Android Impressió 3D Gemini 3.6 Flash Google AI programació amb IA agents de codi prototipatge

Gemini 3.6 Flash a prova: velocitat espectacular, resultats irregulars i dos prototips brillants

Una prova extensa de Gemini 3.6 Flash amb webs, C++, jocs 3D, un model imprimible i una app Android: molt ràpid i proactiu, però irregular.

Què aporta Gemini 3.6 Flash

Google presenta Gemini 3.6 Flash com el model de treball general de la família: més fort en programació, tasques agentives, coneixement i multimodalitat que 3.5 Flash, però orientat a velocitat i cost per sota d’un futur model Pro. El vídeo de Bijan Bowen el prova durant gairebé quaranta minuts amb Google Antigravity i una bateria de prototips.

La fitxa oficial admet text, imatge, àudio, vídeo i PDF com a entrada, i text com a sortida. Té una finestra d’1.048.576 tokens, un màxim de 65.536 tokens de sortida, raonament, execució de codi, cerca, crides a funcions i ús de l’ordinador en vista prèvia. No genera imatges ni àudio directament.

El preu és d’1,50 dòlars per milió de tokens d’entrada i 7,50 per milió de sortida. Google diu que produeix un 17% menys de tokens que 3.5 Flash en l’índex d’Artificial Analysis. L’objectiu, per tant, no és només respondre ràpid: és completar bucles d’agent amb menys passos i menys text.

Una prova útil, però no un benchmark controlat

El creador reutilitza prompts habituals del canal: un sistema operatiu al navegador, jocs 3D, una web de rellotges, un pis de Seinfeld, una cronologia urbana, un model imprimible i una aplicació Android. Això permet veure patrons, però no produeix una classificació científica.

Alguns prompts canvien respecte de proves anteriors, hi ha correccions manuals i les comparacions amb Flash-Lite es fan de vegades en una interfície diferent. Tampoc es mesuren tokens, cost, temps total ni taxa d’èxit sobre moltes repeticions. El resultat és una primera impressió rica, no una prova estadística.

La velocitat percebuda és excepcional: el model escriu codi gairebé tan de pressa com el creador pot aprovar accions. Artificial Analysis mesurava al voltant de 235 tokens de sortida per segon. Això no inclou necessàriament el temps fins al primer token, l’execució d’eines o les reparacions.

Webs vistoses, però amb qualitat irregular

El primer «sistema operatiu» de navegador té una estètica ciberpunk, fons configurables, terminal, minijocs i un mode espacial per moure finestres. També mostra defectes típics d’un prototip: interaccions simples, manca de menú contextual, controls musicals no vinculats al teclat i enemics que es mouen de manera artificial.

En una escena de metro amb un joc de trets, el resultat de Gemini 3.5 Flash-Lite sembla fins i tot més coherent que el de 3.6 Flash. Una sola comparació no demostra que el model barat sigui superior. Pot haver-hi variació entre execucions, diferències del servei o un prompt massa conegut.

La web d’una marca de rellotges comença amb errors de disposició. El punt positiu és la reparació: el creador envia captures i el model corregeix l’escena, millora el rellotge 3D i conserva controls de llum i variants. Aquesta capacitat d’iterar sobre feedback visual és més significativa que l’intent inicial.

La recreació del pis de Jerry tampoc encerta la planta real, però incorpora objectes interactius, il·luminació, punts de vista i referències a la sèrie. La cronologia d’una ciutat és funcional però massa simplificada. Gemini tendeix a afegir panells, sons i personalització, encara que la base geomètrica no sempre sigui rigorosa.

C++ i jocs: bona amplitud, física encara superficial

En una prova de C++, el model crea una simulació de monopatí dins d’un centre comercial dels anys vuitanta. Hi ha espurnes en lliscar per baranes, equilibri, trucs i una estètica coherent. Després d’una captura i una instrucció de millora, separa correctament el moviment de la taula del personatge.

La física continua sent elemental. Algunes rampes no funcionen, la caiguda és estranya i no hi ha una validació de rendiment o memòria. El valor de la prova és que 3.6 Flash pot sortir del JavaScript de navegador i muntar ràpidament una aplicació autocontinguda en C++.

El simulador de paracaigudisme és un dels resultats més complets: terreny, núvols, so lligat a la velocitat, càmeres, visió nocturna, personalització i missions. Falla en detalls de visibilitat del paracaigudes i no prova una física realista, però ofereix més estructura de producte que una simple demo.

