS architecture and then the font. What required reverse

31:36 engineering? Undocumented partition table and storage architecture. White label Android TV boxes like the T95 Super don't publish board schematics or storage maps. We had to inspect and

31:45 reverse engineer the flash layout from the running device. Okay. Vendor boot image layout device tree. Stock TV bootloadaders reject standard desktop Linux or generic DOSs boot sectors. We

31:56 had to reverse engineer this specific boot chain. And that was one of the things when folks were putting Linux on these like back during the supply chain constraint days. There was a lot of like

32:06 talented people figuring out how to do that because these models like didn't exist. All right, next up we're going to do a simple subway FPS test. All right, so surprisingly our Subway FPS result

32:17 had a few issues where basically clicking to enter station never actually worked. Okay, good. Now it does. Um, okay.

32:30 We have a clean but basic station. It's for some reason snowing indoors. Those are just like particles, but still it no escape. Sector 4 quarantine. Those are security cameras there. Overall, it's

32:45 Oh, hey, there is a subway. It's just kind of like creepily lit, I guess. Now, something I'm wondering is Okay, good. Oh, wow. Good. I like seeing that. Why is the sound not working? The speaker's

32:58 on. It's just reloading. All right. This is simple. Okay. There are bullet holes in the subway car, which I like. And in the

33:14 walls, we do have a good good vending machine here. Can we change weapons? Oh, cool. Look, we can. The weapons definitely cyber suppressed

33:28 pistol. I would just imagine that's probably what the sounds of the ability to change weapons is something that oddly I don't know that I've ever seen. All right.

33:38 Overall, it's that's actually kind of cool. The effects are accurately representative of the specific selected weapon in terms of the holes that get put in the

33:49 environment. So, this is actually kind of fun. It's not visually great, but it's These do like weird spin moves. What happens if we lose? Oh, it just respawns us. Oh, all

34:04 right. Cool. I don't know why the sound is not working. But next up, I'm going to give it the beautiful watch website where it needs to create a good-looking 3D model of a watch, have it in the hero

34:14 section with a cinematic panning like shot over it, and then just a good-look front end. All right, here is our luxury watch website. Okay,

34:26 I'm noticing a few problems of UI. However, let's maybe just judge the watch as it sits. All right, it's definitely I will say this is an increase in capability over the previous

34:39 Gemini models that I've seen. We have our watch face. It even has additional dials here. 30 minutes 60 seconds. We can really zoom in on this. Okay, slept watch co.

34:52 Oh, so that is supposed to be placed flat here. Second hand is moving. There are straps. There is a crown. There's even some oddly positioned additional buttons there. Hey, corresponding with

35:02 the extra dial. So, that's some interesting attention to detail. The table has those luxurious squiggles that we're all used to. And then additionally, I see some

35:13 interesting things here like we can actually change the color. This is almost like an interactive I like that arctic mirror steel DLC stealth onyx. So I've not really seen it's almost made

35:26 this a very very interactive like viewer shopping experience. It seems like we also have some okay exploded view. Oh that I've not seen before. Interesting. That's actually kind of

35:40 cool and arguably a better presentation of the watch. I think the problem is the assembled view and the exploded view. This piece right here never quite it back. Loom mode.

35:51 Okay, that is Hold on a second. Check that out. It put like little glow-in-the-dark things and you can kind of see them when you turn that on here.

36:04 I will say this is a very the model is not the best one I've seen, but the amount of features it's included here is like even presets for like different things. It's very very feature packed

36:20 and that I've not really seen before to that degree. Submicron precision. Oh no. Okay, this though, like this just kills it because normally we like to see the

36:32 watch models also down here, but it went so hard with like the display and presentation up here that it it probably was like, "No, I don't need to put them elsewhere on the page." So interesting.

36:43 I see signs and and pieces that I've not seen before, like the exploded view and the customization in the hero section, but then at the same time, the models, it's just not all there like this. So, a

36:56 leap in capability over the predecessor though, definitely. All right, so we have our Street Yee game. This is going to be hopefully pretty funny. It again, it was just super fast. All right, so

37:07 our Street Yeek game is finished with the cut scene to >> That was the last one, >> was it? That was the last one.

37:18 >> Okay. speed as well. That was the last one. The sound is a little like violent.

38:12 Then we have to press space to it's it's a bit glitchy. Let's just see if we can eat this cab. I don't Why do models sometimes do that? They put in that like uhoh. All right. This

38:31 is a little troubled. Okay, here's the thing. So, I'm probably going to conclude this video here because I've realized that the entirety of the audio for this video from this mic right here

38:43 is not existent and I have no audio for the entirety of this video that is of any reasonable quality. So, I now need to spend some time figuring out how to make the iPhone audio to work well. Due

38:54 to that, I think the conclusion and closing thought section of this video is going to be a bit more concise than normal as I try not to spin out on camera. So, basically, overall, this is

39:06 definitely definitely an improvement over 3.6 Flash. I didn't see anything that was completely like, oh my goodness, this is mind-blowing. However, I am going to maintain that I think what

39:16 it did with this little Android TV box was pretty darn impressive. I may have massively overestimated the complexity of that task, but it seemed like something I didn't think was going to

39:26 work, and it did, and it did a really decent job at it. So, I know sometimes folks want to see more tests like that and less of the game stuff, but like what can this thing actually do with

39:36 something that is not going to be in the training data? It's like, oh, here's how to get a T95 Android box to run like a DOS operating system. It's just not in the training. So, that I wanted to

39:48 include even if it took a bit of time. The pricing for this model, at least until December 31st, is pretty cheap. So, it's comparable to like a lot of cheaper models in price, but

40:00 quality-wise, I would say it's definitely at this price point something to cause problems for maybe some other providers, especially I don't know cuz Sonnet 5, the test was a while ago, but

40:11 I think this model definitely has a place and we have to remember this is the flash model now because there hasn't been a pro model update for quite a while. I do think that folks are likely

40:21 going to be looking to these newer flash models and assuming like, okay, well, this is their absolute state-of-the-art top tier model right now because there is no updated pro. And I think that's a

40:31 fair assessment to make. So, I understand why some folks are looking a bit negatively at some of the recent coding performance because there is no frontier competitor right now from

40:41 Google, but there are some mid-tier competitors in Gemini 3.7 Flash. So, that is probably going to conclude today's video. If you have any questions that are not at all related to anything

40:52 sound related in this video, please feel free to leave them down below in the comments. And thanks for watching.