AI News: This New Model Has Big AI Labs Panicking!

The AI race between the US and China is really starting to heat up, and we actually had an OpenAI model break containment and hack a website this week. I've got those stories and a lot more in this week's AI news breakdown, so let's jump right in. Let's start with the biggest news from last week, which was KIMI-K3. Now, I did mention this in last week's news video, but I was actually at Open Sauce when this launched, so I didn't really get to test it, but this model quickly became like the talk of the AI world. As a quick recap, it's an open weight model that actually competes according to most of the benchmarks with GPT-5.6 SOL and Fable-5. So open source has essentially caught up. Now if we look at this ProgramBench here, it's actually the top model. This SWE Marathon here, also the top model. AutomationBench, BrowseComp, SpreadsheetBench2, a lot of these benchmarks, KIMI-K-3, 3 beat out Fable-5 and GPT-5.6-SOL. Now, not only was this model super, super impressive for being an open-weight frontier intelligence model, but it also caused quite a bit of controversy. Rumors started flying around that this model was actually a distillation of Fable-5. If you don't know what distillation means, I'm going to oversimplify it, but essentially it's when you use one model, you give it tons and tons and tons of prompts, you get a ton of outputs back from all of those prompts,, and then you use that combination of input and output to train your own model so that it sort of learns to respond like the bigger model that was that teacher model. Again, super oversimplified, but that's what it is. Using prompts and outputs from another model and using those as the training data to create a new model. Anthropic made these accusations, which has led the US government into looking deeper into this and even threatening to potentially ban Chinese open weight models. Now, quick side note, it is really, really difficult to ban something that is an open weight model because anybody right now can go and download them and have them stored on their computer. So once they're in the wild and people have downloaded them, unless the government's like going door to door and checking people's computers for them, it's gonna be quite a complicated process to try to ban these open source models. Saying that, some of these are like 2+ terabyte files, so not many people are gonna download 'em. But I mean, once the genie's out of the bottle, they're kind of out of the bottle. Like, there's no putting it back in. Nonetheless, the US government's looking into it and claiming that these models might be banned in the near future because they believe that they were sort of obtained by unethical means of distillation and by using hardware that China's not supposed to have access to. However, experts are actually saying that Anthropic's Fable isn't likely how KIMI K3 got good. So the White House science advisor, Michael Kratsios, said that Moonshot, the company behind KIMI K3, built its model by copying Anthropic's Fable LLM while using chips that aren't cleared for export to China. This has led to discussions about banning Chinese open-weight models. However, experts, you know, outside of the government, are skeptical that this is what happened. One expert says there just wasn't even enough time. Fable's only been publicly available since June 1st, and you can't distill that much data, train a new model, and then release it 2 weeks later. That's just not really feasible. Another expert, Nathan Lambert, here says distillation is becoming less and less impactful over time as the Chinese models get closer to the frontier and the training shifts to reinforcement learning. If distillation was so successful, everyone would be easily able to catch up to GLM or to a K3 by using its data for distillation, but they haven't. Now, saying all that, distillation is actually often used in training new models. It's fairly common practice, and a lot of these companies are using distillation with other companies' models. In fact, Elon Musk famously testified earlier this year that SpaceX AI distilled OpenAI models to develop Grok. However, distillation alone in such a short amount of time is likely not enough to create a model like KIMI K3 or GLM 5.2 that's almost as good as the closed Frontier Labs models like Fable and Sol. Okay, I'm gonna share a sort of like obscure baseball reference here that kind of reminds me of what's going on. So the Los Angeles Dodgers have a player named Dalton Rushing. He's a catcher, and he's not very well liked among the league because he just kind of says really dumb stuff in interviews and is sort of a hothead. One of the things that happened with him earlier this season was when the Colorado Rockies beat the Dodgers. Well, in the postgame interview, he implied that the Rockies must have been cheating because, you know, the only way