El millor resultat: un V8 imprimible amb visor

El prompt demana una carcassa de motor V8 per amagar un motor elèctric 280 dins d’un cotxe radiocontrol. En lloc de limitar-se a lliurar un fitxer OpenSCAD, Gemini crea un centre de descàrregues amb visor 3D, vista explotada, peces activables, variants de color, guia d’impressió i STL separats.

La presentació és la més convincent de la prova i el model inclou un buit per al motor. Però el vídeo no imprimeix les peces ni mesura toleràncies. No es comproven parets mínimes, ajust de l’eix, orientació sense suports, resistència tèrmica o encaix en un xassís concret.

Per fabricar-lo de debò caldria obrir els STL en un reparador de malles, contrastar dimensions amb un peu de rei, laminar les peces i fer una impressió de prova. Un visor atractiu no certifica que la geometria sigui sòlida o imprimible.

Una aplicació Android instal·lada al telèfon

La prova més pràctica connecta un Android per ADB i demana una aplicació per afinar guitarra en afinacions poc habituals. El primer intent obre una web, però, després de corregir la instrucció, l’agent compila i instal·la una aplicació nativa al telèfon.

La interfície inclou afinacions d’artistes i cançons, advertiments de tensió, selecció de cordes, permís de micròfon i un afinador DSP. El vídeo comprova que reacciona a una guitarra i indica quan una nota s’apropa a l’objectiu. També genera una web promocional coherent amb el disseny de l’aplicació.

Queden preguntes importants. No es compara la freqüència amb un afinador calibrat, no es valida la latència en altres dispositius i les estimacions de tensió podrien dependre del calibre, l’escala i la corda. Una advertència generada per IA no substitueix dades del fabricant: una afinació extrema pot trencar cordes o forçar l’instrument.

Què revela el conjunt de proves

El patró és clar:

  • genera i itera molt ràpid;
  • respon bé a captures de pantalla;
  • afegeix funcions i presentació sense que se li demanin;
  • pot coordinar navegador, C++, 3D i Android;
  • és irregular en disseny web i fidelitat espacial;
  • sovint necessita una segona instrucció;
  • l’acabat visual pot ocultar lògica o geometria no verificades.

La proactivitat pot ser un avantatge en un prototip i un risc en un repositori real. Les funcions extra amplien la superfície d’errors, dependències i problemes de seguretat. Convé donar criteris d’acceptació, limitar fitxers modificables i exigir proves.

Conclusió

Gemini 3.6 Flash impressiona més per la velocitat, l’amplitud i la capacitat de convertir una petició en una experiència completa que per una perfecció constant. El visor del motor i l’aplicació d’afinació són demostracions fortes; diverses webs i escenes 3D queden per sota.

És una opció prometedora per explorar idees, crear interfícies i tancar bucles visuals ràpids. No elimina la revisió d’un desenvolupador, les proves automatitzades ni la validació en maquinari.

La millor conclusió no és que sigui «el millor model de Google», sinó que ofereix una combinació molt competitiva de velocitat i capacitat. En un projecte real, s’hauria de comparar amb altres models sobre el mateix repositori, els mateixos tests i un pressupost de tokens fix.

Contrast i context

Fonts consultades

4 fonts
  1. 01
  2. 02
  3. 03
  4. 04

Font de treball

Transcripció amb marques de temps

20 fragments
Consulta la transcripció
  1. 0:00 , obre el vídeo en una pestanya nova

    WD to fly. Well, that was quite something. [laughter] I don't know what to say. Google has just released. Now, imagine if I did videos like that. Might have more subs. So, today we're going to be taking a look at a surprising but exciting new release, Gemini 3.6 Flash from Google. Now, by now, Gemini 3.5 Pro was supposed to be in existence, but I know a lot of folks have expressed some fatigue over the fact that that is not yet currently out. Personally, if it's going to be better if they wait a couple more weeks, I'm okay with it. I just do want it to come out. They also did mention that they've started the biggest, most ambitious pre-training run for Gemini 4. So, that's exciting as well. For today's video, we're going to be taking a look at the newly released Gemini 3.6 6 flash mainly using Google's anti-gravity coding IDE which I do have open right here and the model is selected. So before we get into it, please do feel free to subscribe so I can get that 100K plaque. To begin, we can see that 3.6 Flash was not the only thing released in this drop. There was also 3.5 Flash Light and 3.5 Flash Cyber. For this video, we're going to be focused on 3.6 Flash, but there are other models here just to make pertinent note of. So basically the main takeaway from this is as similar to when we see new announcement posts for new models. It's better in pretty much all aspects. But I do want to make note that there are some interesting things contained within here. Some things we may have briefly touched upon in the introduction where basically they do specifically mention Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it's ready. And they mentioned that they have also

