In this video, Poolside co-CEO and co-founder Eiso Kant explains the launch of the company's latest openweight model, the Laguna S. He talks in depth about how he was inspired to enter the language model field by Andrej Karpathy's blog post in 2015, why the AGI Lab decided to release an open source model, and the question of ownership of intelligence. Kant cautions against the concentration of intelligence in the hands of a few companies and emphasizes that open source is critical to creating a future in which more companies can develop and utilize intelligence.

1. Concerns about intelligence monopolies and the importance of open source

Poolside's co-CEO Iso Kant shares his thoughts on the recent changes in San Francisco and the growth of the AI industry, saying it all feels like science fiction. Especially now in 2026, he was concerned about dystopian scenarios when imagining the future in 2035. This is a warning about a situation where only a few companies build artificial general intelligence (AGI) and monopolize everything that is economically valuable and scientifically interesting. He says:

"It's like... we're going to look back in 2035 and say, 'It was an oligopoly of two or three companies. They built artificial intelligence, they got to AGI, everything that was economically valuable and everything that was scientifically interesting ran on top of these companies.'"

Kant said that intelligence would become the most important commodity after energy and that he felt great discomfort about having to purchase this intelligence from a small number of companies. Accordingly, Poolside explains that it has begun to open the open source model to prevent this intelligence monopoly phenomenon and create a world where numerous companies can develop intelligence.

2. The LLM journey started from Karpati's writings 🚀

Isokant recalls that he "nerd sniped" on Andrei Karpati's 2015 article "The Unreasonable Effectiveness of Recurrent Neural Nets." This article instilled in him a strong belief that neural networks can generalize everything and that language is the key to improving intelligence.

"I think I was nerd-sniped by Andrei Karpati's 2015 article 'The Irrational Efficiency of Recurrent Neural Networks'. I still don't know why it captivated me, but it instilled in me a strong belief that neural networks can generalize everything and that language is the key to improving intelligence."

Within a few months, he shifted his company's focus to building code-based language models. Although no one was paying attention to this field at the time between 2015 and 2019, he and about 40 team members invested $12 million to modify NVIDIA gaming GPUs to train RNN, LSTM, and transformer models to develop code prediction technology.

He remembers that moment and says it felt like "I believed in an idea that no one else believed in." However, I have deep respect for companies like Google and OpenAI, who confidently scale despite small signals. After leaving the field for a while at the end of 2019, he returned with the launch of ChatGPT and is impressed by the amazing rate of progress in the field over the past few years.

3. Response from software engineers and acceleration of AI development

Kant says some software engineers still tend to be "denial" about their ability to write AI code. But he understands this reaction and explains that it's not easy to see AI outperforming something that has long been part of his identity.

"Some software engineers are almost in denial, saying, 'Oh, AI isn't that good at writing code.' But I think their opinions will change more and more."

Nonetheless, he emphasizes that software developers have a great ability to quickly adopt new technologies and will recognize that LLM brings a huge speedup**. The intelligence of current AI models has a "very jagged edge," meaning they excel in certain areas but are limited in others, but the pace of progress is exponential compared to what it was 6 or 12 months ago, he says.

This rapid development is a powerful incentive for companies like Poolside to focus on the "model" itself and ignore everything else. Kant emphasizes that to win in the current AI industry, you need to be extremely focused on "just one thing"**.

4. Laguna S born in 8 weeks: Amazing efficiency and performance 🚀

ISOKANT is very proud of its latest open source poolside model, the Laguna S. This model is the result of Poolside's "Model Factory", which explains that it is the result of a three-year obsession with the industrialization process of model building, which takes data as input and outputs an evaluable model.

• Remarkable speed of development: Laguna S went from start to finish in 8 weeks, with 4 weeks of pre-training, 3 weeks of post-processing, and 1 week of launch preparation. This is an example of how effectively Poolside has automated the model building process.
• Scale and Performance: Laguna S is a mixed expert (MoE) model at 118 billion parameters scale, with 8 billion parameters active. Kant says the model is "not a very large model, but it's just as powerful as much larger models," and especially highlights its outstanding performance on long-term agent coding benchmarks.
• Unexpected results: The Laguna S performed much better than expected. Kant says this model outperforms other open models of the same or twice the size, reminding us how much more intelligence can be squeezed out of bits**.
• "America's first open weight competitor": Kant says the Laguna S is one of the first open weight models released in the West and is proud of the fact that it outperforms other models in its weight class. However, due to the rapid pace of development in this field, this advantage may change within a few weeks, he added.

5. Laguna S's innovative architecture and advantages

Isokant explains the architecture of Laguna S in more detail. The Laguna S is based on the architecture of its predecessor, the Laguna XS, and belongs to Poolside's "2.1 generation architecture".

• Hybrid attention mechanism: This model is optimized for inference by combining sliding window attention and global attention at a 3:1 ratio. This is an important architectural decision that balances performance and efficiency.
• Sparse MoE: The Sparse MoE structure, in which only 8 billion parameters out of 118 billion are active, maximizes model efficiency.
• Context Length Extension: Through techniques such as RoPE scaling and context length extension, Poolside has focused on helping models better remember information from early stages when performing long-term tasks.
• 1 Million Token Context: One of the most impressive features of Laguna S is its 1 million token context length. This is a very useful feature, especially for long-term tasks.

Kant explains that most companies in the open weight model space are using similar architectures (sparse MoE, improved attention mechanisms, etc.), and emphasizes that Poolside is focused on maximizing the intelligence of the model through these technologies.

6. America's open source research institute, is it special?

Kant expresses some confusion about calling Poolside "America's open source lab."

