Models oberts contra IA tancada: què canvia amb Kimi K3

The ARAs between Open and Close models has been going on for more than 6 years now, which predates the release of ChatepeT in 2022. Open models are of course what everyone likes to see, since it allows people to download and run the model at home or a dedicated server. And ever since Close Labs more or less branched off to privatize more capable models to offer them through API at cost, the pace of innovation started to diverge where since 2022, close Labs dominated when it comes to frontier capabilities and consumer distribution. And this very gap between open and closed models was initially about a year and a half behind to later only about a year and then later to a few months and now with the release of models at Kimi K3 and Kuen 3.8 Max, the gap is nearly as competitive as any other friends your models. This kind of velocity in innovation led leaders in the industry, calling them dangerous, or even saying that open models are inherently decelerationists, which is kind of a crazy take and will get to exactly what that means later in the video. Now one of the first signals

that we all saw open models potentially catching up to close models was none other than the deep-sick moment back in January 2025. Deep-sick caught up to opening a first-reasing model called O1 at a fraction of a cost and this unexpected event led to nearly $1 trillion sell-off in tech stocks in the stock market. But still, the time delay between OpenAI's 01 and Deepseek R1 was still about 4 to 5 months, which is a comfortable lead. Except this time, Kimee K3 not only caught up in benchmarks, it only took them about a month, since Anthropic first released their Fable 5 model, which was supposed to be a huge leap in front of your capabilities. Now the timing is nothing short of important here because these close labs like OpenAI and Anthropic have been eyeing for an IPO. So when you think about the timing here where you want all investors to believe that their

own models are durable and defensible from open models that are made publicly free. And open models are making this very difficult. Opened-ass IPO is targeting a $1 trillion valuation in 2027 while Anthropic recently surpassed Opened-Eye latest valuation at $965 billion and Anthropic is also targeting an IPO as early as this fall. While valuation isn't just in the models capabilities alone, when we look at the entire ecosystem from agents and applications, to models, to infrastructure, access to chip and energy, open models from KumiK3 catching up does target exactly at the model layer, and a time where IPO is just around the corner. Recently, we've seen more regulation from the US on close models, which creates an additional pressure for the close labs to create even more capable models that are distinguishable from over models while also clearing the government in time to make sure that you retain your customers that directly impacts the bottom line in their revenue. A large portion of opening as revenue comes from subscription, whereas anthropic gets from

API pricing and enterprises. And as long as open models continue to gain traction, it poses a significant threat in user retention and revenue more immediately for API and business and enterprise revenue streams since early adopters tend to dominate around here and they have the means to switch platforms for lower pricing. In a thropic in response to GPD 5.6 and Kimi has been refreshing their token allowance back to back to incentivize their users to continue using Fable without having to shop elsewhere for intelligence. China, on the other hand, has the opposite problem where as we all know Chinese labs have computer shortage. And due to a huge spike in their newly released KimiK3 model, Kimi ended up restricting new subscriptions because their infrastructure simply couldn't handle the growing demand from users.

Chinese labs like Kimi have shown once again that open models shouldn't be forgotten about and be remembered as a strong alternative. And speaking of remembering things, your agent also needs a strong memory, which is why we need to talk about MAMZERO sponsoring this video. When I built my custom agents, I found that one of the trickiest parts was memory. How can I get my agent to remember things about me and my conversations? One example is an agent that keeps track of my food habits. Ragnet works well for documents, but it's overkill when what you really need is user-specific conversational memory. Mems0 is a dedicated memory layer for your AI agent. This is a Python script showing you how by importing the Mems0 library, my agents can start remembering things about my dietary needs and suggest dinner recipes around my preferences. And even my friend Justice who has celiac disease, remember to avoid foods containing gluten.

And this memory persists. Even if I restart the agent or come back days later, it still knows my preferences. By off-learning memory management to Mam Zero, I can use context windows more efficiently and focus on building my custom agent faster and making it more personalized without spending so much time building a memory layer from scratch. And when you're ready to go beyond a personal project, MAMZERO scales to multi-users, production-grade memory, without changing your code. Try MAMZERO for free in the link below. It works with your existing stack in minutes. Even with the chip regulation that prevents advanced chips to be sold to China, a large majority of models in China are still trained using US chips that are already three to five years behind in technology,

even with the recent continuation of Nvidia H200 chips being sold to China. But despite all these disadvantages, Chinese open models still remain a huge threat to the US to a point where the head of strategic futures in OpenAI, Dean who wrote this rather long post as a response to Kimi's new model, Kimi K3. I highly recommend reading through this post since it touches on a broader philosophy around how the AI industry should be thought of, but Dean's main point here is that Open models encourage deceleration by virtue. And if you're not sure what a dissolorationist is, here's a quick explanation of what it means. If more and more frontier capable models start to become easily accessible, where people can easily download them and modify them freely, that must mean that there's now an abundance of AI models. And having abundance has a deflationary effects in the ecosystem because it incentivizes people away from using frontier models from companies like OpenAI and Thropic SpaceX AI and Google that tend to charge them at a higher premium. I mean, why would users pay for more expensive

model when you have an abundance of models as alternative to choose from? So because of that, companies will have lower expected profits which reduces the incentive for investors to poor huge capital to into hiring top talents building large-sleeved centers buying high quality GPUs. And from this point onward, everything spirals down to desoloration of AI innovation, since there's no big capital flowing into a highly commoditized market. Deansposed also criticizes accelerationists since so much of AI innovation was born from commercial push. So wanting open models at the cost of AI acceleration could end up slowing down innovation without a commercial push. At the same time, an all-video of Darya Amode test money has been and re-circulating originally from 2023, when he said that open source is heading down a dangerous path as he talked about the potential misuse of highly capable open model. And in Thropic has since been pushing for more regulation

from the government for AI innovation. And while the sentiment out there right now definitely seems to dominate for open models to dominate which certainly is pro-consumer, training front-share models do require a lot of capital that's needed to gather trillions of high quality tokens, building data centers with high quality GPUs, hiring top talent in AI for model architecture and training techniques, and all these are currently spearheaded by frontier labs. So we do need a healthy balance of both open and closed models to encourage close labs

from price-gouging end-users and be pressured to keep innovating and open labs to continue help increasing AI adoption while keeping the cost manageable. And one company that sort of stands unique in this is Google, who doesn't really have to stretch and bend over backwards to compete. Google has custom ASIC like their TPUs, pop researchers, and the infrastructure and strong ecosystem to release their products. And Google doesn't really stand to lose as much as other front-year labs if the model layer truly ends up becoming a commoditization at the end. Since their mode is not in a single model that they release, but their control over the the entire stack that actually benefits from commoditization rather than being threatened by them.