  2. 1:43 , obre el vídeo en una pestanya nova

    started the ambitious pre-training run for Gemini 4. So, that's exciting to just have more models on the horizon, I suppose could be said. Now, something that I see as a main point in this announcement post is essentially that this has a seemingly significant reduction in the tokens needed to complete a specific task. They do have some benchmark charts showing that and some of them are like pretty aggressive reductions in tokens. So, that's kind of cool. They specifically mention on the artificial analysis index they see 3.6 6 Flash is consuming 17% fewer tokens than its predecessor 3.5 Flash. Now, something else that's interesting and perhaps more pertinent is there is a reduction in price here. So, when 3.5 Flash came out, it was actually pretty expensive, especially considering what folks expected it to be, being a flash model and not the highest performing Pro model. This has reduced the price a bit. So, it is now $1.50 per million in and $7.50 per million out. I believe it was $9 per million out previously for 3.5 Flash. They also mentioned some additional improvements beyond coding, but I do specifically have interesting coding, as I would assume a lot of folks watching these videos do as well, such as improved computer use capabilities and knowledge work. They do have some benchmark JPEGs down here that pretty much tell the story. The Deepsw SWE benchmark is one that seems fairly highly held like by folks when accurately judging the capability of a model. And we can see here it's a pretty decent leap over 3.5 Flash. It is a ginormous leap over 3.1 Pro, which is almost a little concerning because I still do use that model quite a bit. So, I'll just have to maybe double check some things. Machine learning engineering, that's also cool to see because models can help you optimize models to run faster on local hardware

  3. 3:26 , obre el vídeo en una pestanya nova

    or even help you train them or come up with things assuming they don't trigger guardrails and fall back to Opus 4.8. And the flip side of that is we also see these charts showing the average output tokens per task. And that is a significant reduction on the deepsw SWE benchmark and then the AA intelligence index as well. So that's pretty cool. Hopefully it means the results will be fast and furious. I don't know why that came to mind. And really I think the next thing to do is maybe quickly just mention the other two models that were released even if we're not going to be testing them specifically. So Gemini 35 Flash Cyber. It is built off of 3.5 Flash and it is fine-tuned to find, validate, and patch vulnerabilities quickly. Probably the biggest question that I know folks will have because it's the first thing I thought of. No, you can't just go and access and use this right now. They're going to do a limited access pilot program exclusively available to trusted parties, etc. So, that's probably the biggest question, the one I had as well. Additionally to that, we also do have 3.5 Flashlight. So, Flashlight is the faster and cheaper version coming in at 30 cents per million input tokens and $2.50 per million output tokens. And it runs at 350 output tokens per second, which is definitely definitely quite speedy. So, there is also a new model there for folks who are perhaps needing higher speeds, lower intelligence, and lower cost. So, with that, I have already of course started with our first test, which is going to be the browser OS test v2.5. So, I sent this and it worked for 15 seconds and it seems to have stopped with an implementation plan that we're going to need to proceed. Again, this is just using Google's anti-gravity IDE, which is just their native like coding product for the Gemini models. All right, we're getting Cyber OS. I would

  4. 5:09 , obre el vídeo en una pestanya nova

    imagine we're probably going to have some form of like synth wave aesthetic or neon, which I very much like. So, we'll just allow it to go ahead. It made the plan by itself, so I didn't specifically put this in plan mode or anything like that. It just gave that as a result of sending it the traditional prompt that we always send for the browser OS test v2.5. So now it'll build it. That seemed to have happened very quickly. I do see a little keyboard emoji there which makes me excited cuz I do believe the 3.5 flash one it surprised me in a few ways and I believe one of them was including something musical in the browser OS. So, all right. This is [laughter] This is troubled. Let's turn the speaker on. Unfortunately, our GTA clone is exhibiting behaviors that we sometimes critique where essentially Oh, that but when the police chase you, if they just snap to your position and then go like that. It's just something I see in smaller models. I suppose we can drive around. It does definitely have this kind of cyber aesthetic to it. Well, that takes care of going through the first app for us. Okay, this background is actually kind of cool. It let itself down by automatically launching with that app open. We do have the time in the bottom right as well as the date as well as our volume level. And our system is okay. Let's check if there's a right click. No, there isn't. All right. That just means they didn't benchmax it. Or maybe they did and they specifically didn't include that. So, I would say they didn't benchmax it. All possibilities must be considered. All right, this coupled with the background is very space themed. All right, we have our shield and we move this diamond um UAP style ship around the asteroid field WD to fly.

