As Rio de Janeiro's ambitiously unveiled large-scale AI model turns out to be a patchwork of existing models rather than an original development, the practical challenges of building true 'local AI' have come to light.
Introduction: The Truth Hidden Behind the Glamorous Debut
Imagine this: a world-famous magician declares that after years of painstaking training, he has invented a unique levitation trick. Under the bright lights, a huge crowd is cheering and giving a standing ovation. But then, a boy who happened to peek backstage shouts, “Hey! He just tied two ropes together and hung them from the ceiling!”
This intriguing story recently became a reality in the AI industry, the battlefield for the world’s most cutting-edge technology. The city government of Rio de Janeiro, Brazil, proudly unveiled a large-scale AI model that they claimed was developed from scratch. However, suspicions that it was actually a “patchwork” result created by cleverly mixing technologies others had already built have turned out to be true.
Today, many countries and cities are struggling to have their own AI to avoid being dependent on foreign tech giants. So, what exactly happened in Rio de Janeiro? Why did people become convinced that this technology was fake, and what message does this incident send for our future? Let’s dig into the full story as if a smart friend were telling it over a warm cup of coffee.
Why It Matters: The Dream of ‘National AI’ and 397 Billion Stars
The voice assistants on the smartphones we use every day or services like ChatGPT operate based on massive “Large Language Models (LLM),” which are AI systems that understand human language and generate sentences by learning from vast amounts of text. Recently, countries around the world have been staking their futures on developing independent AI under names like “Local AI” or “Sovereign AI” to protect their data sovereignty and unique cultural characteristics.
Last week, IplanRIO, Rio de Janeiro’s IT agency, announced a historic achievement. They proudly released a giant model named ‘Rio-3.5-Open-397B’ on Hugging Face, a library-like platform where AI developers worldwide share their code Source: Rio de Janeiro’s ‘Homegrown’ AI Was Someone Else’s Model Wit….
The number ‘397B’ at the end of the name is what matters. It means the AI has 397 billion parameters. To put it simply, parameters are like the “dials” in a photo app that finely adjust color or brightness. Inside an AI model, these dials turn constantly to remember vast amounts of knowledge and make decisions. The number 397 billion is an awe-inspiring scale, comparable to the number of stars in our entire galaxy. This weight class means it stands shoulder-to-shoulder with the cutting-edge models built by the world’s top big tech companies like Google or Microsoft at an astronomical cost Source: Rio de Janeiro’s ‘Homegrown’ AI Was Someone Else’s Model Wit….
If a city government agency had truly developed such an enormous AI completely “on its own,” it would have been a monumental achievement in the history of human technology. However, this great celebration was immediately engulfed in fatal suspicions.
Breaking It Down: The Crucial Difference Between ‘Original Development’ and ‘Model Merging’
To get to the heart of this incident, one must understand the fundamental difference between “independently training” an AI and simply “merging” it.
Think of it like introducing a special curry with a completely new flavor to the world. ‘Original Development (Self-training)’ is a grueling journey of growing your own potatoes and onions in the field, importing spices from the rugged lands of India, and testing the blending ratio thousands of times to create your own perfect curry powder. it requires an immense amount of time, a huge amount of money, and the sweat of many experts. In the AI world, this is a solitary and harsh process of running thousands of ultra-expensive computers (GPUs) day and night for months, teaching it by spoon-feeding vast amounts of data from the beginning.
On the other hand, ‘Model Merging’ is a completely different story. It’s like buying “Brand A’s solid curry” and “Brand B’s spicy curry,” which are already best-sellers at the local big-box store, putting them into a large pot together and boiling them. A reasonably plausible and tasty result might come out as the two curries mix. But what if you put this mixed dish in front of the public and advertised, “This is an innovative new curry product that our city government has been researching for years and developed independently from the bottom up!”? That would be a clear act of deception.
Unfortunately, the “homegrown” AI model announced by Rio de Janeiro was not an independently trained system built on a completely new foundation Source: RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan….
Current Situation: GitHub Detectives and the Feeble Excuse
Surprisingly, it wasn’t a major media outlet or a government auditing agency that first caught this giant technical bluff. It was ordinary developers on ‘GitHub,’ a software development platform where tens of millions of programmers worldwide are active. Pandora’s box of truth was opened when someone asked a sharp question on the ‘Issues’ board, GitHub’s space for reporting errors Source: Cosmic Rundown: Billion Dollar Essays, Rio’s LLM Drama, Context Window Limits.
As a result of the community’s analysis, the truth was fully revealed: this “homegrown model” was actually a sophisticated merge of the ‘Nex-AGI’ and ‘Qwen3’ models, which were already publicly available for free on the internet Source: Rio LLM Exposed: Major Model Merge, Not Original AI, Source: RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan….
