Open AI Trojan Horse or Gift

DeepSeek, a little-known Hong Kong AI startup, released a new model series, DeepSeek-R-1 on 20th January 2025. The new Open-source large reasoning model has reshuffled the market. It stirred up fears of a Trojan horse and hastened no shortage of hand-wringing and existentialism among Western-block techies.

2025 is the confluence of Open AI Vs. Closed AI models in the sprint to dominate as the global AI provider. DeepSeek-R1 offers the freedom to modify and build whatever you want. Just imagine, if you are a chocolatier, it is like having unhindered and unlimited access to all the single-estate fine Trinitario cocoa powder you need to make whatever candies, cakes, and chocolates you can imagine for free.

The capital risks are immense given the record spending of huge sums on new AI infrastructure that depreciates swiftly due to rapid chip, software, and hardware advancement. The race to dominate as the global AI provider is on.

The alternative may turn out to be a multiplicity of models each with a smaller market share. Which model of the future will win in the end is still unclear. However, Open-source AI models are outstripping proprietary ones using open research.

According to the Open Source Initiative, Open source AI offers the freedom to use a model, inspect its components, modify the system including changes to inputs, and share the system with others to use with or without modifications. Closed Source AI, on the other hand, keeps the source code and underlying algorithms private, and they cannot be modified or built upon.

The antibodies argue that DeepSeek-R-1 is a Trojan horse, but the evangelists see it as a gift to the world. To power Facebook’s new version of its open-source AI model Llama 4, Meta is building a 2GW+ datacentre in Manhattan. Meta will invest US$60-65B in capex in 2025 while growing its AI teams. This unlocking of innovation, digital infrastructure, investment, and talent attraction will drive its core products and business in the years ahead.

Present generative AI and large language models (LLMs) are already fading. A whole new paradigm of AI is emerging. Most of the familiar components will not be included in the central architecture of future systems.

While dazzling, present models inhibit truly intelligent behaviour, and lack understanding of the physical world, lack complex planning capabilities, lack common sense, lack reasoning, and do not have persistent memory. The next wave of AI will interface with machines and will have memory, intuition, reasoning capabilities, and common sense – traits far beyond what we see now which is largely pattern recognition.

In his prison notebooks, Antonio Gramsci (1971) defines “common sense” as a chaotic aggregate of disparate conceptions of truths that are historically and socially situated. How AI models assemble this aggregate of tacit knowledge across cultures and social strata remains a challenge for those building taxonomies of ethical guidelines for AI, and those designing AI projects for social good. Contemporary LLMs are great at language but not thinking. This provokes the question – What do generative AI models know about the real world?

AI is about pattern-matching computation using materials scraped from books, libraries, and the internet. If you enter the prompt: “If I drop a feather and a stone simultaneously from the top of the Eric Williams Financial Complex will they reach the ground at the same time under the force of gravity? As the output appears it is clear that the works of Galileo and Newton shape the answers generated. But generative AI has never observed a feather or a stone falling freely from any height. It produces responses using data but not first-hand experiences. This looks pedestrian. However, if AI is going to be connected to robots, appliances, and tools in homes, offices, and factories it is important that it can do so safely.

AIs trained on text-based data may not integrate well into a mass event like the Trinidad Carnival or in a factory alongside human workers.  Such mechanizations must be compatible with humans and operate in various environments without being harmful or endangering. Physical AI is one strand of three foundation AI projects evolving alongside Agentic AI, and Generative AI.

We can describe agentic AI in one word: proactive. In the coming era, humans will interact with AI-powered agents to plan overseas work missions. They do not rely on human prompts. Rather, they are set to optimize. Unlike generative AI, they can independently execute complex sequences of activities such as searching databases, triggering workflows to complete activities, or proactively managing complex IT systems to pre-empt service disruptions.

AI-powered supply chain specialists will optimize inventories in response to fluctuations in real-time demand. To achieve this level of autonomous decision-making, agentic AI uses a complex ensemble of machine learning, natural language processing, and automation technologies.

These models will harness the creative abilities of generative AI Models but focus more on making decisions rather than on creating content. The future of Latin America and the Caribbean is now in the hands of developers who may wish to consider the benefits of contributing to – and benefitting from – a LAC Open AI ecology with clear disclosures about datasets and safety mechanisms.