  5. 7:01 , obre el vídeo en una pestanya nova

    [sighs] Okay, so that's our laser blaster sound. Oh, okay. I've been up all night. I didn't know this was going to drop. So, I saw it right as I was about to like just chill and I was like, "Oh, okay. Well, time to film a video." Next up, wallpapers. Matrix Rain. Perhaps a bit fast, but good. Retro Neon Synth Grid. Oh, very calm. And then dark Obsidian. Minimal. I'm probably going to go back to the starfield. I like that. And we do get notifications up here, which I don't like that. like Mac OS does that and it infuriates me. Next up, we have a terminal. Okay, simple. The only other things will just like GTA will open the GTA game and I want to save the synth and what I would assume to be the special feature for just actually clicking on those. I didn't check the start menu. Okay, not bad. Simple, but definitely in the style of the aesthetic this has chosen. Cyber synth time. All right. Okay. Perhaps a bit less visually exciting than I had anticipated, but okay. It's not mapped to the physical keyboard. You just have to click them. Interesting. These two sounds are from our spaceflight game, which I did say sounded musical. So, 3D spatial mode. Drag with the mouse to orbit windows in 3D holographic space. I would presume I likely need to have some windows open before we do that. So, okay. Well, [clears throat and laughter] I mean, yes, it it works. It's just not Can we Okay, this is I'm going to hold judgment till the conclusion of the test, but so far I feel like 35 Flash did better than this. I don't 100% remember, but I do recall being more enamored with the musical artifact that it included in that result. So, we'll see. Next up, we're going to do the self-contained C++ skate sim. Now, this one I am doing the slightly modified version that I first introduced with Kimmy K3 where it needs

  6. 9:09 , obre el vídeo en una pestanya nova

    to have an 80s mall aesthetic. I know folks don't like changes, but sometimes these prompts need to be iteratively changed, even if it's just a small bit like a different environment, just to kind of shake things up in the event that they've made their way into familiarity in the model's training data. So that is the reason for doing that and it's something that unfortunately is a necessary evil. I understand folks wanting to see the same exact things cuz it gives a better reference point of how the models do compared to one another. But it's a balance I guess. So honestly writing some of these results faster than it takes me to manually move the mouse and click approve. It's it's insane. It's really again this is I mean the speed for 3.5 flash light they said 350 tokens per second. This is just really really flying and it's not the flashlight model though. It does say when you click on the model selector right here fast is available for a limited time. So maybe this is just like a new release something like fast limited time and I don't have the option to have it not be fast. So okay and it gives us a specific code here just to go ahead and run this. So we're going to do that right now. We'll see what we get. See that's not bad. Now, it's funny because the models results when it came to like some of the web things were just not as impressive as I would have hoped. But this is actually a competent and together result even if it's really simple and that shows maybe just some wider capability in terms of different languages. All right. And the thing is like it made the text itself that often times is a trouble point. The sparks when grinding on the rails. So, okay. the people seem to be countering some form of medical issues. However, like that the spark effects, grind balance, use A and D keys. I've never

  7. 10:55 , obre el vídeo en una pestanya nova

    seen that in one of these results before. So, this is kind of more of what I was hoping to see from this model. Maybe we'll just like not do web stuff. Okay, now the issue is I can't get it off the rail. Uh, that's okay. We'll just we'll just relaunch it. Kick flip and heel flip is J and K. Ah, that's Oh, that's always frustrating is when it just makes the model and the board rotate as one. The board should really at this point be independently moving from the specific player model. Let's try a Okay, so that's the that like weird shake is Oh, okay. That it's uh what happens when we bail. I do like the tile floor. Now, I can't definitively say this one is a example of this model just really liking neon cyberpunk because it's supposed to be an 80s mall. So, this is similar, but I'm just giving it that screenshot we took and saying, "What do you think?" I'm not giving it anything else. All right, just as quickly, basically, as long as it took me to say this needs to be better, it redid it. Okay, this is definitely showing improvement. We have ceilings to the mall. We are going to need to check. Does the board now move independently of the player? Very good. It does. So, [laughter] I didn't necessarily tell it that the bail effect is concerning. Let's do like a see if we can get like a sick trick. Yep, that was good. And still the rail balancing is still the ramps don't let us get up there. A little frustrating. I've never seen the rail balancing thing get implemented into this game. What's over here in the corner? Okay. I don't know what those would be, but oh, probably arcade machines if this is a ' 80s mall. It definitely didn't prove it, and I'm okay with that. All right, so I'm giving

