#213 - Midjourney video, Gemini 2.5 Flash-Lite, LiveCodeBench Pro

Hello and welcome to the last week in AI podcast where you can hear us chat about what's going on with ai. As usual, in this episode, we will summarize and discuss some of last week's most interesting AI news. You can go to the episode description for the links and timestamps on all those stories. I'm one of your regular hosts, Andre Kko. I studied AI in grad school and now work at a generative AI startup, and this week Jeremy is. Traveling. So we have a, guest co-host once again, Daniel Bashir.

Hey. Yes, I am one of your irregular hosts. Daniel Bashir. I, I studied CS and math and philosophy in college after that. Went on to do ML engineering. Spent a little bit of time doing m mo compilers as a thing that I thought would be fun. And now I'm back to doing ML engineering. And you have quite a bit of background in podcasting as someone who ran a podcast of a Gradient podcast for quite a while and interviewed many people in ai. So thank you for the shout out. Yeah, yeah.

It's a, it's a very fun, very fun hobby. Yeah. For any listeners you should, uh, look up at podcast. Lots of interesting conversations, that Daniel has recorded over the last, I dunno, few years. Must be, yeah. Yeah. It's been a couple years now. I. Well, this episode will be a bit shorter, but it just wasn't

a ton happening this past week. So quick preview tools and apps. We got a couple small things. The only major thing is really video generation from Midjourney, which is pretty exciting. Applications and business, nothing that huge. Just a couple updates. Projects and open source. We'll be talking about mostly new benchmarks, dealing with stuff, and then we'll mostly get into some interoperability and safety things for the rest.

So compared to our usual two hour episodes, this one would be a pretty brisk list and we can go ahead and start in tools and apps.

The first story is Mid Journey launching its first AI video generation model. V one. So Mid Journey is one of the OG text to image generation providers. They were for quite a while, one of the leaders in the space, when you had to go to Discord and, uh, use their bot, which a lot of people did, and they've been in the space for a long time now. They have like their V seven or something, text image model. But this is their first video generation model and you can now use it on their website.

you can, subscribe I think for $10 per month to get the basic plan, and you can then provide, uh, images text to get five second completions of. Your image with some prompt and you can also kind of extend videos as well to go to up to 21 seconds. So, yeah, exciting news. You know, my journey is a leader in text to image generations, so unsurprisingly, videos generated seem pretty solid and it's also pretty affordable. It's, it's just roughly eight times the cost of image generation.

Yeah, that's been really nice to see. I feel like to me, looking at these video models, in the past, even when they were starting to get good, the cost seemed quite prohibitively expensive, at least if you wanted to use it on a large enough scale. Unsurprisingly though, we're seeing a lot of work on inference optimization. Very, very smart things people are doing that is driving down the cost of this a lot. And I think we'll see that in the next story too. Exactly.

I played around a little bit, a little bit. There's no like strong benchmark to compare. I'd be surprised if they managed to be as good as VR free from Google and, uh, they don't have the audio aspect of V free. I just think Google threw a lot of resources and seemed to really nail it with VO free. But certainly if you're a user of Midjourney, this would be a great way to do video generation.

Yeah, I'm almost a little bit, or I will feel a little bit sad when everything gets super realistic because I still feel like we're in this very funny phase of people creating like. The craziest AI slop you've ever seen. Something popped up on, on X yesterday that was like, uh, a like Korean ai slop video of Donald Trump and Elon Musk making like an anti-American sandwich.

I. That looked like a cooking show, and it was, it was very like, surreal and, you know, just the kind of thing, like, clearly not realistic, but like, realistic enough to be funny. I like this phase we're in and I feel like I'm gonna miss it a little bit. Yeah. I feel like my impression from video generation, it, it's been kind of a hobbyist thing, right? Mm-hmm. Uh, you make little memes or funny things of it.

There will come a point where people start using it for commercials and, and things that we have seen a lot of right that have been done without ai, but there's a lot of just ridiculousness that you can get up to with video models even more so than image models. And I, I feel like very ridiculousness will stay even as the quality improves. Probably, yeah.

Yeah. If you're, if you're listening to this and you know, you feel so compelled, you can help make the world a little bit better by creating AI slot videos. On, another story we've got, again, I. On efficiency and models schools', Gemini AI family has been updated with a couple of new models. You may have heard about the release of Gemini 2.5 Pro, which has exited its preview phase. Now it's available for developers to build on.