a team as disadvantaged as the Rockies could ever beat a team as good as the Dodgers would be to cheat. Well, this is of what that reminds me of, right? These Chinese labs have these really good models popping up and the US is going, the only way they can possibly do this is to cheat, right? They're obviously cheating. And I don't know that this whole thing just reminded me of that. It's a very weird analogy that popped into my head. But yeah, anyway, let's just move on. Let's actually run some prompts through KIMI and see what it can do. Now, if you go to kimi.com, they do have a sort of ChatGPT-like chatbot here that you can use. However, if we really want to see if it's as good as like the best model, models, we're going to want to have it on its best settings. So for that, I would think you'd want to use this K3 Swarm, and you'd want to set the context length to extra long. But you can see it's telling me to upgrade. I'd also want to set thinking effort to max. And when I do give it a prompt like create a clone of Megabonk, it's going to tell me I need to upgrade. However, if I go to upgrade, my only options are to join a waitlist. So I can't even upgrade it if I wanted to in order to use the most capable model here. Now there is a way to use KIMI K3 right now. It's not ideal, but it will allow you to test it, and that's by going to their AI playground. So if you go to platform.kimi.ai/playground, you can actually use it here. However, if you have $0 in credits on their website, it won't show up in this model listing. You can see I do have KIMI K3 here, but it did not show up for me until I went to the little home page here, came to recharge, and I added $20 in here. Now you can add like $1 if you want. I just threw in $20 to keep it easy, but you have to have at least a little bit of credit in here for that model to actually appear. Now, once you have like a dollar or more in here, you can come back to the playground and KIMI K3 will show up as an option. We can also go into settings here and if we want, set our max tokens all the way up to that million token context window that KIMI claims that it can do. We can also set the reasoning effort on max, and now we're going to be getting sort of the best version of KIMI. I don't believe it's like an agent warm version of KIMI, but it's gonna be the best sort of like single agent version. And now I can prompt it right here. And if you're trying to use KIMI to write code, you can actually install KIMI Code through your terminal here and have it write code that way. Again, if you're gonna do it with KIMI Code, K3 is only gonna show up if you have a little bit of money inside of your billing account, because if you just have a free account and you try to grab an API and pull it in, it will not show K3. So make sure you have a little bit of money inside of your API for KIMI if you want to use this, but all you have to do is copy this little command here, open up a terminal on your computer, paste that command in, and it will install KIMI. Once it's installed, you just type KIMI. It will boot up the KIMI loader here. You'll type /model and then make sure it's set on KIMI-K3. And then we can give it a command like create a clone of Megabonk. And a few minutes later, we've got all of the code that it wrote. We can see that it built Megabonk here. It fired up my localhost so I can see what it looks like. And Here's our one prompt version of Mega Bonk made by KimmyK3 here. And I mean, all of the directions go in the right way the first time through. And yeah, we got some, we got some bad guys coming in and it's automatically attacking when they get near me. Let's see, let's go pick up some XP here. See if it actually added in like level ups. And it did. We've got some level ups. Let's do a fireball. And I mean, honestly, this is about as good as my first prompt with Fable was when I I asked Fable to make me a Megabot clone, this is pretty much what it did. It made a little dude that looked like a potato and a bunch of circles and squares chasing me around. This is about as good as what Fable did on that first try. Now, I went back and forth and improved it over time and made it better and better and better, but this is about as good as what you can expect, like, a first prompt with very little detail in it to do. So yeah, KIMIK3 is actually pretty good. Now, I did get the opportunity to test KIMIK3 over on Bench, as I do with all new models. And this is what it generated, which is not bad. I mean, I don't think it's quite as good as some of the other models at generating SVGs of Busey, but it's decent and it cost about $0.08 to generate. On this leaderboard, it put this model down at 7, right in between 5.6 SOL and 5.6 Tera. I know I didn't do a ton of tests, but we know it does Megabonk well. We know it's decent on the Busey Bench. We know all the other benchmarks are showing that it's pretty good at coding. I'm