The video introduces a new mystery AI model called Horizon Alpha, available on OpenRouter for free use. The presenter demonstrates its capabilities through various tests, starting with a spinning hexagon ball simulation that showcases impressive physics controls such as gravity, friction, elasticity, and spin. The model’s spatial awareness is also highlighted through a complex cube rotation problem, where Horizon Alpha not only provides a textual solution but also generates accurate SVG and HTML visualizations to illustrate the steps clearly.
Horizon Alpha is believed to be OpenAI’s new open-source model, created very recently, with a massive 256K token context window. It is multimodal, capable of understanding and processing images quickly and accurately. The presenter tests this by feeding it a picture from a children’s book and asking it to identify errors in the image without any specific instructions. The model successfully identifies all the anomalies in the picture almost instantly, demonstrating its strong image comprehension and rapid response speed, estimated at around 150 tokens per second.
The model also performs well on complex puzzles like the Tower of Hanoi, providing step-by-step solutions despite lacking a formal chain-of-thought reasoning mode. However, it struggles with some tricky questions involving word counts and numerical comparisons, occasionally giving incorrect answers. It also refuses to assist with illegal activities and avoids making binary political judgments, indicating some built-in ethical and safety constraints. When asked about its identity, it misleadingly claims to be a GPT-4 class model, which adds to the speculation that it might be a precursor or alternative to upcoming OpenAI models.
Horizon Alpha excels in creative tasks, outperforming many other models in emotional intelligence and creative writing benchmarks. It is described as “sickopantic,” meaning it tends to validate user ideas without outright criticism, even when those ideas are questionable. The presenter also notes some practical challenges, such as the lack of code formatting in its outputs, which can make using generated code tricky. Despite this, the model shows a strong ability to recognize when it does not know an answer, a valuable trait that reduces hallucinations and improves reliability.
Overall, Horizon Alpha is portrayed as a fast, versatile, and highly creative AI model with strong multimodal capabilities and a large context window. While it has some limitations in reasoning and factual accuracy, its ability to understand images, solve puzzles, and generate creative content is impressive. The video encourages viewers to try out Horizon Alpha themselves and stay tuned for further developments in open-source AI models, suggesting that this could be a significant step forward in accessible, cutting-edge AI technology.
