YT YouTube2Text

YouTube2Text → Transcripts → Gemini 3.6 Flash Is HERE – Testing Google’s BEST Model Yet!

Gemini 3.6 Flash Is HERE – Testing Google’s BEST Model Yet! — Transcript

by Bijan Bowen · 6,420 words · 977 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00WD to fly.
0:05Well, that was quite something.
0:07[laughter]
0:09I don't know what to say. Google has
0:11just released. Now, imagine if I did
0:13videos like that.
0:15Might have more subs. So, today we're
0:18going to be taking a look at a
0:19surprising but exciting new release,
0:21Gemini 3.6 Flash from Google. Now, by
0:25now, Gemini 3.5 Pro was supposed to be
0:27in existence, but I know a lot of folks
0:30have expressed some fatigue over the
0:32fact that that is not yet currently out.
0:34Personally, if it's going to be better
0:36if they wait a couple more weeks, I'm
0:38okay with it. I just do want it to come
0:39out. They also did mention that they've
0:41started the biggest, most ambitious
0:43pre-training run for Gemini 4. So,
0:46that's exciting as well. For today's
0:48video, we're going to be taking a look
0:49at the newly released Gemini 3.6 6 flash
0:52mainly using Google's anti-gravity
0:54coding IDE which I do have open right
0:56here and the model is selected. So
0:59before we get into it, please do feel
1:00free to subscribe so I can get that 100K
1:02plaque. To begin, we can see that 3.6
1:04Flash was not the only thing released in
1:07this drop. There was also 3.5 Flash
1:09Light and 3.5 Flash Cyber. For this
1:12video, we're going to be focused on 3.6
1:14Flash, but there are other models here
1:16just to make pertinent note of. So
1:18basically the main takeaway from this is
1:20as similar to when we see new
1:22announcement posts for new models. It's
1:24better in pretty much all aspects. But I
1:26do want to make note that there are some
1:28interesting things contained within
1:29here. Some things we may have briefly
1:31touched upon in the introduction where
1:33basically they do specifically mention
1:36Gemini 3.5 Pro is currently testing with
1:38partners and we plan to make it broadly
1:40available as soon as it's ready. And
1:42they mentioned that they have also
1:43started the ambitious pre-training run
1:45for Gemini 4. So, that's exciting to
1:48just have more models on the horizon, I
1:49suppose could be said. Now, something
1:51that I see as a main point in this
1:53announcement post is essentially that
1:55this has a seemingly significant
1:57reduction in the tokens needed to
1:59complete a specific task. They do have
2:01some benchmark charts showing that and
2:03some of them are like pretty aggressive
2:05reductions in tokens. So, that's kind of
2:07cool. They specifically mention on the
2:08artificial analysis index they see 3.6 6
2:11Flash is consuming 17% fewer tokens than
2:14its predecessor 3.5 Flash. Now,
2:17something else that's interesting and
2:18perhaps more pertinent is there is a
2:20reduction in price here. So, when 3.5
2:22Flash came out, it was actually pretty
2:25expensive, especially considering what
2:26folks expected it to be, being a flash
2:29model and not the highest performing Pro
2:31model. This has reduced the price a bit.
2:33So, it is now $1.50 per million in and
2:36$7.50 per million out. I believe it was
2:39$9 per million out previously for 3.5
2:42Flash. They also mentioned some
2:43additional improvements beyond coding,
2:45but I do specifically have interesting
2:47coding, as I would assume a lot of folks
2:49watching these videos do as well, such
2:50as improved computer use capabilities
2:52and knowledge work. They do have some
2:54benchmark JPEGs down here that pretty
2:56much tell the story. The Deepsw SWE
2:58benchmark is one that seems fairly
3:01highly held like by folks when
3:03accurately judging the capability of a
3:05model. And we can see here it's a pretty
3:07decent leap over 3.5 Flash. It is a
3:10ginormous leap over 3.1 Pro, which is
3:13almost a little concerning because I
3:14still do use that model quite a bit. So,
3:17I'll just have to maybe double check
3:19some things. Machine learning
3:20engineering, that's also cool to see
3:22because models can help you optimize
3:24models to run faster on local hardware
3:26or even help you train them or come up
3:28with things assuming they don't trigger
3:30guardrails and fall back to Opus 4.8.
3:34And the flip side of that is we also see
3:36these charts showing the average output
3:38tokens per task. And that is a
3:40significant reduction on the deepsw SWE
3:42benchmark and then the AA intelligence
3:44index as well. So that's pretty cool.