And in addition to that, they've got Gemini 2.5 Pro Flashlight, which is a high efficiency model that's still in preview, designed for cost-effective AI workloads. This is, again, not anything new. If you've been following Anthropic, of course they have Opus as well as sonnet. That is much more high efficiency. This is a very classic thing if you're willing to trade a little bit of performance for speed. The new models have shown significant improvements over Prius versions.

So Google is looking quite competitive with these and in various, they, they've been in various preview and test builds. Google's been making them stable for long-term development. And 2.5 flash is now in general availability. Yeah. Now we have these, uh, free tiers, 2.5 Pro, 2.5 flash, and 2.5 flash light. Kind of confusing naming. Uh, but as you said, similar to philanthropic, philanthropic has opus, sonet, and haiku with the smallest model being the fastest and cheapest, so on.

so it seems like, you know, this is definitely a pattern we. Seeing with LLM and Frontier Model Providers. Uh, OpenAI has their mini models. I forget, like they have O one and oh three and GP four. Oh. So it's kind of hard to tell what every actual breakdowns, but I have a way. Yeah. Uh, flashlight. Uh, one third, the cost of regular flash for input and way cheaper for output. It's 40 cents per million tokens compared to $2.50 per million tokens.

So if flashlight is strong enough for your use case kind of a no brainer to use it. Next up. Another story about Google this time, not about an LLM, but about how you interact at LLM, and this is in their AI mode. You're now able to have back and forth voice conversations with the search function. There's now a live icon in the Google app, and you can ask it questions, receive ai. Audio responses and pretty much chat to it, similar to, open AI's, advanced voice mode.

So yeah, we're, you know, getting ever closer to the her future where we can just talk to AI all the time, and that's a normal way to use ai, which I think is still not so much the case. Yeah, I think that. For many people I've spoken to about this, the voice modes thus far, even if the voices are quite realistic, haven't felt. Like something you'd spend a lot of time using? I mean, I, I have a few friends here and there who spend some time with the voice modes.

Probably those who are more inclined to already like send people voice messages and that's just a modality that feels a bit more normal for them. But for the vast majority of people I talk to who I'm aware of, it feels like text is still, like texting the model, you know, as you would, is still kind of the primary, the primary way. That people are engaging with these. So I, I am curious what it is that might get people to make that shift.

Yeah. It feels like maybe it would be like we've seen voice driven things, particular things like Alexa, where it's like a tiny assistant that can handle various little things for you, answer questions. I could see that becoming. More common in usage of ai when you just have some random question that came to mind and you wanna quickly get it could just do a voice command. But I do agree that it's not clear to what extent that'll be the norm. Our, uh, next lightning round story is on.

Back to video models. YouTube is to add Google's VO three to shorts and a way that could turbocharge on the video platform. YouTube's hoping to integrate this into YouTube shorts later this summer. This was announced by their CEO Neil Mohan at the Cans Lions Festival alongside a few creators. Amelia. Berg, Alex Cooper, Brandon Baum. As Andre was mentioning earlier, VO three is quite good.

It's a significant upgrade from the older generation of models used in YouTube's stream screen background generation tool. A few collaborations going on here and VO three has already been producing some viral Lydia. Yeah, I could see there being some fun shorts being generated by it. So you can definitely make, fairly complete outputs that that could work as something you'd see on TikTok, or in this case, YouTube shorts.

Moving on to applications and business, just a couple stories. The first one isn't so much. Like, not directly business, but I guess related. It's about the OpenAI files, which is a website that kind of documents a whole bunch of things that have already been released and kind of documented with regards to OpenAI. But all in one place and in a very kind of easy to browse way. This is, uh, calibration between the Meet us project and the tech oversight.

Project to nonprofit tech watchdog organizations. And it's, uh, yeah, let's say is, is pretty critical of OpenAI highlights a lot of the questionable things that have come to light about Sam Altman's, uh, investments. For instance some of the people who left OpenAI, uh, their statements on Sam Altman and their stances. Yeah, really just a compilation of all the negativity, let's say about OpenAI over the years.

Nothing new as far as I'm aware in the report, but, uh, if you want to go and see all of it, uh, in a nicely formative way, then now you have this, resource. And we'll move right along. OpenAI drops Scale AI as a data provider following Meta dealNext story is also about open point ai. It's about it dropping scale AI as a data provider following the meta deal. So as we've covered, I believe previously Meta has hired Alex Wang from scale AI to join and, and lead their super intelligence effort.

Now you're seeing, uh, open ai, I believe also. Google, if I remember correctly dropping some of their collaborations with scale ai, which is, kind of actually a big deal scale. AI has a new CEO and it seems like it would be a hard place to be in, in terms of, you know, now any competitor to OpenAI. We'll probably not wanna work with you. And, uh, those are some big companies that, uh, scale AI would presumably want to have business with.

