Gemini 4 Still Isn't What I Need (Yet) Transcript, AI Summary & Key Points

00:00 The day is finally here. Despite my complaints, Google is still pretty chill with me and gave me early access to Gemini 4 Carbon. I've been using this model a ton. And I'm going to be fully honest. Promise it'll make sense in a second. I was not that impressed. And my plan for this video was to be not particularly nice to it because when I used Argan, it was not great.

00:17 And I did not see as big of an improvement with this release as I was hoping. Especially considering the fact that when it was announced 2 weeks ago, it was the best model in the world cuz it was cut off there with Gemini 4 carbon getting a 60, Opus 55 at a 59 in artificial analysis. The problem was that Fable just dropped, which is a huge bump over even the benchmaxing nature of Gemini 4 carbon.

00:42 It does look great if you look at the Gemini models of recent by looking at it on something like here in slopolytics. But the only reason these numbers look so good is because I turned off all the other models. Like, yeah, there are better options for a lot of things in a lot of places. Or so I thought, because I continued to dig deep into this model and its capabilities, and I found things that nobody told me about that were genuinely kind of surprising.

01:04 And I can't believe I'm saying this. I think I'm going to land some changes tonight using this model. Yeah, really. I'm shocked, too. I got a lot of explaining to do with this one because this release is certainly not what I expected at all. And I promise you're going to like the ending on this one. It's a it's a chaotic journey to say the least. And before any great journey, one must equip themselves properly with useful things like today's sponsor.

01:28 Let's play a game. What do these companies have in common? Companies like OpenAI, Anthropic, FAL, Parallel, Cursor, Clay, Whisperflow, Firework, Base 10, Replet, Gamma, Scribe, Laravel, Perplexity, Hex, Factory, Turbo Puffer, Planet Scale, Granola, Bolt, Post Hog, Exa, Warp, Luma. Sorry, I'm struggling to say all of the companies without running out of air because Work OS has won the industry.

01:54 And there's a reason why. Getting O right might seem easy, but getting off wrong is insanely expensive. And when a company like Microsoft comes to you and says, "Hey, we want to use your product. Do you have all our weird SSO things we need to sign up?" You better be ready to do it. And don't think your agent can because good luck getting an agent to set up SAML, Octa, and all these other annoying providers directly properly.

02:14 Now that every O provider is easy to set up, thanks AI. It's important to have one that actually solves the problems that matter to your business. And if you don't recognize every single one of these logos, you're in for a rough time when an enterprise comes to work with you guys. Set your business up to work with every other business at swive.link/workos.

02:34 So, first and foremost, let's see what Google has to say. Carbon is now available. Remember, they announced it two weeks ago. I'm sorry I couldn't talk about it then. Was waiting until I got the thumbs up to be allowed to. I've been using it for a good bit of this time and was not super duper impressed initially. They call out at the start here that they're making Gemini 4 carbon available to developers, enterprises, scientists, and consumers.

02:51 Carbon is our strongest model with frontier performance. Frontier performance a bold statement especially when it falls within cyber security capabilities and long horizon agentic workflows across complex knowledge work software engineering and scientific reasoning. Huh. This is an interesting sentence that I did not properly understand the implications of until about an hour before I was ready to film.

03:12 Yeah. They called out that two weeks ago they announced the model and released it to a set of cyber security defenders via the Fairwin program. They also gave early access soon after that which is how I got a hold of it. And I will say I ignored the cyber security aspect for far too long. It ended up being the best score ever on cursor bench 4, which is kind of wild and I'll be real makes me question this benchmark because in my benches it did not perform anywhere near as well.

03:37 My recent code benches have been a strange angle. They are focused on analysis of code. Whether it is a code base in order to better understand it or if it is poll requests that it has to review and find potential issues in. This model did not impress me much here. If we look at how it performed in my Rust codebase doing an architecture review of the existing code, it performed better than Argan.

03:57 It's a much lower percentage of lies where Argan made 36 claims, 24 of which were false. Carbon made 79 claims, so it said way more. It had a lot more opinions and it also got there faster, but it still had 23 false claims, which is insane because a lot of the other models have between zero and one, and a handful have three to four if they're a little too uh loosey goosey, so to speak.

04:19 But this is just one of a few. Here we see in my T3 code analysis, Gemini 4 argan got a 47 of 92 for fails. That is over half of its claims were false. In Carbon was only 23 of 86. Across all of my evals, the percentage of false claims dropped from 36% with argan to 18%. And I find this particularly fascinating because I was told by some deep mind people that they were very aware of the hallucinations that I was seeing and they saw about a 2x decrease in hallucination rates from argan to carbon.