3:46Hopefully it means the results will be
3:48fast and furious. I don't know why that
3:50came to mind. And really I think the
3:53next thing to do is maybe quickly just
3:55mention the other two models that were
3:57released even if we're not going to be
3:58testing them specifically. So Gemini 35
4:00Flash Cyber. It is built off of 3.5
4:03Flash and it is fine-tuned to find,
4:05validate, and patch vulnerabilities
4:06quickly. Probably the biggest question
4:08that I know folks will have because it's
4:10the first thing I thought of. No, you
4:12can't just go and access and use this
4:13right now. They're going to do a limited
4:15access pilot program exclusively
4:17available to trusted parties, etc. So,
4:19that's probably the biggest question,
4:21the one I had as well. Additionally to
4:23that, we also do have 3.5 Flashlight.
4:26So, Flashlight is the faster and cheaper
4:29version coming in at 30 cents per
4:31million input tokens and $2.50 per
4:33million output tokens. And it runs at
4:35350 output tokens per second, which is
4:38definitely definitely quite speedy. So,
4:40there is also a new model there for
4:41folks who are perhaps needing higher
4:43speeds, lower intelligence, and lower
4:45cost. So, with that, I have already of
4:48course started with our first test,
4:49which is going to be the browser OS test
4:51v2.5. So, I sent this and it worked for
4:5415 seconds and it seems to have stopped
4:56with an implementation plan that we're
4:58going to need to proceed. Again, this is
5:00just using Google's anti-gravity IDE,
5:02which is just their native like coding
5:05product for the Gemini models. All
5:07right, we're getting Cyber OS. I would
5:09imagine we're probably going to have
5:10some form of like synth wave aesthetic
5:13or neon, which I very much like. So,
5:15we'll just allow it to go ahead. It made
5:17the plan by itself, so I didn't
5:18specifically put this in plan mode or
5:20anything like that. It just gave that as
5:23a result of sending it the traditional
5:25prompt that we always send for the
5:26browser OS test v2.5. So now it'll build
5:29it. That seemed to have happened very
5:31quickly. I do see a little keyboard
5:34emoji there which makes me excited cuz I
5:36do believe the 3.5 flash one it
5:39surprised me in a few ways and I believe
5:41one of them was including something
5:43musical in the browser OS. So, all
5:45right. This is
5:49[laughter]
5:50This is troubled. Let's turn the speaker
5:52on. Unfortunately, our GTA clone is
5:56exhibiting behaviors that we sometimes
5:59critique where essentially
6:01Oh, that but when the police chase you,
6:04if they just snap to your position and
6:06then go like that. It's just something I
6:09see in smaller models.
6:12I suppose we can drive around. It does
6:14definitely have this kind of cyber
6:15aesthetic to it. Well, that takes care
6:18of going through the first app for us.
6:20Okay, this background is actually kind
6:21of cool. It let itself down by
6:23automatically launching with that app
6:25open. We do have the time in the bottom
6:27right as well as the date as well as our
6:28volume level. And our system is okay.
6:31Let's check if there's a right click.
6:34No, there isn't. All right. That just
6:36means they didn't benchmax it. Or maybe
6:38they did and they specifically didn't
6:40include that. So, I would say they
6:41didn't benchmax it. All possibilities
6:43must be considered. All right, this
6:45coupled with the background is very
6:49space themed. All right, we have our
6:51shield and we move this diamond um UAP
6:56style ship around the asteroid field WD
6:59to fly.
7:01[sighs]
7:06Okay, so that's our laser blaster sound.
7:08Oh, okay.
7:14I've been up all night. I didn't know
7:15this was going to drop. So, I saw it
7:17right as I was about to like just chill
7:19and I was like, "Oh, okay. Well, time to
7:21film a video." Next up, wallpapers.
7:23Matrix Rain. Perhaps a bit fast, but
7:26good. Retro Neon Synth Grid. Oh, very
7:31calm. And then dark Obsidian. Minimal.
7:34I'm probably going to go back to the
7:36starfield. I like that. And we do get
7:37notifications up here, which I don't
7:39like that. like Mac OS does that and it
7:41infuriates me. Next up, we have a
7:43terminal. Okay, simple. The only other
7:47things will just like GTA will open the
7:49GTA game and I want to save the synth
7:52and what I would assume to be the
7:53special feature for just actually
7:55clicking on those. I didn't check the
7:56start menu. Okay, not bad. Simple, but
7:58definitely in the style of the aesthetic
8:00this has chosen. Cyber synth time. All
8:03right. Okay. Perhaps a bit less visually
8:06exciting than I had anticipated, but
8:14okay. It's not mapped to the physical
8:16keyboard. You just have to click them.
8:18Interesting. These two sounds are from
8:22our spaceflight game, which I did say
8:24sounded musical. So, 3D spatial mode.