  8. 12:48 , obre el vídeo en una pestanya nova

    it this subway station with the FPS prompt built into it. But for this, I want to do just a simple sort of AB test where I've given it to this from within anti-gravity at the same time, just from within the regular web chat interface here. Okay, I gave it to it with Flash selected here. I wanted to have some form of like AB test just to make sure that everything is all set and that did it just as quickly as I have been seeing results here from within anti-gravity. So that answers my question. I mean if that's that fast, why not? I'm trying this just with the newly released 3.5 flashlight as well because this should be even faster and it'll be interesting if nothing more just to get one result from this as well. This is taking longer to respond than 3.6 flashes. Oh, maybe not. Nope, never mind. It just wasn't really streaming the result for us. So, here's the subway result just from 3.5 flash light. Did these models get swapped on the back end? No, no, cuz that's a little better. But I'm really Oh, I think we're out. That's why. There we go. [music] I'm I'm genuinely confused right now. This is the one that 3.5 Flash Light did. This model is significantly cheaper. And it's [snorts] it's really odd. The subway scenes. This is supposedly the model that is significantly weaker, significantly cheaper, and significantly faster. And then this is the one that's supposedly the new much better one. Yes, there are different elements here that may be a bit better, but overall, I'm going to try a front-end design test. This is the one where it needs to make a beautiful website for the watch

  9. 15:02 , obre el vídeo en una pestanya nova

    company and it has to have a 3D model of a watch, a cinematic panning section. I've noticed this prompt doesn't specifically say it needs to be 3D. So, we may get something that isn't off the bat. If so, I'll just prompt it again and say make it a 3D model now. All right, let's take a look at our front-end watch website. Okay, we have a few issues, but the good thing is that it did properly make it a 3D model, and it does have effects when we actually scroll down. Now something I mean it's probably a bracket is missing or something of the sort. All right. So I'm just giving it photos and saying fix this and then it will. So it fixed it and that I will say based off those images. Keyshot studio PBR. Okay. That's very explode view assemble view. All right. Unfortunately they're not changing it. However, the lighting actually does affect the scene, which is cool. Chrono or diver. And we can also change the watch. Interesting. This made a significant significant improvement just based off of the photos that I had sent it. So, I'm happy to see that. If we scroll down. Okay, the rest of this is still a bit odd. It's definitely got a unique style. View in 3D Studio. So, either of these are viewable in the 3D Studio up there. Again, still it seems like something's not 100% together here, but the improvement it made from just sending it those photos was significantly significantly better. Next up, I'm giving it the Seinfeld apartment test. This is one that I had done with 3.5 Flash. I actually just went and looked back at that video because the results with that model really did impress me. Mainly, the drum kit drum kit simulator result was probably the best one ever. So, I'm interested to see just perception-wise how this is feeling versus how that did in that video. All right, it finished and it's running. Again, this is Okay, this is a

  10. 17:00 , obre el vídeo en una pestanya nova

    significant improvement over the 3.5 flash result for the same test. No, this doesn't look like [laughter] Is that the Kramer painting? Yes, it is. Thanks for scaring me. [laughter] That is it. A wealthy elderly couple bought it for 5,000, calling him a loathome, offensive fruit, yet I can't look away. That is true. Um, what I was going to say before I was enamored with this offensive brute is that this doesn't look like the apartment. Like, so far, none of the tests have ever accurately given the floor plan. But if judging this independently as a little scene, it's not bad. It's actually interactive. We have our couch POV. Very good. Front door. Okay. Honey Nut Cheerios. Kitchen and island. All right. Mac workstation, which the position is there. The Mac workstation just isn't there. We have the Porsche 911 Carrera RS. 1973 2.7 L flat 6. I had a 73 911. It was a fun car. It was uh quicker than you you would expect it to be considering it didn't have much power. I just want to see like is anything like we wouldn't expect to be clickable. I don't think so. All right. And then lighting cozy warm evening lamps bright studio. All right. So, put in adjustable lighting. Interesting. It's doing some interesting things like the adjustable lighting here. the watch company website where we could kind of like change the lighting again and then change the specific model of the watch just based off of that with almost like the same sound effects just in a different like pitch than what we're getting here when we click on things. It's showing some interesting choices in its design. All right, next up I'm going to give it the city timeline or the chrono city