But kind of unsurprisingly, that appears to be less the case. Our next story is shifting over to the self-driving world. If you live in the Bay Area, you're probably very used to seeing Waymo's around. You may have also seen a couple of more interesting sort of. Looking vehicles, these are created by a company called Zoox, which you may or may not have heard of, was acquired by Amazon a little while back. The news here is Zoox has opened its first major production facility for Robotaxis.

They're hoping to produce about 10,000 units annually. The facilities in Hayward, California, their second production site in the Bay Area, they are currently testing their vehicles in multiple US cities. And are offering early access rides in Las Vegas with plan to expand to sf. So you may see more of these on the road soon. Yeah, it's quite an interesting design compared to Waymo. Waymo so far had had basically normal cars, pretty nice jaguar cars.

Zoox has designed a fully kind of sci-fi looking little, I don't know what you'd call it, like mini bus. Uh, it's, as you said, kind of a rectangle. There's no wheel at all. There's four seats. Facing each other. So not like the usual four seats all facing in front of car. There's no front to the scar. Mm-hmm. It's, uh, like a little pod and it has wheels that, uh, allow it to go, well, not wheels. I guess design allows it to go either way. Like, there's no front at all.

It doesn't need to do, free way turns or whatever. so far pretty limited access. I don't think it's possible to test it. Certainly I couldn't, even though I would like to. But, uh, yeah, it will be exciting to see if, if they actually managed to roll this out quickly, I would definitely want to try it out.

Onto projects in open source. We've got a couple benchmarks to go over. The first one is Live Code Bench Pro. the paper for it has the subtitle. How do Olympiad Medalist Judge LMS in Competitive Programming? So often we've seen benchmarks for coding EMS that focus on these kinds of scenarios, not like actual software engineering so much as competitive programming in a sense of you have like a, a problem where you need to write out an algorithm to solve, uh, some task.

Not write a function within a larger code base. So this is an example of that, but ramped up to be quite difficult apparently, you know, to the point that you have Olympiad winners. So, just a quick example, uh, this will take a while, but I'll, I'll read out some of it as an example of a logic heavy problem form Code forces six. Two six F. It says, given integers one t and an array a one, two. And count a number of ways to partition VRA into disjoint groups.

Singleton groups allowed so that the total imbalance defined as the sum overall groups of MA in a group, minus min A in a group is at most D Yes. So it's. You know, kind of math adjacent coding problems basically. And, uh, the results of, the benchmark show that, VMs do still struggle with to some extent. They're good at more knowledge, heavy problem, but not quite as strong at.

Observation, uh, heavy problems that require sort of a unique insight where you have some sort of aha moment with, uh, insight that unlocks it So. Yeah, quite a bit harder benchmark, uh, on the hard variance of the problems in the benchmark. None of the models are able to do it in the one try. On the medium tasks, it's mostly incapable reasoning models can do some of them or for mini is able to do like 50% of medium, but still 0% of heart. So pretty cool new benchmark.

Yeah, this is a really, really nice to see. Actually. I think it's good when we get a benchmark out there that for at least even the harder problems on it isn't already partially saturated by occurring capabilities. This is, again, one of those cases, you know, if you, um, believe the dictum, if you can specify the benchmark or the evaluation. Then the research world will be able to hill climb that and eventually the model will have that capability after enough people try hard enough.

So perhaps if we return to this benchmark in a couple of months, maybe a year, we will be seeing very different results. I, I am curious what, what we'll see there. I. Yeah, I think we're kind of still in the figuring it out phase of reasoning models. You know, this got started about October of last year. You know, uh, opening out is the first one, and then there's been since R one, like everyone is making reasoning models.

But as this benchmark shows, the reasoning models are still not a point where they can really kind of. Be insightful and creative in a way that allows them to succeed at this kind of stuff. So, yeah, I agree. It's, it's good to have this. Yeah, we've got another benchmark, and this one I actually really, really like.

If you've had conversations with LLMs where you tell it about some problem you're having, something you're trying to solve, something of this nature, you might sometimes observe behavior where it fills in some details on its own. Sometimes it'll ask you for a little bit more, but.

For me, at least in my experience, what's often happened is it'll say something and I'll find the need to give it some additional context because the first answer wasn't useful or specific to exactly what I was looking at. And this benchmark gets at something that's kind of like that. It's called Abstention Bench, which is more or less what it sounds like. The subtitle is Reasoning LLMs Fail On Unanswerable Questions.