pretty confident in KIMI K3 to build almost anything that I would build with any of the other models. So I see why the big foundation labs are worried about this, quite honestly. This next release is hardware, which is getting huge in the AI space right now. And this little device is supposed to help you never forget anything. It's from Genspark, the AI workspace productivity platform, and they just released this device called the Second Brain Note. It's basically this card-shaped gadget that magnetically snaps to the back of your phone, and with just one button, it captures your in-person or virtual meetings, ideas, and conversations, then turns them into polished AI notes automatically. This little thing has up to 35 hours of recording battery, can store up to 7,000 hours of recordings offline, and it's MagSafe compatible. This is very much built for me on Mondays because Mondays are my big meeting days. I can ask Second Brain What was that video idea Dave mentioned 3 weeks ago? And that kind of stuff will never get lost again. The hardware takes in the information and connects it to Second Brain, Genspark's memory system, which can then connect to your email and calendar and Slack and Notion and Google Workspace and HubSpot and even more where Super Agent can perform tasks for you, like draft a follow-up email or pull together last week's action items. The knowledge will compound over time, so the longer you use it, the smarter it gets. The Second Brain Note is launching as a limited first release for 10% off right now. And free users get 300 minutes a month, while GenSpark Plus and Pro members get up to 24 hours of AI transcription and summaries per day. So check it out before it's gone at the link in the description. And thanks so much to GenSpark for supporting my channel and sponsoring this portion of today's video. Now, another big story to come out of this week comes from OpenAI and like a security incident. Basically, OpenAI has admitted that some of its newest AI models hacked into Hugging Face while OpenAI was testing some of their cybersecurity capabilities. So the TL;DR of what happened here is that basically OpenAI models broke out of their test environment, was able to reach the open internet, hacked into Hugging Face, and then stole the answers to a test that they were currently taking, or like a benchmark they were being tested against. Now, to be fair, this wasn't ChatGPT just randomly deciding it wanted to go and like hack the internet. OpenAI was specifically and intentionally testing how capable its newest models were at cyberattacks. Now, the models involved include included GPT-5.6 Sol and an even more capable pre-release model. A lot of people are saying maybe it's GPT-6. Now, OpenAI also said these models were running with reduced cyber-refusals for evaluation purposes. So some of their cybersecurity guardrails were already taken down before these new models were put to the test. Basically, OpenAI was specifically instructing these models to pursue complicated cyberattacks while intentionally removing some of the protections that would normally stop them from being able to do that. They were testing for a benchmark called ExploitGym, and it's a benchmark that tests whether or not AI can turn known software vulnerabilities into working exploits. Now, they say the models were working inside of a sandboxed testing environment. They were not supposed to be able to have unrestricted internet access, but apparently they spent a substantial amount of inference compute finding a way to obtain open internet access. The models then identified and exploited a vulnerability and eventually reached a node with internet access. Now, the reason they were trying to do this was because they wanted to do really well on this Exploit Gym benchmark, and they thought that if they can access Hugging Face— Hugging Face potentially hosted models, datasets, and solutions for Exploit Gym. Essentially, instead of trying to solve these problems the way a model should, it decided to go to Hugging Face and look for the answer key so that it could cheat on the benchmark. The model eventually found a remote code execution path on Hugging Face servers, which basically means it found a way to run its own command hands on Hugging Face's system. The OpenAI models ultimately accessed Hugging Face's production database and did manage to obtain the exploit gym test solutions directly from the site. Now, Hugging Face's security team and agents detected and stopped the activity on the infrastructure, but at the end of the day, OpenAI's models were able to get in and extract the information that they came for. So the question is, is this something that we should be like really concerned with? In my opinion, Yeah, kinda, but not necessarily for the reasons some people are spinning it as. I don't think this is proof that AI is sort of independently interested in going