04:56 And there it is. It's an exactly 50% decrease. I'm a little frustrated with this one because a lot of the like official hallucination benches don't measure this same failure that I'm seeing, but I'm seeing it pretty damn clearly here. If I'm being real, it's like, yep, there it is. It dropped in half, but it's still way worse than anything other than the Muse line.

05:17 Actually, it's still a lot worse than the Muse line. Even Wonder is still pulling a 14.5 versus the 18 to 36. Obviously, the non-ful version of Wonder is garbage, as I've mentioned many times, so take that as you will. It is genuinely obnoxious how slow this model is. The reasoning effort levels don't really change how slow it is either. They're all obnoxious cuz that roundtrip time is garbage.

05:37 So on low the average task time was 22 minutes in artificial analysis and on high it was 27 versus just for a comparable range something like opus the low was only 3 minutes and 1 second and the max was 25 minutes. We switched to linear instead of log you can see how much worse this is. It is consistently the slowest model. It is slower than Max on literally everything other than Sonic cuz it's stupid in Fable 52 cuz Max 52 makes no sense at all.

06:08 Yeah. So, everybody's saying, "Well, it's a fast model." Shut the up. You have no idea what you're talking about. Like, I'm sorry if you think TPS means speed. Look at the numbers. This model is slow and it sucks to use cuz it's slow. Ah. Anyways, I told you there's things that are good about this model, but I have a little more crashing out I got to do first.

06:23 They finally introduced their equivalent of workflows with SLS swarm. a little cringe to like push this as a new feature. I also noticed that they pushed this to their brand partners for the announcement like the Verge. And here's where things start to get Gemini. The price is $2 per million input tokens and $10 per million output tokens. But this is their introductory price.

06:45 What is an introductory price, might you ask? It's a Googlem. The price is $2 per mill in and $10 per mill out with a 10 cent context cash read cost until December 1st at which point it becomes $4 per mill in and 20 per mill out and 20 cents per mill cash read doubling in price. Normally intelligence is supposed to get cheaper over time but Google truly innovates on shitty things.

07:14 I think they do this because they are trying really hard to get people to adopt the new model and they're hoping they can get some people stuck on it and then they will charge the higher price. Previously, my assumption was that they did this in order to disincentivize people from sticking with an older model so they could focus their provisioning better.

07:31 But they would give like two to three plus month timelines for these discounts. Now it's under a month, which is uh yeah, that's something. I do not think this model will be worth either of those prices, but the 4 in and 20 out price is Opus price for a model that much much more so is unpleasant to use and the subscription subsidies are significantly worse as well.

07:50 One last thing on cost. Thought this is worth calling out because I did some experiments on it. The Gemini APIs don't do explicit callouts for caching. The OpenAI ones didn't used to, but now you have to like tell the API, I want to write this to cache when you're done. Claude has had it that way forever and you can even choose how long the TTLS are on the cache, like how long will this be cached for.

08:10 Gemini doesn't. Gemini APIs cache when they feel like it, and they don't do a great job at it. I found that in a 5-second window, there was a 20% chance of a cash miss. In a 30-cond window, there was a 34% chance. In 120 second window, it was even more. This results in as much as a 50% increase in requests if they are waiting for long enough that they hit those worst cache read rates.

08:34 Which means that already, as I've said many a time, the token cost does not matter. The cost per task is what matters, especially when you consider how many tokens are being used in a given task. Cash rate efficiency matters a lot too. And normally we don't have to talk about this because the cash rate effectiveness is 100% because the model hits the cache when you write the cache.

08:52 With Gemini, everything is special. So, you cannot trust the number on the tin because it will double in just a few weeks, because the cash rate does not work, and because it burns tokens like nothing I've ever seen. So, if you're picking this model because it's fast and cheap, you're wrong and wronger. So, why the hell do I like it? Well, we'll get there in a bit cuz we got to talk about office work.

09:13 This is the furthest opposite of fish slop. I promise we'll do fish slop, don't worry. As well as some other really fun things. But, uh, yeah, it has higher precision knowledge work, artifact generation, and data analysis. Cool. They say that existing benches don't adequately capture how agents reason natively over dense files with charts, images, blueprints, schematics, and tables.

09:29 So, they partner with artificial analysis to make a new bench, the multimodel knowledge agent. It's funny, they help make a bench, and they're only number two on said bench. And now for what we're all here for, the improvements in agent coding, full stack development, and 3D as well as UI design. I will say up front, not super impressed with UI. Just checked.