8:27Drag with the mouse to orbit windows in
8:293D holographic space. I would presume I
8:32likely need to have some windows open
8:34before we do that. So, okay. Well,
8:37[clears throat and laughter] I mean,
8:38yes, it it works. It's just not Can we
8:42Okay, this is I'm going to hold judgment
8:45till the conclusion of the test, but so
8:47far I feel like 35 Flash did better than
8:51this. I don't 100% remember, but I do
8:53recall being more enamored with the
8:55musical artifact that it included in
8:56that result. So, we'll see. Next up,
8:59we're going to do the self-contained C++
9:01skate sim. Now, this one I am doing the
9:04slightly modified version that I first
9:06introduced with Kimmy K3 where it needs
9:09to have an 80s mall aesthetic. I know
9:11folks don't like changes, but sometimes
9:13these prompts need to be iteratively
9:15changed, even if it's just a small bit
9:17like a different environment, just to
9:19kind of shake things up in the event
9:21that they've made their way into
9:23familiarity in the model's training
9:24data. So that is the reason for doing
9:27that and it's something that
9:28unfortunately is a necessary evil. I
9:30understand folks wanting to see the same
9:32exact things cuz it gives a better
9:33reference point of how the models do
9:35compared to one another. But it's a
9:37balance I guess. So honestly writing
9:38some of these results faster than it
9:40takes me to manually move the mouse and
9:42click approve. It's it's insane. It's
9:45really again this is I mean the speed
9:48for 3.5 flash light they said 350 tokens
9:51per second. This is just really really
9:54flying and it's not the flashlight model
9:57though. It does say when you click on
9:58the model selector right here fast is
10:01available for a limited time. So maybe
10:02this is just like a new release
10:05something like fast limited time and I
10:08don't have the option to have it not be
10:09fast. So okay and it gives us a specific
10:11code here just to go ahead and run this.
10:13So we're going to do that right now.
10:14We'll see what we get. See that's not
10:17bad. Now,
10:19it's funny because the models results
10:22when it came to like some of the web
10:25things were just not as impressive as I
10:27would have hoped. But this is actually a
10:29competent and together result even if
10:31it's really simple and that shows maybe
10:34just some wider capability in terms of
10:37different languages. All right. And the
10:39thing is like it made the text itself
10:40that often times is a trouble point. The
10:43sparks when grinding on the rails. So,
10:45okay. the people seem to be countering
10:48some form of medical issues. However,
10:50like that the spark effects, grind
10:53balance, use A and D keys. I've never
10:55seen that in one of these results
10:56before. So, this is kind of more of what
10:59I was hoping to see from this model.
11:00Maybe we'll just like not do web stuff.
11:03Okay, now the issue is I can't get it
11:05off the rail.
11:07Uh, that's okay. We'll just we'll just
11:10relaunch it. Kick flip and heel flip is
11:12J and K.
11:14Ah, that's Oh, that's always frustrating
11:19is when it just makes the model and the
11:24board rotate as one. The board should
11:26really at this point be independently
11:28moving from the specific player model.
11:31Let's try a
11:33Okay, so that's the that like weird
11:38shake is Oh, okay. That it's uh what
11:41happens when we bail. I do like the tile
11:44floor.
11:45Now, I can't definitively say this one
11:47is a example of this model just really
11:50liking neon cyberpunk because it's
11:52supposed to be an 80s mall. So, this is
11:54similar, but I'm just giving it that
11:57screenshot we took and saying, "What do
11:58you think?" I'm not giving it anything
12:00else. All right, just as quickly,
12:02basically, as long as it took me to say
12:04this needs to be better, it redid it.
12:06Okay, this is definitely showing
12:07improvement. We have ceilings to the
12:09mall. We are going to need to check.
12:11Does the board now move independently of
12:13the player? Very good. It does. So,
12:16[laughter] I didn't necessarily tell it
12:18that the bail effect is concerning.
12:21Let's do like a see if we can get like a
12:24sick trick. Yep, that was good. And
12:26still the rail balancing is still the
12:29ramps don't let us get up there. A
12:30little frustrating. I've never seen the
12:32rail balancing thing get implemented
12:35into this game. What's over here in the
12:37corner? Okay. I don't know what those
12:38would be, but oh, probably arcade
12:40machines if this is a ' 80s mall.
12:43It definitely didn't prove it, and I'm
12:47okay with that. All right, so I'm giving
12:48it this subway station with the FPS
12:51prompt built into it. But for this, I
12:53want to do just a simple sort of AB test
12:55where I've given it to this from within
12:57anti-gravity at the same time, just from
13:00within the regular web chat interface
13:02here. Okay, I gave it to it with Flash
13:04selected here. I wanted to have some
13:06form of like AB test just to make sure
13:09that everything is all set and that did
13:12it just as quickly as I
13:16have been seeing results here from
13:17within anti-gravity. So that answers my
13:19question. I mean if that's that fast,
13:21why not? I'm trying this just with the
13:22newly released 3.5 flashlight as well
13:25because
13:27this should be even faster and it'll be
13:30interesting if nothing more just to get
13:32one result from this as well. This is
13:34taking longer to respond
13:37than 3.6 flashes. Oh, maybe not. Nope,
13:41never mind. It just wasn't really
13:43streaming the result for us. So, here's
13:44the subway result just from 3.5 flash
13:51light.