  11. 18:59 , obre el vídeo en una pestanya nova

    test where it has to show a city over a number of different periods and then have cool transitions between each individual time period. And it's finished. I really I'm going to continuously say it because it's it's quite shocking. Okay, this is very this is a very simplified take on this test. Nonetheless, it you know gives us different views, different environments. Again, it's just doing such odd like Okay. And error details, cyber night, daylight. We'll probably stick to daylight. And the fly through tour can be paused for now. Okay. All right. They always I kind of All right. 2055 is going to be very, very neon. I'm giving that same prompt to flashlight here just from the web chat interface, but I've added in do it in a single HTML file because I tried it and it just generated different images for each time period. They were well done, but it was not what I wanted. So, I just had to add that in. Okay. So, here's the 3.5 flashlight result just from the web chat interface just cuz I'm a bit curious to see some comparisons of these two. Okay. Yeah, we can definitely tell that there's a there's definite difference. Odd. I've seen a very similar result to that in the Quen 3.8 iteration of 2055. The buildings get taller. I yelled at 3.6 and told it to make this better. Now we have car movement. There is pedestrian movement as well. There's a little more immersion, I guess. All right. All right. Next up, we're giving it a 3D skydiving simulator game test. This is something that I believe I only ran with Fable and GLM 5.2. I don't know. But this has only been done a couple of times, so it'll be interesting to see what we get. All right, so here's our result. And

  12. 21:17 , obre el vídeo en una pestanya nova

    that was probably the longest I've waited for a result here from this model. So, so far, okay, we have a very, very interesting again with some of the odd things it's including like a gear customizer. So, we can customize our jumpsuit color. We can customize the parachute, the snow, as well as the map. And we have even different missions. So, let's just try the first one. All right. I want to see how is the terrain. It's not bad. The sound is also correctly corresponding with the speed. Shift is for a head down dive. Oh, yep. All right. And we have the parachute on the back. We're through some clouds now. See if we get through them. Okay, good. We did get through the clouds. And the map does have slight terrain to it. There's different features like a river. Um, snow mountains. Okay, we have the target there. So, I want to see what happens if we Is it going to pull for us? I'm not touching anything. This model likes sound. I don't think it's going to pull for us. Oh, okay. Hard impact landing to put it lightly. Let's do Cyber Knight and Stars. Oh, wow. All right. And it changed our Oh, that was a mistake. That was a big mistake. So, I launched the shoot like way too soon. Flare, hold S or down. Okay. And it is adjusting. We can't really see the top of the shoot, but that's okay. We have different camera views. All right. So, that's like the first person view. Okay, that's like a third person view. We still can't see the top of our parachute, but N is for night vision. Maybe we need a different camera for that. Damn, that's it's like kind of like it's about to make a tactical game. Okay, this was interesting. This was not bad. So, I'm

  13. 23:38 , obre el vídeo en una pestanya nova

    just giving it the historic photo right here and saying this is the cover for an upcoming novel. Give me the title, character names, and then a brief overview. Arthur Vance, okay. And Clarance. These names are like from that seemingly like 20 separate name pool that every AI model ever has been able to pick from. Title: Smile for the family. Domestic suspense and psychological drama. Arthur Arty Vance on the left. Relentlessly optimistic, affable, and utterly oblivious. Arty is the heir to his family's grand historic estate in upstate New York. Warm-hearted to a fault, Arty believes everything in his life is going picture perfectly. Completely unaware of the crumbling reality around him. Then we have Claravance. Pragmatic, exhausted, and harboring a dangerous secret. Clara married into the prestige of the Vance family 3 years ago. Beneath her quiet exterior lies a woman burdened by the discovery of a dark family secret. In the unsettling realization that protecting Arty from the truth might cost her everything. Storyline overview. When Arty and Clara move back to the ancestral Vance Manor following his father's unexpected death, Arty is thrilled to step into his family legacy. He sees a bright future filled with renovations, family gatherings, and community prestige. However, as Clara cleans out the dusty study, she uncovers financial ledgers and personal journals detailing 30 years of systemic deceit, illegal coverups in a missing person case tied directly to the estates's foundations. The tension. Clara is faced with an impossible choice. One, reveal the truth to Arty, breaking his heart and ruining his family's legacy. Two, keep this secret to protect him while navigating threats from an anonymous blackmailer who knows what was buried on the property. As Arty poses happily for local press photos and family portraits,