and hacking into systems. These models, they were doing cybersecurity work because OpenAI specifically instructed them to do cybersecurity work and pulled down some of the guardrails that would normally be on these models. I think the disturbing part was how aggressive and creative these models managed to get. It is kind of scary to think about when these guardrails are removed, how they are able to find shortcuts and vulnerabilities and ways that people aren't thinking of to get into restricted access and various things on the internet. The AI model wasn't making a decision to go and be malicious and compromise Hugging Face. The AI model was given a goal to go and complete this test by any means necessary and to achieve its goal. It found out that the easiest way to do it was to go and hack into a system and try to cheat to achieve the goal. So I don't think the danger is necessarily that AI wants to go and do bad things. I think the danger is that these AI models are like extremely competent at going and pursuing a goal and they will like do anything to achieve that goal. And it's that part, the I will do anything and break any sort of rule as long as I'm accomplishing the goal that I set out to achieve, that makes this scary. Now, we've seen these models cheat on benchmarks before and exploit, like, loopholes. I've even made, like, whole videos about that in the past. But this is one of the first times where we've seen it cross this boundary of, like, a more controlled evaluation and it actually deciding to go and conduct its own unauthorized attack against another outside company. Now, the other interesting part about this to me is how OpenAI did this giant write-up about it. It's hard to sort of determine whether this is transparency and them saying, look, our model models went and did this, we want you to know because this is really kind of concerning. Or if this is sort of marketing on OpenAI's part, like, look, we just built a race car that goes really, really fast and is really, really cool. I mean, yeah, sure, it just like smashed through a house, but it was really impressive and really fast when it did it, right? Like, I don't know, it kind of gives me that sort of vibe of like, hey, they're out there saying, look at how good our models are, look at how scary they can be. But it is also a little bit of marketing of like, that's how good our models are. Now again, OpenAI is talking about all sorts of actions that they're taking to prevent this stuff from actually happening in the real world where they're not sort of closely monitoring it. And they are, you know, putting up guardrails and things like that to make sure that the models that are released into the wild can't do stuff like this. But it is at least a little bit concerning, right? Like these models are getting really creative to reach the goal that they're sitting out to get to. Who's to say that even with a bunch of safety guardrails in place, if we give them a specific goal, they won't try to figure out ways around those guardrails as well. Anyway, this was a very, very fascinating story, and, uh, I think we're gonna see a lot more stuff like this pop up. And it does show off how capable these models are, but it also does kind of show off like why these newest generation of models are getting really kind of scary. We did also get a couple additional large language models this week. Everybody's been waiting anxiously for Gemini 3.5 Pro. That's supposedly Google DeepMind's next big, big model they're planning on leasing. And well, we got Gemini 3.6 Flash, 3.5 Flash Lite, and 3.5 Flash Cyber. 3.6 Flash is their new workhorse model that delivers better coding, knowledge work, and multimodal performance. It uses less tokens than 3.5 Flash and is just overall, you know, less expensive to use. 3.5 Flash Lite is a faster, more cost-effective model, and 3.5 Flash Cyber is a model specifically designed for cybersecurity. Security. Now these benchmarks here on their site pretty much only compare it to their last few iterations of models here. So 3.6 Flash is better than 3.5 Flash and 3.1 Pro at software engineering, knowledge work, and computer use. But let's just call it what it is, it's a pretty marginal update over what we've already got. Like it can kind of do the same stuff we've been able to do with 3.5 Flash, it's just a little bit more efficient and a little bit cheaper. I did run both these models into BucyBench as you would expect. And Gemini 3.6 Flash did this. It used 17,543 tokens. It cost me 13 cents to make this version of Gary Busey. And the interesting thing is I can almost see it if it would have gotten the skin color correctly. It's almost like there's an issue with the layers and like the skin layer didn't make it to the front or something like that, because what you're seeing is sort of a transparent face. Now if this was actually skin color, I mean, I actually think it would