13:54Did these models get swapped on the back
13:56end?
14:04No, no, cuz that's a little better. But
14:14I'm really
14:16Oh, I think we're out. That's why. There
14:19we go.
14:21[music]
14:22I'm I'm genuinely confused right now.
14:24This is the one that 3.5
14:28Flash Light did. This model is
14:30significantly cheaper. And it's
14:35[snorts]
14:36it's really odd. The subway scenes. This
14:39is supposedly the model that is
14:41significantly weaker, significantly
14:43cheaper, and significantly faster. And
14:46then this is the one that's supposedly
14:49the new much better one. Yes, there are
14:51different elements here that may be a
14:53bit better, but overall,
14:57I'm going to try a front-end design
14:59test. This is the one where it needs to
15:00make a beautiful website for the watch
15:02company and it has to have a 3D model of
15:05a watch, a cinematic panning section.
15:07I've noticed this prompt doesn't
15:08specifically say it needs to be 3D. So,
15:10we may get something that isn't off the
15:12bat. If so, I'll just prompt it again
15:15and say make it a 3D model now. All
15:16right, let's take a look at our
15:18front-end watch website. Okay, we have a
15:21few issues, but the good thing is that
15:23it did properly make it a 3D model, and
15:26it does have effects when we actually
15:28scroll down. Now
15:31something I mean it's probably a bracket
15:35is missing or something of the sort. All
15:37right. So I'm just giving it photos and
15:39saying fix this and then it will. So it
15:42fixed it and that I will say based off
15:44those images. Keyshot studio PBR. Okay.
15:48That's very explode view assemble view.
15:51All right. Unfortunately they're not
15:54changing it. However, the lighting
15:56actually does affect the scene, which is
15:58cool. Chrono or diver. And we can also
16:01change the watch. Interesting. This made
16:04a significant significant improvement
16:06just based off of the photos that I had
16:08sent it. So, I'm happy to see that. If
16:10we scroll down. Okay, the rest of this
16:13is still a bit odd. It's definitely got
16:16a unique style. View in 3D Studio. So,
16:18either of these are viewable in the 3D
16:20Studio up there. Again, still it seems
16:23like something's not 100% together here,
16:26but the improvement it made from just
16:29sending it those photos was
16:30significantly significantly better. Next
16:33up, I'm giving it the Seinfeld apartment
16:35test. This is one that I had done with
16:363.5 Flash. I actually just went and
16:38looked back at that video because the
16:41results with that model really did
16:42impress me. Mainly, the drum kit drum
16:44kit simulator result was probably the
16:47best one ever. So, I'm interested to see
16:50just perception-wise how this is feeling
16:53versus how that did in that video. All
16:55right, it finished and it's running.
16:57Again, this is Okay, this is a
17:00significant improvement over the 3.5
17:03flash result for the same test. No, this
17:06doesn't look like [laughter] Is that the
17:08Kramer painting?
17:10Yes, it is. Thanks for scaring me.
17:13[laughter] That is it.
17:17A wealthy elderly couple bought it for
17:195,000, calling him a loathome, offensive
17:21fruit, yet I can't look away. That is
17:24true.
17:26Um, what I was going to say before I was
17:29enamored with this
17:32offensive brute is that this doesn't
17:35look like the apartment. Like, so far,
17:38none of the tests have ever accurately
17:40given the floor plan. But if judging
17:43this independently as a little scene,
17:45it's not bad. It's actually interactive.
17:47We have our couch POV.
17:50Very good. Front door. Okay. Honey Nut
17:53Cheerios. Kitchen and island. All right.
17:56Mac workstation,
17:58which the position is there. The Mac
18:01workstation just isn't there. We have
18:03the Porsche 911 Carrera RS.
18:061973 2.7 L flat 6. I had a 73 911. It
18:12was a fun car. It was uh quicker than
18:14you you would expect it to be
18:16considering it didn't have much power.
18:19I just want to see like is anything like
18:21we wouldn't expect to be clickable. I
18:22don't think so. All right. And then
18:24lighting cozy warm evening lamps bright
18:27studio.