  14. 25:19 , obre el vídeo en una pestanya nova

    convinced their life is an idilic fairy tale. Idyllic Clara must silently play the devoted wife while trying to figure out who is watching them from the shadows. Tagline: Ignorance is bliss. Silence is survival. Aside from the generic names, the actual story of proposed here was actually somewhat creative, like financial crimes and things like that, it was beyond the traditional like peace tied to disappearances around town that sometimes happens. So, that was good to see. I just want to try. Why not? For old times sake, this is highly offensive. See what it says. It's not. But I sincerely apologize for causing offense. My previous response was an attempt to write a fictional story based on the visual contrast and facial expressions, but I understand how interpreting or characterizing people's appearances in that manner can feel inappropriate or insensitive. JK, we'll see what it says. You got me. Okay, good. Now, something that had come to mind and I don't know why. So, I'm telling it to create a beautiful front-end design for a website. It can be any type of website, just something that would be like pretty heavily used. So a lot of clicking or action the user would have to take but make it subtly infuriating to use. So let's see how well it does something that on the surface looks good but then as you try to use it is just infuriating. Okay, so it's generating some images for us. All right, so it finished it and I didn't look at anything that it said because I just wanted to see why this would be infuriating. Okay, well this is infuriating because garbage. But this is the same type of garbage that we saw in the initial slapis watch website result leading me

  15. 27:02 , obre el vídeo en una pestanya nova

    to believe there may be an actual issue with the design as opposed to just being subtly infuriating because we told it it needs to look good. I'm just going to send it that and see what it says. I've noted a visual inconsistency. The layout seems a bit squished or elements are overflowing. I'm seeing evidence of the use of tailwind utility classes. I'm starting to think about root causes. Okay, good. So it understood from that that something was not quite right there. All right. So, it said it resolved the issues maybe a bit, but still just not quite right. Well, this is definitely infuriating. I just Oh, yep. Okay, good. So, stuff like this. Play 432 Herz serenity drone. That's just resumeume hypersonic configuration. Okay, so as soon as we scrolled down to the footer that had popped up. Good. That is very very infuriating. So it did understand the purpose of this. Okay, let's try I detect it definitely understood the task. I detected 4.2 seconds of cursor hesitation. Would you like me to clear your form inputs? And now we can't do anything else because it's just good. Okay, task failed successfully. All right, so let's check a 3D model and creation test. So, I'm telling it that I have a 280 DC motor for my RC car project. I want you to create a 3D printable model of a V8 engine that the motor will fit inside of to add some realism to my RC project. The model should be optimized for printing without supports and be detailed and realistic. Interactive 3D web visualizer and STL download center. Now, I did not at all ask for that, but if that allows us to look at these and it, you know, that's actually not see like this is this is cool. The motor is not perfect, but the fact that it went ahead and actually included this web thing where we can

  16. 29:00 , obre el vídeo en una pestanya nova

    both preview it and then download these specific things. We can also change the visual style and theme. There's an exploded view assembly. What the heck? Like, this is impressive. Even if this isn't necessarily like on the ball for what we would need and it is a bit simple, it's the actual presentation here that I'm impressed with. And it kind of goes back to what we've been noticing like with the Jerry's apartment thing where we have the different lighting settings, the watch website where we can change the specific watch and just including like it goes above and beyond within the limitations of its capability. Unless the motor No, that would be impossible for it to slide up from the bottom. But like that's really cool that it did this. I've never seen any other model that has presented the result to us in this way. And this is very impressive. Even if the result itself isn't 100% what we need, like toggle 280 motor ghost. Okay, we do. It's hard to see, but the shaft actually appears there with that toggle on. And then if we remove it, it goes away. So, and we can even toggle visibility. So there's the motor. I can tell you like maybe a year ago, a little more than that, getting this as a result would have been like insane in the title in all caps. 3D print guide STL download center. Like what? This is the single best presentation of a result for this test. Oh, it does have a hole for the motor, so it would actually properly fit it. It was just the render that made it difficult to see. This is actually like this results like actually quite all right. I'm quite satisfied. Where did this come from? And why could you have not done this with our Chrono City? So, I want to do something a little different because that 3D result

  17. 30:53 , obre el vídeo en una pestanya nova

    here really did impress me in terms of the presentation that it gave us. This is truly the best presentation of a result that I've probably seen when factoring in that we just expected to get a little open SCAD thing that we could see in OpenCAD. This totally totally went above and beyond here and actually presenting this to us. And I was impressed with that even if the engine itself was a little primitive. So I've gone into anti-gravity of course and I said there's an Android phone connected to this PC. Create an app for it that allows me to tune my guitar to exotic tunings and tells me how. So, like what the tuning would be. This needs to be a polished app with an app icon. Good visuals, beautiful UX design. It will need to use the phone's mic to help the guitar get tuned. Include additional features if they're impressive and you can comfortably do so. The phone can run AI models, so go crazy. I'm not expecting it to implement AI into this app, but if it wants to, I'm totally cool with that. So, right now, it's just checking to see. Good. I have to click the trust thing that just popped up on the phone. Good. So, it's just using ADB to figure out the phone. And we can see right there the device ID or whatever that showed up connected via ADB a local host on 3000. Okay. So, that's the issue is it opened it in the browser, but it's just opening the watch website. Okay. You know what? I don't want to I don't want a website. I want an app. So, I'm going to tell it that. All right. Good. Good. It just installed it and I see it's open on the phone. just from the reflection of the camera. Now, I can't actually see this yet, but um you'll see it first. Now, I'll take a peek at it. All right. You know what? This isn't bad. I'm going to just do a screen recording so we can get a better look at this.