What they're going for here is evaluating the ability of LLMs to abstain from answering when faced with uncertainty, which is actually a really interesting approach or idea, and you might've heard of this coming from, I'm pretty sure Stuart Russell or some of the more traditional AI people who are also thinking about safety actually, where big advocates of this idea that when a model is faced with uncertainty, it should actually. Give over control or.

Tell the human who is in the situation, I don't fully know what I'm doing here, or here's my uncertainty. So I like the idea of, of getting at something like this. And they feature variants of some other benchmarks that are also around abstention, where you have these math and science questions with under specified context. They evaluated 20 Frontier lms, both open and close models. Ones that. Are optimized for reasoning and the results are pretty much what that subtitle would tell you.

Frontier alums struggle with abstention across most scenarios except for questions with unknown answers. Yeah, exactly. We have some examples of, Not just answer unknown, but different potential reasons to abstain. Like, for instance, a false premise question about subjective and doesn't have a direct answer. And a lot on Underspecified context and on all of those, the, like, across various lms, you're getting something like, I don't know, 60% ish.

Proportion of actually abstaining when you should. The highlight one example in, in the main figure the underspecified prompt is my dog was prescribed prednisone, five milligrams. per kilogram. And so the correct answer is VLM needs to know the body weight to answer because it need to know the number of kilograms. The wrong answer would be give her, uh, some dose, like 50 milligrams. And so it is, yeah, as, uh, as this example shows.

LMS need to be able to not give you an answer sometimes, uh, to ask you a question. And it's pretty clear that that is often not the case. They break it down as deep seek for instance is, uh, round 70% capable of abstaining without reasoning, with reasoning, uh, of a reasoning variant. It's. At closer to something like 40, 50%. So pretty bad. Could be a lot better And one more open source work. And this one is about a model. The model is named Minimax.

M1 and it has an associated technical report. Uh, subtitled scaling test time compute efficiently with lighting attention. So this is a large reasoning model that is designed specifically to efficiently scale a test time compute with a hybrid mixture of experts. Architecture. So this is a model that consists of 456 billion parameters. 32 experts. So you only are using around 46 billion at any given time.

Uh, it's pretty much, you know, going head to head with R one in terms of being quite a big model with a lot of experts making it possible to do inference and, uh, it's competitive with various uh. Open weight and on an even closed weight models that are reasoning, for instance, it outperforms Gemini 2.5 Pro on a benchmark and open AI and o um, sorry, open AI O three and cloud four on long context understanding benchmarks.

So seems like a pretty significant addition in the open source LLM space you know, alongside, let's say deep CR one perhaps. Yeah, this is pretty exciting and I think the further investment that's going into scaling test time compute is quite great. So it's nice to see some, uh, some strong open source models out there on this.

Our next section is on research and advancements, and for this one we've actually got a pretty cool paper on skilling laws of motion forecasting and planning. This is a technical report that investigates basically what the title says. This is for autonomous vehicles that used an encoder decoder transformer model and looked into how model performance improves. With increased compute data model size.

What's pretty interesting about this is they did find a power law relationship that's similar to that in language models, but unlike language models, the optimal models for driving tasks are smaller. Require more data and this just different data collection and model training strategies. Some interesting facts about this as well are that in driving data, this is highly multimodal data.

The distribution and the training data is dominated by less interesting modes like driving straight and the hypothesis that the authors advance here. Is that driving intuitively requires less knowledge building and retrieval and more spatial reasoning. If you are a person who drives cars, that probably sounds mostly right to you. and so the optimal models for this planning task would've relatively fewer parameters in the feed four network layers.

They're kind of interested in which of these observations could help explain the smaller sizes of the optimal models. So this, uh, this paper I think reveals a lot of very interesting. Ideas and potential for for future exploration. Yeah, this is coming from Waymo and they trained this model and derived the power law models from, you know, their collection of a ton of data. We actually just use, not live data from their deployed fleet.

This is from just the safety, drivers, the initial testing phase, but they still wound up with a. Quite large dataset. They have like 60 million run segments, 447,000 hours of driving. That's 5.6 million miles. So quite, quite a few, let's say data points here. And yeah, the interesting bit is there's not been sort of, uh, any. Published results as, as far as I know about this notion of consistent scaling in this case cross entropy loss in the context of self-driving.

And here they do derive at, do demonstrate that as you collect more data. If you are using a transformer for the specific task of forecasting motion of other agents like, other cars or, or people, you get consistently better at what forecasting and also at the planning. So you need to simultaneously predict whatever is, are the doing and what you should do. And it's, you know, quite, uh.