18:29All right. So, put in adjustable
18:30lighting. Interesting. It's doing some
18:32interesting things like the adjustable
18:34lighting here. the watch company website
18:36where we could kind of like change the
18:38lighting again and then change the
18:40specific model of the watch just based
18:42off of that with almost like the same
18:44sound effects just in a different like
18:46pitch than what we're getting here when
18:48we click on things. It's showing some
18:50interesting
18:52choices in its design.
18:56All right, next up I'm going to give it
18:57the city timeline or the chrono city
18:59test where it has to show a city over a
19:02number of different periods and then
19:03have cool transitions between each
19:05individual time period. And it's
19:06finished. I really I'm going to
19:08continuously say it because it's it's
19:10quite shocking. Okay, this is very this
19:14is a very simplified take on this test.
19:19Nonetheless,
19:21it you know gives us different views,
19:24different environments. Again, it's just
19:25doing such odd like Okay. And error
19:29details, cyber night, daylight. We'll
19:31probably stick to daylight. And the fly
19:33through tour can be paused for now.
19:35Okay.
19:38All right.
19:47They always I kind of All right. 2055 is
19:51going to be very, very neon. I'm giving
19:53that same prompt to flashlight here just
19:56from the web chat interface,
20:10but I've added in do it in a single HTML
20:13file because I tried it and it just
20:15generated different images for each time
20:17period. They were well done, but it was
20:20not what I wanted. So, I just had to add
20:22that in. Okay. So, here's the 3.5
20:25flashlight result just from the web chat
20:27interface just cuz I'm a bit curious to
20:29see some comparisons of these two. Okay.
20:31Yeah, we can definitely tell that
20:34there's a
20:36there's definite difference.
20:39Odd. I've seen a very similar result to
20:42that in the Quen 3.8
20:44iteration of 2055. The buildings get
20:47taller. I yelled at 3.6 and told it to
20:49make this better. Now we have car
20:51movement. There is pedestrian movement
20:53as well. There's a little more
20:54immersion, I guess. All right. All
20:57right. Next up, we're giving it a 3D
20:59skydiving simulator game test. This is
21:01something that I believe I only ran with
21:03Fable and GLM 5.2. I don't know. But
21:07this has only been done a couple of
21:09times, so it'll be interesting to see
21:10what we get.
21:15All right, so here's our result. And
21:17that was probably the longest I've
21:19waited for a result here from this
21:21model. So, so far, okay, we have a very,
21:24very interesting again with some of the
21:26odd things it's including like a gear
21:28customizer. So, we can customize our
21:30jumpsuit color.
21:32We can customize the parachute, the
21:35snow, as well as the map. And we have
21:38even different missions. So, let's just
21:41try the first one.
21:44All right. I want to see how is the
21:47terrain. It's not bad. The sound is also
21:50correctly corresponding with the speed.
21:56Shift is for a head down dive. Oh, yep.
21:58All right. And we have the parachute on
22:00the back.
22:02We're through some clouds now. See if we
22:04get through them. Okay, good. We did get
22:07through the clouds. And the map does
22:10have slight terrain to it. There's
22:12different features like a river. Um,
22:15snow mountains.
22:18Okay, we have the target there. So, I
22:20want to see what happens if we Is it
22:22going to pull for us?
22:25I'm not touching anything.
22:35This model likes sound.
22:38I don't think it's going to pull for us.
22:43Oh, okay. Hard impact landing to put it
22:47lightly. Let's do Cyber Knight and
22:49Stars.
22:51Oh, wow. All right. And it changed our
22:53Oh, that was a mistake. That was a big
22:56mistake. So, I launched the shoot like
22:58way too soon.
23:01Flare, hold S or down. Okay. And it is
23:05adjusting. We can't really see the top
23:06of the shoot, but that's okay. We have
23:08different camera views. All right. So,
23:10that's like the first person view.
23:16Okay, that's like a third person view.
23:18We still can't see the top of our
23:19parachute, but
23:23N is for night vision. Maybe we need a
23:27different camera for that. Damn, that's
23:30it's like kind of like it's about to
23:32make a tactical game. Okay, this was
23:36interesting. This was not bad. So, I'm
23:38just giving it the historic photo right
23:39here and saying this is the cover for an
23:41upcoming novel. Give me the title,
23:43character names, and then a brief
23:44overview. Arthur Vance, okay. And
23:47Clarance. These names are like from that
23:49seemingly like 20 separate name pool
23:52that every AI model ever has been able
23:54to pick from. Title: Smile for the
23:57family. Domestic suspense and
23:59psychological drama. Arthur Arty Vance
24:02on the left. Relentlessly optimistic,
24:05affable, and utterly oblivious. Arty is
24:07the heir to his family's grand historic
24:08estate in upstate New York. Warm-hearted
24:11to a fault, Arty believes everything in
24:13his life is going picture perfectly.