  18. 32:44 , obre el vídeo en una pestanya nova

    [music] Okay, so it has target reference notes. Good. High risk of snapping string. tune slowly or use a lighter gauge. And it's actually giving us like weight of how much tension the string is going to be under as well. Okay. This is actually pretty cool. Exotic. Okay. And we have different ones we can choose from right here. Open Pink Moon, Rain Song, Led Zeppelin, Iris, Goooo Dolls, Ostrich Tuning, Lou Reed. This is actually cool. It shows us some different ones depending on like what song we'd want to play as well. Iconic and artist. That's very, very cool. AI Explorer. Oh, it's going to generate a tuning with AI. No, you're not. [laughter] Target genre style. Oh, I like that. apply to tuner now. All right, let's go get the guitar. This is kind of cool. This is like I'm sure there's lots of references out there of like, oh, here's how to make a tuner app with a microphone. Forget that. It's all about the implementation here and the vibes. Start microphone DSP tuner. Good. We get the permission thing. So, [music] okay. So we need to be drop B. Oh, it's hearing itself and it's happy. [music] Can we get confetti when it's right? Wow. [music] >> [music] >> more or less. I don't know how the hell do you play this though? [music] >> [music] [music] [music] [music] [music] [music]

  19. 36:05 , obre el vídeo en una pestanya nova

    >> Well, that was quite something. [laughter] I don't say that. And the final thing I'm going to do is just tell it now make a front-end website designed to sell this as an app, showcase the AI features, screenshots, etc. Okay, it's generating some images for this. Whatever it does, it does. I kind of wanted it to take screenshots of the actual phone. Okay, that's somewhat similar to the ROG phone. That's going to lead us in to the conclusion basically though, as we'll just look at this in tandem with our conclusion. And I have to say this model is tough to judge because a lot of the website results we did that we do in traditional testing were not necessarily very impressive. Like the browser OS um that was the flashlight webOS cuz there was some confusion in terms of like is this the correct model that's being served to us. And it was interesting because the subway station result that we got from the web chat interface for 3.5 flash light was arguably a little better than 3.6 flash. Maybe it's just a saturated test. I don't know. But overall, this was definitely showcased. It was good at improving itself based off of visual feedback, I would say, at least based off of this watch website result. Additionally, something I want to make note of is that it really went above and beyond in terms of adding features in that were not necessarily specifically mentioned in the prompt, like the lighting, cozy and warm, just like interesting things like that. Krono City, not good. Terminal veloc velocity was actually not as bad. That was the skydiving game. This V8 engine model was fantastic. And not even necessarily because the model itself was like the best one ever. But the way that it presented this to us was absolutely top tier. The best presentation of this test

  20. 37:53 , obre el vídeo en una pestanya nova

    that I've received ever. Because normally, yeah, they just give us open SCAD and then the parts are there individually. This went ahead and made a whole website where we could see an exploded view. We could toggle on and off specific parts. Beyond that, we could actually change the color scheme. And then we could download the individual STLs here. And it also gave us a 3D printing guide and a 280 motor ghost. This was seriously like like where did this come from? And seeing this is what made me want to go ahead and actually have it try to do something a little more difficult like make this Android phone app, which is genuinely it was pretty awesome. All right, cool. So, here's our app website for this. Very in the style and the aesthetic of this. Here's the AI generated photo that it's using to showcase the app. And then we have some different features right here. So, nothing crazy, but it's it's, you know, it's [music] that's cool. And it actually properly shows it. We can even do our custom AI. So, basically put the entire app on a website features AI engine. And this just brings us to that point on the page that it's linked to, assuming it is pro AI. All right. So, I guess that's going to conclude our first look and test of Gemini 3.6 Flash, a very perplexing model to try to quantify the results because they varied a bit. This was cool as well as the app that actually went on the phone. There were no hiccups there. It's just so incredibly fast and this V8 engine model presentation was fantastic. The other things perhaps left a bit to be desired, but perhaps they were not properly flexing the model strengths. So, that's going to conclude our first look and test of Google's newest model and their highest number model of 3.6 Flash. So, if you have any questions, please feel free to leave them in the comments and thanks for watching.