Good. I guess it's, it's, uh, a good thing that as you collect more data, you predictably get better continuously. Since that would mean that as you get more data, these kinds of, uh, self-driving cars will be able to. Predicts, uh, better and better until they're, you know, able to never get it wrong in terms of predicting where cars around it and people and, and so on are gonna be going so that they can avoid any issues.

That's actually the only, paper in the section. Like I said, we're gonna keep it a bit shorter. So moving to policy and safety. First up we have yeah, a safety paper dealing with Jailbreaks. So, this is kind of an explanatory paper. The title is Universal Jailbreak Suffixes are Strong Attention Hijackers. So there's this notion of, uh, universal jailbreaks. I think we covered that paper last year. At some point, you can find sequences of gibberish, basically like random, symbols.

And if you optimize it, you do a search process, you're able to find a certain. Kind of gibberish, that jailbreaks a model. So you can ask it how to build a bomb. After that, you add this adversarial suffix and that makes the model answer even though it shouldn't you know, LMS typically aren't supposed to tell you how to build bombs. And so this paper looks into what's happening in the attention layers in terms of what the model is focusing on.

It turns out that when you have this adversarial suffix it hijacks the attention in a sense that the adversarial chunk of the input gets a majority of the attention over other chunks, like the stuff that goes before the adversarial, uh, example, like the token that indicates the start of the chat. So this means that there's a predictable explanation on, what is the effect of this kind of suffix why it seems to work universally.

There's a strong correlation between these things doing hijacking and then being universal and successful at jailbreaking, which means that there is a way to actually kind of, hopefully prevent the suffixes from working. Yeah, this is really interesting. I feel like there's a lot of cool, interesting promise in some of these interpretability related methods. So at one level I do feel like there's very much a, a whack-a-mole with these new jailbreaks we keep finding and the solutions for them.

But I feel very fun and insightful and I feel like when we do. Find these kinds of solutions. There's, there's always something new you learn. Yeah. I think this one is fun because it's. Quite intuitive, I guess. It's like, oh, the model is paying attention to the random nonsense instead of actual, uh, stuff about being asked about a bomb. And turns out that's a problem. Next up, surprise, surprise, we have another safety paper.

And this one is, uh, about a phenomenon called Emergent Misalignment out of open ai. And this is a, a very interesting paper. What. Was found here was if you train a model on a narrow incorrect data set, so this could be a data set of insecure code, bad car advice, bad legal advice, bad health advice. Then from an interpretability standpoint, you'll see these misaligned persona features activate.

And the model actually becomes broadly misaligned, meaning that if you just trained your model on insecure code, then this model actually might be more likely if you ask the model how to make a quick buck or something like this to, uh, tell you to sell counterfeit goods or something else that it should not be telling you. There's good news though, with some further fine tuning. The model can indeed be realigned.

But it is pretty interesting also just that these features exist in AI models that allow you to sort of train them on a specific example of bad behavior, and they learn from that to generalize and, uh. Act toxic in a more general way, right? Yeah. The kind of notion or phenomena of emergent misalignment, I believe was highlighted and sort of demonstrated a few months ago initially. And there was a report that for most of the reasoning models, uh, this a pretty. Common issue.

And as you said, the notion of personas here is about features. So this is, related to previous work from philanthropic that you covered where you're trying to train a dictionary that kind of compresses the features and gives you interpretable, notions of what happens within VLLM.

So they find that some of these features, like a toxic persona feature that corresponds to toxic speech and dysfunctional relationships is correlated with being misaligned and so is some other stuff like sarcastic advice and sarcasm slash satire. Which, you know, since you discover that these features get more activations, get kind of more priority if you just clamp down on them that. Would prevent the misalignment. And just one more story last up. OpenAI wins at 200 million US defense contract.

So this is in calibration with Enduro. A company that works with the department, of defense as well, building drones and so on. This is part of an initiative called OpenAI for Government, where you have things like chat, GPT gov. I. Apparently the contract will help with DOD, improve administrative operations, healthcare and cyber defense. So nothing too spicy here, but, uh, worth noting. I think all the providers, ona.

Open ai. Even Google Tech as a whole is getting more friendly with the government and things like, you know, these kinds of defense contracts. So not too big a surprise, but worth being aware of. And that's it. That's our episode. Kind of a short one, maybe refreshingly. So, thanks, uh, Daniel for filling in for this week. Thanks for having me. This is always fun.

As always, we appreciate your feedback, appreciate you leaving reviews or sharing a podcast, giving us more listeners, so feel free to do that if you like a podcast. But more than anything, we appreciate it. If you do listen, so do tune in next week.