When they examined the internal brain structure of the AI model, which consists of computer code and mathematical values, not a single piece of evidence was found that it had studied from scratch. Instead, only clear evidence of physically mixing others’ models poured out Source: Rio LLM Exposed: Major Model Merge, Not Original AI. GitHub, a playground for developers, was used like a whistleblowing forum or a sharp investigative blog Source: Hacker News 20 on X: “Rio de Janeiro’s "homegrown" LLM appears to be a merge of an existing model https://t.co/G1dBFWiQcO (https://t.co/Uht1ZUEPrL)" / X, Source: RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan….
As the voices of criticism spread like wildfire, IplanRIO, which led the development, quickly released an explanation. They apologized, saying, “We made a mistake in uploading the wrong file during the process of uploading a previous version. We should have uploaded the final ‘Distilled model,’ but we mistakenly uploaded the ‘Base merged version’ that was an intermediate stage of the work” Source: RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan….
What is ‘Distillation’ in this context? Think of extracting a single cup of very dark and fragrant espresso by strongly pressing a huge amount of coffee beans in a large coffee machine. In the field of AI, distillation is a high-level technology that extracts only the core knowledge of a genius AI (teacher model) that is too bulky to handle and compresses it into a lightweight AI (student model) that runs quickly on small devices like smartphones.
In other words, the city government’s excuse is: “It’s true that we mixed other models and boiled them in a pot (Merge), but what we originally intended to release to the public was the final espresso (distilled model) that beautifully compressed the result.” However, even if we generously assume it was a simple upload mistake, the essence remains unchanged: the backbone of the public AI, into which citizens’ taxes were invested, is ultimately “something made by mixing others’ models.”
Outlook: The Era of Repackaging—How to Spot Real Innovation
This “Rio de Janeiro Scandal” has made many local governments and companies around the world, who are trying to build their own independent AI ecosystems, realize the heavy wall of reality.
A famous tech expert active on X (formerly Twitter) watched this farce and pointedly noted: “Rio de Janeiro’s supposedly homegrown model? It eventually turned out to be a patchwork of existing models. The hype around ‘local AI’ keeps running into the same massive wall. Building something completely new that has never existed in the world is incredibly harsh and difficult, while repackaging what already exists is much easier” Source: Anto Patrex on X: “Rio de Janeiro’s supposedly homegrown LLM? Turns out it’s a merge of existing models. The hype around ‘local AI’ keeps running into the same wall: actually building something novel is hard. Repackaging is easier." / X.
As AI technology becomes deeply integrated into our lives, many institutions will compete to blow their trumpets and shout, “We have finally completed our own independent AI!” But we must now look into those glamorous packages with a very careful and critical eye. Because even if it looks like a great invention with hundreds of billions of parameters on the outside, the inside might just be a sly mixture of free open-source models that someone else sweated to create.
AI’s Perspective: Transparency is the Ultimate Technical Strength
We are in an era where the massive scale of AI models or the astronomical number of parameters represents the technical pride of a city or a nation. We cannot unconditionally criticize Rio de Janeiro’s desire to secure its own technical capabilities. Smartly combining and utilizing existing open-source technologies to overcome the limitations of massive capital is also a natural and efficient trend in modern software development.
However, we must not forget that in the transparent open-source ecosystem watched by many genius developers with sharp eyes, clumsy repackaging and exaggerated advertising only serve to seriously undermine trust. True technical independence and securing sovereignty do not come from mixing seasonings according to recipes floating on the internet and giving them grand names. The flower of real innovation can only bloom through the patience of honestly collecting high-quality local data even in a barren environment, transparently sharing one’s own limitations, and moving forward step by step. Rio de Janeiro’s brief and hollow day of glory has left a painful but valuable lesson for all of us living in the AI era.
References
- Rio de Janeiro’s ‘Homegrown’ AI Was Someone Else’s Model Wit…
- RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan… (Deep Intellica)
- Rio LLM Exposed: Major Model Merge, Not Original AI
- Cosmic Rundown: Billion Dollar Essays, Rio’s LLM Drama, Context Window Limits
- Hacker News 20 on X: “Rio de Janeiro’s "homegrown" LLM appears to be a merge of an existing model https://t.co/G1dBFWiQcO (https://t.co/Uht1ZUEPrL)" / X
- RiodeJaneiro’s”homegrown”LLMappearstobeamergeofan… (Hacker News Discussion)
- Anto Patrex on X: “Rio de Janeiro’s supposedly homegrown LLM? Turns out it’s a merge of existing models. The hype around ‘local AI’ keeps running into the same wall: actually building something novel is hard. Repackaging is easier." / X
- It was found to be the world's first model to perfectly understand human emotions.
- It was a patchwork model created by simply mixing existing AI models.
- It was praised for disclosing the entire development process transparently.
- The original file was manipulated due to a hack.
- It was an unavoidable choice due to a lack of data.
- They accidentally uploaded a base merged version instead of the final version (distilled model).
- Building a brand-new AI from scratch that has never existed before is extremely difficult.
- There is a lack of electricity required for AI development.
- Legal regulations are too strong.