09:44 And of course, our boy D has got Witchai updated so we can look at the front ends quick. Here are some of the ones it made using the anthropic design skill, which I found it handles decently well. I don't love the layouts here if I'm being real. Let's turn that off and see what it does by default. Oh, actually, ah, his native design is less bad than with the skill.

10:07 But it still really likes this particular template of like the pill section here, the two line centered text, the logo, the begin, the four links that don't really matter in the middle, the screenshot of the app cut off. Like these are the same just with different colors. And that's the same as well. It looks like only two of these are different. Yeah.

10:28 So it does get very templaty. Take that as you will. They claim that it can perform exceptionally well in engineering tasks that require multi- aent collaboration like multi package migrations, cross-system root cause investigations, and complex architectural redesigns. It's also good at web development and 3D game development stuff. Allegedly, they got some quotes from our friends over at Open Code, who as much as I love them, I am very skeptical of their love of the Gemini models.

10:54 They also really liked Argan, which was not a good model for real world code work. I had to revert every single thing I merged from it and I didn't really merge a lot from it. So that says quite a bit blah. We struggled to get our app looking consistent with an existing UI framework. We gave Carbon a list of requirements before closing our day. Overnight, Carbon built a new Rust UI library from scratch and rewrote the Open Code desktop app with working demos.

11:15 It's a brainy model and it's a preferred planning model for our hardest tasks. That is a choice because when you give this model long things, it hallucinates more. I just don't think this is a fair statement to make and I would be very curious how they chose to do things because that others said it was good at difficult 3D tasks like cloth simulation.

11:33 It treats software contracts with respect. Yeah. Yeah. You get the idea. People seem to like it quite a bit. They show some fun demos here. Fancy 3D things. It made underwater Atlantis scene in 3JS. Me. Here is a random concert hall that it mocked up. Looks fine. Just this all feels like six plus months ago if I'm being real. Oh, they snuck in a pelican on a bicycle.

12:03 Of course, shout out to Simon. Beautiful. And here we get to the section that I kind of slept on. Frontier cyber security capabilities, which we'll get to right after Fishlop. Here is the first version that it made. Not the most impressive I've seen. It does things different. Like again, it's one of those where it's a flavor that is different from the other models, which is cool in its own way, but like compared to what we just saw with Fable, this is embarrassing.

12:35 I wanted it to have a better chance, though, cuz I I know Google's models allegedly are good at 3D stuff. So, I did. I told it to make it beautiful, and it certainly made uh something here. Might just be Firefox. Nope. That is hilarious. It made the water look good, but it uh broke everything else. Just told T3 Code to find the thread I made this in because I did that days ago and bump it and tell it to go fix that with Gemini 4 Carbon High just so we see what it can do and I'm not patching its mistakes with other

13:09 models. I want to see its capabilities. I wanted to get you guys data on how much usage you can get from the Gemini models with the subscription because I have the $200 tier. First and foremost, I need to crash out a tiny bit at the anti-gravity team because I've told you guys many times, do not show my email all over the place in your UI and no matter what I do, they spam it everywhere.

13:28 So, thanks to my editor for making sure my email here does not become visible. Thanks to Anti-Gravity for continuing to do this. Very annoying. One of the many reasons I will never open Antiggravity in my content unless I'm filming offline. Yeah, so much it. Anyways, you'll notice that my limits are at 100%. That's because I am on carbon with early access, which I don't think goes against my usage, which means it's very hard for me to get real numbers here.

13:51 I will follow up with those numbers when I am able to get them. Just know that historically the subsidization that you get with the Gemini models on the $200 plans is not particularly great, especially when compared to what you get from Anthropic with the $200 plan there. Okay, back to Frontier cyber security capabilities. This is one of the things I slept on a bit.

14:08 I will show some fun examples later, but for now, let's just take them to the for their word, where they say, "Gemini 4 carbon for cyber delivers our strongest cyber security capabilities to date, giving security teams endto-end intelligence that they need to protect critical software." When assessing through the Fairwind program, Carbon leads on key benches across findings like SEC Bench Pro, Exploit Gym, and more.

14:30 Early testers inside of Google are already seeing strong results with Carbon for cyber in practice. Project Fortify. They went through and found a bunch of things in open source projects. It's a 25% relative improvement to what Argan was seeing. Apparently, they were able to find two bugs in popular open source projects that are similar in severity to PHP to shell, which is terrifying.