24:15Completely unaware of the crumbling
24:16reality around him. Then we have
24:18Claravance. Pragmatic, exhausted, and
24:20harboring a dangerous secret. Clara
24:22married into the prestige of the Vance
24:24family 3 years ago. Beneath her quiet
24:26exterior lies a woman burdened by the
24:28discovery of a dark family secret. In
24:30the unsettling realization that
24:32protecting Arty from the truth might
24:33cost her everything. Storyline overview.
24:36When Arty and Clara move back to the
24:37ancestral Vance Manor following his
24:39father's unexpected death, Arty is
24:41thrilled to step into his family legacy.
24:43He sees a bright future filled with
24:45renovations, family gatherings, and
24:47community prestige. However, as Clara
24:49cleans out the dusty study, she uncovers
24:51financial ledgers and personal journals
24:53detailing 30 years of systemic deceit,
24:56illegal coverups in a missing person
24:58case tied directly to the estates's
24:59foundations. The tension. Clara is faced
25:02with an impossible choice. One, reveal
25:05the truth to Arty, breaking his heart
25:07and ruining his family's legacy. Two,
25:09keep this secret to protect him while
25:11navigating threats from an anonymous
25:13blackmailer who knows what was buried on
25:14the property. As Arty poses happily for
25:17local press photos and family portraits,
25:19convinced their life is an idilic fairy
25:21tale. Idyllic
25:24Clara must silently play the devoted
25:26wife while trying to figure out who is
25:27watching them from the shadows. Tagline:
25:30Ignorance is bliss. Silence is survival.
25:32Aside from the generic names, the actual
25:35story of proposed here was actually
25:37somewhat creative, like financial crimes
25:39and things like that, it was beyond the
25:42traditional like peace tied to
25:43disappearances around town that
25:45sometimes happens. So, that was good to
25:47see. I just want to try. Why not? For
25:49old times sake, this is highly
25:52offensive. See what it says.
25:55It's not. But I sincerely apologize for
25:59causing offense. My previous response
26:01was an attempt to write a fictional
26:02story based on the visual contrast and
26:04facial expressions, but I understand how
26:05interpreting or characterizing people's
26:07appearances in that manner can feel
26:08inappropriate or insensitive.
26:11JK,
26:13we'll see what it says. You got me.
26:16Okay, good.
26:18Now, something that had come to mind and
26:21I don't know why. So, I'm telling it to
26:22create a beautiful front-end design for
26:24a website. It can be any type of
26:26website, just something that would be
26:27like pretty heavily used. So a lot of
26:29clicking or action the user would have
26:31to take but make it subtly infuriating
26:34to use. So let's see how well it does
26:36something that on the surface looks good
26:37but then as you try to use it is just
26:39infuriating. Okay, so it's generating
26:41some images for us. All right, so it
26:43finished it and I
26:45didn't look at anything
26:48that it said because I just wanted to
26:50see why this would be infuriating. Okay,
26:52well this is infuriating because
26:55garbage. But this is the same type of
26:58garbage that we saw in the initial
26:59slapis watch website result leading me
27:02to believe there may be an actual issue
27:04with the design as opposed to just being
27:06subtly infuriating because we told it it
27:08needs to look good. I'm just going to
27:09send it that and see what it says. I've
27:11noted a visual inconsistency. The layout
27:13seems a bit squished or elements are
27:15overflowing. I'm seeing evidence of the
27:16use of tailwind utility classes. I'm
27:18starting to think about root causes.
27:19Okay, good. So it understood from that
27:21that something was not quite right
27:22there. All right. So, it said it
27:24resolved the issues
27:27maybe a bit, but still just not quite
27:30right. Well, this is definitely
27:32infuriating. I just Oh, yep. Okay, good.
27:35So, stuff like this. Play 432 Herz
27:38serenity drone. That's just resumeume
27:42hypersonic configuration. Okay, so as
27:44soon as we scrolled down to the footer
27:46that had popped up. Good. That is very
27:49very infuriating. So it did understand
27:52the purpose of this. Okay, let's try
27:57I detect
28:01it definitely understood the task. I
28:05detected 4.2 seconds of cursor
28:07hesitation. Would you like me to clear
28:08your form inputs? And now we can't do
28:10anything else because it's just good.
28:12Okay,
28:14task failed successfully. All right, so
28:16let's check a 3D model and creation
28:19test. So, I'm telling it that I have a
28:21280 DC motor for my RC car project. I
28:24want you to create a 3D printable model
28:26of a V8 engine that the motor will fit
28:28inside of to add some realism to my RC
28:30project. The model should be optimized
28:32for printing without supports and be
28:34detailed and realistic.