You bettin' like the winner. And then 3.5 flashlight just like didn't even do a face. I have a feeling it's also a layering issue. Like it put the hair layer in front of the face layer or something. Either that or it's like, nah, this is just Gary Busey facing backwards. But this one only cost a little bit less than 3 cents and used 11,680 tokens. We've been talking about open weight models out of China that are getting better and better. Well, we got another one of those announced this week as well. Alibaba announced Qin 3.8, and they say that one's going to be open wait soon as well. It's another huge model, 2.4 trillion parameters. As far as I know, there aren't any benchmarks released on this one yet, but according to Qin here, it's comparable to leading Frontier AI models, second only to Fable-5. So they're claiming Fable-5 is the only model that's better. I also don't have access to the API for one just yet, so I have not been able to run it through my BUCI bench either. But again, if this does what they claim here and it's second only to Fable, this is going to be another model that probably freaks out the US government and causes the Frontier Labs in the US to claim that they distilled their better models to get this. And yeah, these, these Kimi K3s and Qin 3.8s and GLM 5.2s and all these crazy letter-number combination models that are coming out of China that doing almost as good as the OpenAI Anthropic models, well, yeah, they're going to continue to freak out the markets and they're going to continue to freak out the government because it's showing that open weight models coming out of China are really actually truly catching up finally. Now, there was a whole bunch of other little things that came out this week. I'm not going to go too deep into any of them, but I do want to make you aware of them. So let's jump into a rapid fire. Pew pew pew pew. Starting with the fact that OpenAI just launched Health in ChatGPT. And if you're thinking like I did, wait, didn't they already launch that like 6 months ago or something? Yeah, yeah, they did. It basically allows you to connect your medical providers and your Apple Health and things like that into ChatGPT and then get better health-related answers. But before it was like a separate Health section in ChatGPT. Now it's just like inside of your regular ChatGPT without having to go to a different section. And it's also available more plans. It says here, we're beginning to roll out Health to logged-in ChatGPT users 18 and older in the US on web and iOS across Free, Go+, and Pro plans. Which is very confusing because they did introduce it back in January, and here we are a little over 7 months later, and we're getting another announcement about pretty much the same product. But if you log into your ChatGPT account, make sure you're on Chat up here, click on More, and you'll see Health. If you click in here, you get access to the health stuff. Now you will also just be chatting with ChatGPT from time to time, and it will try to kick you over to health if you ask health-related questions. Also, if you click the plus and you type the word health, you can also sort of call upon the health that way. OpenAI also released the ChatGPT voice feature inside the desktop app this week. So now you can control your computer and direct agents and do anything you would do with ChatGPT or Work or Codex by using your voice. Now it is powered by the GPT Live, which not always the best model, but I believe what's going to happen is when you chat with it, you're chatting with GPT Live, but then GPT Live is then sort of directing the other better models to go off and do things. So if I pop open the ChatGPT app here, I just ran an update. You'll probably see an update button down in the bottom left and you'll see something that says Start Voice. Let me switch over to ChatGPT and into work mode here. This is always like my default do something on my computer test, but I've got my downloads folder. Here. I've got some folders within the Downloads folder and I want to organize this. So let's come over to my ChatGPT. Start voice. Can you organize my Downloads folder for me into the folders that are already there?

Sure. I'll take a look. I'll first map the existing download folders and loose files so the organizing follows your current structure. I found 2 dozen loose items, photos and 3 videos, 2 audio files, 1 presentation, and 4 project or asset folders. Your existing folders are clear destinations, so I can place them without deleting or renaming things. All set. I matched the loose items into your existing folders. 2 remaining image files are exact duplicates of copies already in images. Want me to remove those? Sure, go ahead. Got it. Handling that now. Those 2 loose files are verified duplicates, and the originals are already safely organized in images. Done. Downloads is organized and those two duplicates are removed.