14:48 It's also better at scientific and mathematical reasoning. Cool. Awesome. It does enzyme assay development. Go to another channel for these things. I don't know what I'm talking about here. It's more aligned. It triggers high severity alerts in RL much less often. It cheats on eval less often. The other ones were very aggressive with it. Argan would cheat as much as 4% of the time.

15:08 Was a 9x decrease in cheating. Across prompt injection benches, they got a 0% which is very, very good to see. And yeah, now you can use it. But should you use it? Well, as I was saying before, if the task is stuff like code review and codebase analysis, absolutely not. I've seen this model just make up a lot. It makes up column names for tables in your DB.

15:31 It makes up line numbers for things that it doesn't properly read or understand. It makes up things about your architecture that just aren't true. It It's hard to trust. It's really hard to trust, which is why its strengths are even more annoying. As you can guess, I've been using Gemini almost exclusively through T3 Code. There's been a lot of skepticism around using your Gemini sub and other tools.

15:53 And I'll be direct, you shouldn't do it without Google giving the tool permission. Thankfully, the DeepMind team has been working with us a ton. We've been helping them a lot with their ACP bindings, which is how you can control anti-gravity within other tools because they don't want you to use your sub and your inference in something else, like you can't bring this into PI or some other harness.

16:11 But if you're using the ACP bindings to allow another tool like T3 code to control it, they allow that. Previously, these bindings were a separate binary you'd have to download that we included as an option to go install in T3 Code. They've been landing a bunch of changes on their side to make it way easier by including this in the anti-gravity CLI.

16:28 We chat with them a bunch about this. Believe it or not, only two labs have landed changes into T3 code directly. They are SpaceX who built the Grock build implementation that we use initially. We've worked with them a bunch since. And the team over at DeepMind that has been working hard to improve the ACP bindings in T3 code in order to make it so that we have one of the best possible Gemini experiences.

16:50 Some of the hallucinations I was getting from Gemini were because of issues in the old ACP setup, which was obnoxious. It wasn't sending the right data to the model from the ACP bindings, which meant that it didn't get info like line numbers like it would in the official Gemini ACP AGY CLI. Lots of words there. It's just the Agy CLI. They have too many different things.

17:11 I really wish it was still open source like the Gemini CLI was, which is a different thing. Regardless, I don't want to crash out too bad because these guys have been awesome. Every time I have some weird Gemini issue, the number of genuine hardworking people at DeepMind who reach out and do everything they can to fix it is unbelievable. If I was to rank the labs by likelihood they will give it an honest effort to fix a thing I complain about, I think I would put them in first place, which is crazy because OpenAI has

17:41 been super receptive to my feedback. And obviously now that cursor is part of SpaceX AAI, I have a lot of ins there too to get things fixed when they suck. But deep mind just does it. I was upset about the lack of clarity in the policy for using your account other places. So they fixed it the next day. I was upset that the hallucination rates I was seeing in T3 code were absurd.

17:57 So they audited the ACP, found the fixes, and set it up for us. They even went as far as adding a special T3 code identifier to give us the special treatment that certain providers like Zed had. Awesome. This is all really good stuff. Like I want to give the credit where it is due. The people of DeepMind are some of the best and I genuinely really like working with them.

18:15 Which is why I was sad this video was going to have to be a crash out because using the model sucked. And I'm going to be real. Using the model is still nowhere near good enough to justify reaching for it over Opus. The price is not good enough. It's actually more expensive in most cases. It is way slower. The bindings suck. The behavior sucks. It can write decent code in certain places in certain ways if you're willing to babysit it and put in the effort.

18:37 But then something happened today. We got an email today, an email from an independent developer that had gotten Mythos access through Enthropic security program and had used that to audit T3 code looking for security issues. He found two that were very niche. One was almost impossible to trigger in any realworld case without like really specific targeted attacks.

18:59 The other one was a little more likely for an agent to be tricked into doing. So, we decided to address them as any good maintainers would. By the time this video is live, the fixes are out and the likelihood these can be exploited anyways is still relatively low. Regardless, please make sure you update. I know I'm a bad example here, but I have my reasons.

19:15 Anyways, this report was legitimate and caught some very real things we wanted to fix, so we tried. Julius immediately opens up Fable 52, hands it the report, and says, "What do you think of this? How important is it that we fix this?" Tuisha gets rerouted to Opus 4.8, and nothing gets done because it's Opus 4.8. It wasn't even good when it came out.