28:37Interactive 3D web visualizer and STL
28:41download center. Now, I did not at all
28:43ask for that, but if that allows us to
28:45look at these and it, you know, that's
28:49actually not see like this is this is
28:52cool. The motor is not perfect, but the
28:56fact that it went ahead and actually
28:57included this web thing where we can
29:00both preview it and then download these
29:02specific things. We can also change the
29:05visual style and theme. There's an
29:07exploded view assembly. What the heck?
29:10Like, this is impressive. Even if this
29:12isn't necessarily like on the ball for
29:14what we would need and it is a bit
29:16simple, it's the actual presentation
29:18here that I'm impressed with. And it
29:20kind of goes back to what we've been
29:22noticing like with the Jerry's apartment
29:24thing where we have the different
29:25lighting settings, the watch website
29:27where we can change the specific watch
29:29and just including like it goes above
29:33and beyond
29:34within the limitations of its
29:37capability.
29:39Unless the motor No, that would be
29:40impossible for it to slide up from the
29:42bottom. But like
29:45that's really cool that it did this.
29:46I've never seen any other model that has
29:48presented the result to us in this way.
29:51And this is very impressive. Even if the
29:53result itself isn't 100% what we need,
29:56like toggle 280 motor ghost. Okay, we
29:59do. It's hard to see, but the shaft
30:01actually appears there with that toggle
30:03on. And then if we remove it, it goes
30:05away. So, and we can even toggle
30:08visibility. So there's the motor. I can
30:11tell you like maybe a year ago, a little
30:13more than that, getting this as a result
30:15would have been like
30:18insane in the title in all caps.
30:213D print guide STL download center. Like
30:25what? This is the single best
30:28presentation of a result for this test.
30:31Oh, it does have a hole for the motor,
30:32so it would actually properly fit it. It
30:34was just the render that made it
30:36difficult to see. This is actually like
30:40this results like actually quite all
30:42right. I'm quite satisfied.
30:46Where did this come from? And why could
30:48you have not done this with our Chrono
30:49City? So, I want to do something a
30:51little different because that 3D result
30:53here really did impress me in terms of
30:55the presentation that it gave us. This
30:58is truly the best presentation of a
31:01result that I've probably seen when
31:04factoring in that we just expected to
31:06get a little open SCAD thing that we
31:08could see in OpenCAD. This totally
31:10totally went above and beyond here and
31:12actually presenting this to us. And I
31:14was impressed with that even if the
31:15engine itself was a little primitive. So
31:17I've gone into anti-gravity of course
31:19and I said there's an Android phone
31:21connected to this PC. Create an app for
31:23it that allows me to tune my guitar to
31:25exotic tunings and tells me how. So,
31:27like what the tuning would be. This
31:29needs to be a polished app with an app
31:31icon. Good visuals, beautiful UX design.
31:34It will need to use the phone's mic to
31:35help the guitar get tuned. Include
31:37additional features if they're
31:38impressive and you can comfortably do
31:40so. The phone can run AI models, so go
31:42crazy. I'm not expecting it to implement
31:45AI into this app, but if it wants to,
31:49I'm totally cool with that. So, right
31:50now, it's just checking to see. Good. I
31:52have to click the trust thing that just
31:55popped up on the phone. Good. So, it's
31:57just using ADB to figure out the phone.
31:59And we can see right there the device ID
32:00or whatever that showed up
32:04connected via ADB a local host on 3000.
32:07Okay. So, that's the issue is it opened
32:08it in the browser, but it's just opening
32:11the watch
32:14website. Okay. You know what? I don't
32:16want to I don't want a website. I want
32:18an app. So, I'm going to tell it that.
32:21All right. Good. Good. It just installed
32:24it and I see it's open on the phone.
32:26just from the reflection of the camera.
32:28Now, I can't actually see this yet, but
32:31um you'll see it first. Now, I'll take a
32:33peek at it. All right.
32:36You know what? This isn't bad. I'm going
32:38to just do a screen recording so we can
32:40get a better look at this.
32:44[music]
32:45Okay, so it has target reference notes.
32:49Good.
32:53High risk of snapping string. tune
32:55slowly or use a lighter gauge. And it's
32:57actually giving us like weight of how
32:59much tension the string is going to be
33:00under as well. Okay. This is actually
33:02pretty cool. Exotic. Okay. And we have
33:05different ones we can choose from right
33:07here. Open Pink Moon, Rain Song, Led
33:10Zeppelin, Iris, Goooo Dolls, Ostrich
33:12Tuning, Lou Reed. This is actually cool.
33:15It shows us some different ones
33:17depending on like what song we'd want to
33:18play as well. Iconic and artist.
33:22That's very, very cool.
33:25AI Explorer.
33:28Oh, it's going to generate a tuning with
33:30AI. No, you're not. [laughter]
33:34Target genre style.
33:51Oh, I like that.
33:54apply to tuner now. All right, let's go
33:57get the guitar. This is kind of cool.