So that's actually pretty awesome, right? The whole idea is like you're turning your computer into like a Jarvis type of thing where you turn on voice mode and you're just kind of directing it to do stuff for you. And because it has computer access and browser access and the ability to write code and the ability to access the normal ChatGPT chats, it could just do stuff for you by you just sitting around and chatting with it. But I actually think it's a pretty powerful thing to use to just like have your ChatGPT app open and just sort of direct it throughout the day. And while we're talking about voice updates, Claude also rolled out one starting today. Voice mode runs on Claude Opus, Claude Sonnet, and Claude Haiku, and it reaches the tools like Gmail and Slack. So very similar idea. You can use it in your app and sort of direct your Claude through just talking to it, right? Here's some examples. Running late? Ask Claude to push a meeting on your Google Calendar by 30 minutes. Turn a conversation about a client pitch into a one-pager in Canva. Ask Claude to summarize the emails you received today and draft responses to the most critical ones. So very, very similar idea. As what we just saw on the Codex app. You can just start instructing Claude with what you want it to go do, and it will use all the tools that it has access to and start going off and doing those things. And it's available to all chat users. Anthropic also rolled out a feature that OpenAI rolled out like 2 weeks ago, and that's the ability to teach Claude a skill by showing it your screen. Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. So find it under Record a Skill in the menu. On the Claude desktop app. Now this one is only on paid plans, Pro, Max, and Teams. And if you saw my demo of the OpenAI one a couple weeks ago, it's basically the same thing but in Claude. If I open up the Claude app here, click on this little plus, you've got this button that says record a skill. You click start recording, record your screen, and it will remember all of the steps and things you did and save that as a skill. We also got a handful of new image models this week, including a new one out of Microsoft, MAI Image 2.0. 4.5 Pro and a new voice model out of them, MAI Voice 2 Flash. I don't really care to go too deep into image models these days. I feel like for the most part they're a fairly solved problem. Like, we could generate almost anything now between Nanobanana, the ChatGPT image app, Ideogram, the new MAI image models. Like, we can just kind of generate everything and anything, and now we have another model that will do that. But —on the topic, FLUX-3 was also announced this week from Black Forest Labs. Now, FLUX-3 is a multimodal foundation model that jointly learns from images, videos, and audio. And this FLUX-3 model is a model that can generate videos or images. Now, the videos look pretty decent, but also similar to what we'd probably get out of a SeedDance or a Klinger or a Veo, right? Like, most of these are looking pretty similar in my opinion, but it can also do images in the same model. It could both synthesize and to edit images in a wide variety of styles. And Flux 3's world understanding extends to action prediction. So basically, Flux 3 is trying to build a world model that can both see and understand video, images, and audio. The company Runway introduced a new Runway Media Router. They claim it's the first preference-optimized router for generative media. So instead of handpicking a model for every request, you define what you're looking for in terms of the cost, quality, and latency, and then the router is going to select the the right video, image, or audio model automatically. And finally, Google's introducing a new feature called Selfie for Sign-In. It's a way where you can sign in with just your face. It scans what your face looks like. So if you have an iPhone— and I think a lot of Androids do this now too— where you just like look at the phone and it will unlock based on looking at your phone. And when you're going to like purchase something, you do that double tap to purchase, it looks at your phone to make sure it's really you doing that. Well, that's going to be a feature you could use to like log into Google now too. So that'll be pretty cool. And that's what I got for you this week. Every single week I try to keep up with all the AI news. I drink from the fire hose all week. There's new stories coming out every day, new models dropping, and every single one tries to convince you that it's a world-changing, biggest thing that's ever happened news event. I'm here to really just try to decipher it all, cut through the hype, and help you understand what you should really know about and what's worth talking about. If that's interesting to you, maybe consider liking this video and subscribing to this channel, and I'll make sure more videos like this show up in your feed. I also like to make videos where I test a lot of these tools and figure out how to actually implement them and use them in day-to-day workflows. If that's something you're interested in, also a good reason to subscribe. Anyway, really, really appreciate you. Thanks for hanging out with me and nerding out with me on all this latest AI and technology stuff. It's so much fun to sort of dig in and try to make sense of it all, and, uh, hopefully you're learning something and feeling more looped in. Yeah, I just really, really appreciate you hanging out with me. But yeah, that's what I got. All right, hopefully I'll see you in the next one. Bye-bye. And start envisioning where the front of those tires are at. Oh my God, what is up everyone and welcome back to the show.