19:34 And this was annoying. Thankfully, I had applied for the defensive access for using Mythos for defending our codebase. And thankfully got it. This is a thread I did with Mythos and it addressed the issues. It described the problem. It was a little more verbose than I would have liked. I asked for it the changes that it wants to do. And the change that it wanted to do was to fundamentally overhaul how we handle box linking.

20:00 If you're not familiar, one of the cool early access features in T3 Code in the labs is the ability to have one machine request access to another that's in your fleet. So, if I have a Linux machine that I'm using for a lot of work and it wants to test the changes or have another agent go run the changes out on one of my Mac minis or maybe on my Windows PC, it can request access and a little pop-up comes up in the app saying, "Hey, this thread on this computer wants to have access to this environment.

20:27 Do you want to allow that?" And you can click accept. And you can also choose what permission level it gets. If you're not using our connection layer and instead are rolling your own, which obviously you can do, it's an open source project, then it would be possible for somebody to maliciously connect you to the wrong link. So, what was Mythos's solution here?

20:45 You should have to go to the other machine, open the app, or run a specific command to get a six-digit code that you then run back to the original machine to approve. I don't know how much you know about me and the way I run my machines, but I have not looked at any of them in months. I'm not going to do that. That takes the whole value of the feature and makes it worse.

21:04 I'd rather just copy paste a prompt at that point. Horrible. But when I tried other models with this, like I tried what Open AI was doing, I tried SpaceX AI. Not really much to try there if I'm being real. I would just get reduced down to dumber models or rejected outright because of the nature of the security report here. I ended up going back and forth for a while with Mythos until I got it to agree that we can treat the trusted links from T3 code in T3 connect differently when you're connecting through our stuff

21:32 with O. We shouldn't have to worry about this because you're not going to spoof with a different URL when it's coming through our layer. I eventually convinced it to do this after a lot of back and forth and the changes have since landed and merged. Before it was done, I was getting ready to film this video and I decided, you know what? Why not? YOLO, if you will.

21:51 Let's see what carbon does. I'm curious. It described the issue correctly. It does these new line break things everywhere, and I hate them, but just makes the formatting a bit worse. Whatever. It describes the issue very well. It's a little more verbose than I would like, but I could understand it. And since I already knew the issue well, nothing here smelt too too bad.

22:07 I asked it what it thought we should do. And here is why this video is so different than what I had in mind initially. Gemini 4 carbon high said the following. The best fix is a hard separation between known machine flows and raw URL flows paired with tightening URL retention in the peer link section. You do not need heavy cryptographic redesign to kill this vulnerability cleanly.

22:30 And doing the separation properly actually fixes two extra issues that the report didn't call out. Yeah, it suggested what I wanted. I did not expect that considering every other model I had tried was either refusing to help or was mythos which you just saw. The fact that I had the best experience addressing these reports with this model is kind of insane.

22:52 Like genuinely, I did not expect this. I was already so close to done with the Mythos version. And again, I know this model is fast in the sense that like on paper it does a lot of tokens. It does everything else wrong. It uses too many tokens. It has way too much back and forth. Its connection methods are and the latency and time to first token are all garbage.

23:14 Using this model to complete real work is slow as balls. It is miserable to use. Some of this might be the bindings in T3 code. I don't care though. We've audited the hell out of it. It's still slow in anti-gravity, and I used it there, too. All of this said, I had it get a review from the other models that I was using. I had some feedback. I had it address the feedback and build this.

23:36 I stopped it because I had to come up to film, but I told it after its first pass do one quick round of council review. It got a nine out of 10 from Obus 55, a 9 out of 10 from Muse, and then a 7.5 from 61 Soul, and a seven from Terara somehow. Weird. Regardless, not bad scores. Not bad at all. I'm almost positive I could have let this go for a little bit longer and had nearly identical changes, at least in terms of the capability changes in the security of the changes from this, which I did not expect at all.

24:05 So, uh, yeah, this model might actually be a pretty good tool for incident response in security stuff. If you find yourself getting flagged by Opus, maybe give it a go. I will say there is some potential that my early access status prevents me from getting flagged as hard as others do. Maybe it has booked me under the cyber like whatever the name for their program is.