"The American Open Source Lab is shocking to me. It shouldn't be. This is it, right? I don't know how we came to live in a world where the American Open Source Lab is suddenly special. Again, when we talk about the dystopian science fiction we're reading, the fact that we've come up with a capable model and people are surprised that we're an American open source company is something that shouldn't have happened."

He argues that originally a lot of open source and various foundation model companies should have flourished in the US. The current situation presents a reality that amplifies concerns about a dystopian future in which intelligence development is monopolized by a few companies. Poolside focuses on building the most capable model under these circumstances, and believes that this is ultimately the way to win the competition. He says the real winner will be the company that offers the most capable model at the most cost effective.

7. Economics of the open weight model: commoditization of intelligence

Kant analyzes that the open source model is currently facing a significant opportunity. As the evolution of intelligent models increasingly enables models to perform knowledge labor, which accounts for 25% of global GDP, we predict that as models become more intelligent, they will eventually reach the limit of economically valuable work.

He explains by imagining a graph of model size (Y-axis) versus return on investment per token (X-axis). After all, no one is going to pay the most expensive token for the most expensive model, and for every task there will be an optimal model size and cost point.

"Given that no one is going to use the most expensive token for the most expensive, largest model, this means that there is a sweet spot for everything we do: a sweet spot for the search queries we ask, a sweet spot for the piece of code we want to write, a sweet spot for the progress we want to make in scientific research."

This means that eventually the foundation model will be commoditized. As with the iPhone example, the argument is that AI models will gradually develop into smaller and more efficient forms, similar to how small smartphones can now handle tasks that could have been done by supercomputers the size of huge buildings in the past.

Kant believes that the Laguna S's outstanding performance despite its small size is a good example of this commercialization trend. He predicts that in the future, open source models will commoditize more intelligence, making tasks that now require huge data centers possible with much smaller models.

However, he points out that for open source to be successful, two things must be solved. First, the business model must be clear. A huge amount of capital is invested in model training, but the profit model is unclear because the openweight model is virtually free. Second, an open source ecosystem must be established in the United States and Western countries.

8. Ownership of Intelligence: 12-18 Month Opportunity

Kant emphasizes that the question "Who will build intelligence" will be the most important question in the coming years. We are at a "crossroads", and we must decide whether we will move towards a world where a few companies provide all the intelligence, or whether we will move towards a world where anyone with the talent, will and capital can build intelligence.

"We're at a crossroads. Are we going to get to a world where everyone in the world has three or four companies providing all the intelligence, three or four biases and three or four shutoff switches? Or are we going to get to a world where anyone with the talent and the will and the capital can build it?"

He warns that the "window" for these decisions is very short, around 12 to 18 months. This is because of the influence the model itself has on model development, namely recursive self-improvement (RSI). 90% of model building work is engineering (writing code, analyzing data), and the increased capabilities of AI models are significantly accelerating the speed of this work.

The time it takes for researchers to translate their ideas into real models is being drastically reduced, and soon models will autonomously participate in the model improvement cycle. From this point on, "compute" becomes the key element.

Researchers and engineers have a big role to play today, but that won't be the case in 12 to 18 months. Kant warns that if we don't get more people to build on the foundation model by then, we will live in a world where we can buy intelligence from a few companies and not be able to use that intelligence to build intelligence again.

What surprises him most, especially in 2026, is that foundation model companies are starting to say "You can't use what I create to improve something that can compete with me." He says this is a free market right, but it is also a shocking reality.

Therefore, he appeals to those who want to start a foundation model company to "call to arms"**, emphasizing that now is a critical time to create meaningful change.

9. Amazing abilities of Laguna S: Erdös problems and using the debugger

Isokant shares two of the most surprising experiences he had while using the Laguna S.

1. Erdös #397 Problem Solved: OpenAI presented Laguna S with Erdös #397 Problem, which was solved at the end of last year. The model solved the problem in 30 minutes even though the training data did not contain any data for that problem.

   "I gave this model the Erdös #397 problem that OpenAI had solved at the end of last year, and I knew it was within the deadline range so the data wouldn't be in it. But it solved it in 30 minutes."

   In particular, I was very impressed by the reasoning process in which the model explored the wrong path for 20 minutes, then corrected itself and found the correct path to solve the problem. He confesses that he did not expect that this small model would be able to solve such a complex problem.

2. Transcript extraction using debugger: Laguna S was tasked with extracting the encrypted transcript stored in Kant's laptop. The model found an encrypted database, but recognized that it could not be decrypted because it did not have the password. The model then demonstrated its remarkable ability to attach a debugger on its own, run the application on a Mac, and extract the actual transcript from system memory.

   "The model, and Pool's agent, worked for 30 minutes. I found an encrypted database on my laptop. I realized I couldn't decrypt it because I didn't have the password. Then I attached a debugger. I ran the application on my Mac, connected the debugger, and was able to extract the actual transcript from my system memory."

As a software developer, Kant says he was deeply impressed by this persistence and problem-solving ability. Of course, he would look at the model perfectly like a child, but the capabilities of the Laguna S made him question how much more he could get out of the model. He concluded that in the next few years, tasks that now require large models will become possible with much smaller models, a hopeful sign for a world where intelligence is becoming more commoditized and open source.

Conclusion

Isokant's story, along with the incredible pace of AI technology development, raises fundamental questions about whether the future of intelligence will be dominated by a few large companies, or whether it will take a decentralized form with more players participating. Poolside is accelerating the commercialization of intelligence through open weight models such as Laguna S, and dreams of a future where more people can participate in AI innovation. This can be said to be a journey beyond simple technological development toward a larger vision of democratization of intelligence.