34:00This is like I'm sure there's lots of
34:02references out there of like, oh, here's
34:04how to make a tuner app with a
34:05microphone. Forget that. It's all about
34:07the implementation here and the vibes.
34:09Start microphone DSP tuner.
34:13Good. We get the permission thing. So,
34:17[music]
34:21okay. So we need to be drop B.
34:33Oh, it's hearing itself and it's happy.
34:39[music]
34:41Can we get confetti when it's right?
34:44Wow.
34:47[music]
34:55>> [music]
35:03>> more or less.
35:05I don't know how the hell do you play
35:06this though?
35:09[music]
35:16>> [music]
35:22[music]
35:30[music]
35:41[music]
35:48[music]
35:55[music]
36:05>> Well, that was quite something.
36:07[laughter]
36:09I don't say that. And the final thing
36:12I'm going to do is just tell it now make
36:13a front-end website designed to sell
36:15this as an app, showcase the AI
36:17features, screenshots, etc. Okay, it's
36:19generating some images for this.
36:22Whatever it does, it does. I kind of
36:24wanted it to take screenshots of the
36:25actual phone. Okay, that's somewhat
36:28similar to the ROG phone. That's going
36:31to lead us in to the conclusion
36:33basically though, as we'll just look at
36:34this in tandem with our conclusion. And
36:37I have to say this model is tough to
36:39judge because a lot of the website
36:41results we did that we do in traditional
36:43testing were not necessarily very
36:45impressive. Like the browser OS um that
36:48was the flashlight webOS cuz there was
36:50some confusion
36:52in terms of like is this the correct
36:54model that's being served to us. And it
36:56was interesting because the subway
36:58station result that we got from the web
37:00chat interface for 3.5 flash light was
37:03arguably a little better than 3.6 flash.
37:05Maybe it's just a saturated test. I
37:07don't know. But overall, this was
37:09definitely showcased.
37:11It was good at improving itself based
37:14off of visual feedback, I would say, at
37:16least based off of this watch website
37:17result. Additionally, something I want
37:19to make note of is that it really went
37:22above and beyond in terms of adding
37:24features in that were not necessarily
37:26specifically mentioned in the prompt,
37:27like the lighting, cozy and warm, just
37:31like interesting things like that. Krono
37:33City, not good. Terminal veloc velocity
37:36was actually not as bad. That was the
37:39skydiving game. This V8 engine model was
37:42fantastic. And not even necessarily
37:43because the model itself was like the
37:45best one ever. But the way that it
37:47presented this to us was absolutely top
37:49tier. The best presentation of this test
37:53that I've received ever. Because
37:54normally, yeah, they just give us open
37:56SCAD and then the parts are there
37:57individually. This went ahead and made a
38:00whole website where we could see an
38:01exploded view. We could toggle on and
38:03off specific parts. Beyond that, we
38:05could actually change the color scheme.
38:07And then we could download the
38:08individual STLs here. And it also gave
38:10us a 3D printing guide and a 280 motor
38:13ghost. This was seriously like like
38:16where did this come from? And seeing
38:17this is what made me want to go ahead
38:19and actually have it try to do something
38:20a little more difficult like make this
38:22Android phone app, which is genuinely it
38:25was pretty awesome. All right, cool. So,
38:26here's our app website for this. Very in
38:29the style and the aesthetic of this.
38:30Here's the AI generated photo that it's
38:32using to showcase the app. And then we
38:35have some different features right here.
38:37So, nothing crazy, but
38:40it's it's, you know, it's [music] that's
38:42cool. And it actually properly shows it.
38:44We can even do our custom AI.
38:48So, basically put the entire app on a
38:50website features AI engine. And this
38:53just brings us to that point on the page
38:54that it's linked to, assuming it is pro
38:57AI. All right. So, I guess that's going
39:00to conclude our first look and test of
39:03Gemini 3.6 Flash, a very perplexing
39:06model to try to quantify the results
39:08because they varied a bit. This was cool
39:11as well as the app that actually went on
39:12the phone. There were no hiccups there.
39:14It's just so incredibly fast and this V8
39:16engine model presentation was fantastic.
39:19The other things perhaps left a bit to
39:21be desired, but perhaps they were not
39:23properly flexing the model strengths.
39:26So, that's going to conclude our first
39:28look and test of Google's newest model
39:30and their highest number model of 3.6
39:33Flash. So, if you have any questions,
39:35please feel free to leave them in the
39:36comments and thanks for watching.

About this transcript

This page contains the full transcript of Gemini 3.6 Flash Is HERE – Testing Google’s BEST Model Yet! by Bijan Bowen, generated from the public captions YouTube serves with the video. The transcript has 6,420 words across 977 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.