24:26 I don't know. I don't care. This is a thing I was impressed with and I wanted to share the thing I was impressed with cuz yeah, this was good. Good stuff. Google, quick shout out to Pseudo Brandino. By the way, Brandon here was the one who reported this. He isn't at some company that is encouraging him to do this type of work. I think he's actually in it right now, but he's deep enough that he applied for Mythos, got access, filed a report for us that was super easy to deal with, super actionable, clear, concise,

24:52 useful, and didn't ask for anything other than maybe to tag him in the PR. Sorry, Brandon. You're getting a lot more than that because you are way too good to just be an IT guy. This is his Twitter. If you're looking for somebody who understands their in building with agents enough that not only are they using T3 Code, but they're actually contributing useful information for us to keep T3 Code safe and secure for all our users, this is the type of person you probably want to hire.

25:13 You can hit him up from here and I'm going to follow him back because this was a very, very useful report that he filed. Generally speaking, I found it much more capable of doing real work, doing things like using the T3 Code tools to spin up other models to get feedback on the work it's doing. things like staying on task and actually answering questions and taking responses like how do you feel about this path or ready go do it.

25:34 It handles things well. Here I had a bunch of investigation I was doing on some Windows remote update issues and I asked Carbon to implement it out of curiosity and it has been going quite well so far. In fact, it's getting 8 to 9 out of 10 from all the other models reviewing. To be fair, one of these is adjuster argan. Yeah, it's really not that bad.

25:53 Would I ever pick this model as my default for coding? No. Absolutely not. It's just Opus is unbeatable still. Even after Fable 52 dropped, like Opus is such a good value. Opus is such a good experience. Opus is such a good engineer. And if you're using the $200 plan, Opus feels pretty close to unlimited for real world work. You're going to hit your limits faster with this model.

26:14 You're going to deal with the huge price increase with this model. You're going to have a slower experience. It's going to say things that are harder to understand. It's going to make mistakes that are stupid. It's going to hallucinate. You're going to have a bad time. And if you don't, I envy you because you're working on a very simple codebase. And I miss those days myself.

26:28 In my experience, even the things I slopped together with Opus are often too big for this model to actually properly understand. It's kind of crazy that a model that Google, the company with one of the biggest code bases in the world, made struggles so much in big code bases, but it does and none of the other models do. Yeah, that is what that is. I still think it is a huge bump.

26:47 It is way more pleasant to use. It stays on task better. When I first was trying it, it was impressive. Like my initial response was, "Wow, maybe Google cooked." And then I kept using it and the cracks started to become visible again. It shows it's still a Gemini model, but this is still a huge jump in the right direction. I would go as far as to say the gap from 38 Flash to Argan was smaller than from Argan to Carbon in terms of the vibes I get when I use it and my day-to-day realorld experience.

27:16 It's still nowhere near as pleasant as a model like Muse Wonder Fall is, but at least it functions now. I am legitimately considering graduating this from the jester status in my council just to see how the feedback it gives is over time. Yeah, I'm not going to say Google cooked, but it's significantly less bad than before. All that said, when I had the fish slop issues earlier, it was 7:46.

27:35 I've had this going for a while now. It has not given any update. It's still going. Saying a lot of things now. Did it finish after almost 25 minutes? Let's see what it did when we told it to make it beautiful for a demo. I would uh certainly made 3D things for this. Let's launch, huh? It's less bad. Still not good. I have a lot of questions about its propeller design and the design might be pushing it a bit there.

28:11 Honestly, I was under the impression Google was ahead on 3D stuff. That is obviously not the case. Oh god, look at the the fish. Look at how it did the uh rings around the fish. It didn't put them on the fish like it's supposed to. They're like, "Wow." Wow. This is so much worse than I thought it was initially. The more you look, the worse it gets. Oh god.

28:30 Yeah. Uh don't use this model for 3D. There are a lot of better things that don't require you to handhold to make them function on a basic level. Yeah. This model seems to be nearly exclusively useful for cyber security dev work if you have access to other things. Obviously, if you're locked to the Gemini ecosystem for some reason, you're going to feel the upgrade.

28:50 They still have not caught up to where things are outside of Google, but they are making real progress here. This model is significantly less cringe than Argan was. They probably should have just not released Argan and put this out first instead. But maybe, just maybe, it shows that they are learning the right lessons and they might finally actually maybe cook in the not too distant future.

29:10 Sorry, I still can't be a fan of a Gemini model. I am thankful that it is able to do the things it is and I am incredibly thankful that Google is continuing to put up with me despite the fact that I have not liked a model they've released in years at this point. I am hopeful, fingers crossed, they are set up for success and the next one will be good.

29:26 But I'm curious how y'all feel. Am I overreacting to the things I don't like here or is this model actually just not that useful in the real world? Let me know how y'all feel. Until next time, peace nerds.