➡ OpenAI is developing advanced AI systems, called superagents, that can handle complex tasks independently, similar to a PhD graduate. They’re also building their own robots for real-world use, aiming to produce over a million units. A new method called FAST has been introduced to train these robots more efficiently. Other AI advancements include Luma Labs’ generative video AI, Ray 2, and Mistral AI’s CodeStrull 25.01, which improves AI-driven coding.

Are AI and robotics on the brink of superintelligence? Maybe, as OpenAI is planning to unveil PhD-level superagents by the end of the month, and to start, these AI super systems will be capable of tackling highly complex problems with a level of expertise comparable to that of a doctoral graduate. In fact, OpenAI’s unveiling of these superagents is expected to occur during a meeting between CEO Sam Altman and US government officials in Washington on January 30, with these so-called PhD-level superagents representing a brand new frontier in AI for several reasons. Specifically, these agents are described as the next generation of highly capable systems that can process vast amounts of data, analyze complex scenarios, and autonomously deliver polished, goal-oriented solutions.

But unlike traditional AI models right now, which often require significant human intervention to interpret results or refine outputs, OpenAI’s upcoming superagents are instead designed to independently execute extremely complex end-to-end tasks. And on top of OpenAI’s upcoming superagents, the company has also revealed plans to build its own robots, complete with its own custom sensor suites and AI models developed in-house. As of now, this initiative is being spearheaded by OpenAI’s robotics team, which aims to create general purpose, adaptive, and versatile robots capable of operating in dynamic, real-world environments. Plus, job listings from the company further describe the ambition to blend cutting-edge hardware and software seamlessly, enabling robots to perform tasks with human-like intelligence.

Furthermore, OpenAI’s approach not only includes designing custom sensors and computational hardware, but they’re also considering diverse robotic form factors, ranging from traditional machines to humanoid designs. One job description even hints at the possibility of robots with limbs, suggesting a focus on creating systems that can interact with the physical world in a highly sophisticated manner. But OpenAI is investing in scaling its robotics efforts to an industrial level too, with another job listing indicating plans for full-scale production of its robots, with a target of manufacturing over a million units. This level of ambition underscores the company’s belief in the transformative potential of robotics, which is a field that has historically lagged behind software-based AI due to the overwhelming complexities surrounding the integration of intelligence with physical systems.

And to accelerate its development, OpenAI also plans to employ contract workers to test its robotic prototypes to ensure they are ready for real-world deployment, with this strategy reflecting a new focus on building robust, scalable systems that can operate in unpredictable environments. The company has even stated that it knows how to achieve artificial general intelligence, with CEO Sam Altman saying that AI agents may begin to join the workforce this year, predicting that these systems will materially change the output of companies. But another key is OpenAI’s integration approach to software and hardware development under one unified vision, because while competitors focus heavily on physical engineering and motion dynamics, OpenAI’s approach instead centers on creating a seamless synergy between cutting-edge AI models and custom robotics.

Importantly, this holistic strategy prioritizes adaptability and intelligence over form, enabling their robots to evolve and improve in real-world settings rather than being confined to pre-programmed behaviors or fixed-use cases. And to unlock the next frontier in training vision-language action models for these robots, physical intelligence researchers have just introduced a groundbreaking approach called FAST. Amazingly, this breakthrough addresses the long-standing challenge of reducing redundancy in robot action sequences to enable high-frequency task training. And to finally accomplish this, the FAST method leverages what’s known as the discrete cozine transform typically used in standards like JPEG to instead compress the robot’s actions before feeding them into the model for training.

Incredibly, this approach unlocks the potential for autoregressive VLA training on complex high-frequency tasks by eliminating unnecessary redundancy. Furthermore, researchers using FAST were also able to scale VLA training to the Pi0 level, achieving next-level results. For instance, the Pi0 model combined with FAST demonstrated a five-fold improvement in speed compared to diffusion-based Pi0 models, reducing overall training times from weeks to just days. And even more impressive is that this efficiency has even opened the door to solving intricate robot tasks using simple next-token prediction. But one of the most exciting breakthroughs came from testing the model on the DROID dataset.

With FAST, the trained VLAs exhibited zero-shot performance across diverse robotic environments, including labs at Berkeley, Stanford, and the University of Washington. Equipped with just a camera and a typed instruction, these robots successfully performed tasks in real-world settings, showcasing the versatility and robustness of the approach. And to accelerate adoption, the researchers have already released a pre-trained FAST tokenizer that’s trained on 1 million real robot action sequences and is now available on hugging face. Plus, the tokenizer has shown strong generalization across various robot types, enabling researchers to easily train VLAs for complex tasks.

And beyond robotics, FAST enables seamless interleaving of non-robotic data into VLA training, including web data, sub-goals, and video prediction data. And by treating all inputs as tokens, FAST simplifies the integration of multimodal data and pushes the boundaries of what VLAs can achieve. But generative AI has also had another breakthrough, as Luma Labs just announced the release of their latest game changer in generative video AI called Ray 2, which is now faster, more realistic, and boasts an even better grasp of real-world physics than its competitors, such as Runway’s Gen 3 or OpenAI’s Sora.

Additionally, Ray 2 features a standout ability in understanding and simulating interactions between diverse objects, including humans, animals, and vehicles. This capability adds unprecedented realism to the 10-second high-resolution clips it can generate from text or image prompts, with the model’s cinematic output including smooth motion, advanced cinematography, as well as dramatic visual flair. But unlike its predecessor Ray 1, Ray 2 was trained using Luma’s native multimodal architecture on a video-focused dataset and scaled to 10 times the computational power, with this training enabling it to produce ultra-realistic visuals with coherent motion and logical event sequences, which significantly increases the rate of usable video generations.

As of now, Ray 2 supports text-to-video generation and is available through Amazon AWS Bedrock and is also integrated into the Dream Machine platform, with additional features like image-to-video, video-to-video, and editing being expected to release soon. And finally, Mistral AI just released CodeStrull 25.01, which is set to transform AI-driven coding after achieving top-ranking LMSYS benchmarks, supporting over 80 programming languages and excelling in debugging test generation and fill-in-the-middle tasks. Plus, CodeStrull is powered by an optimized transformer architecture. It processes code twice as fast as its predecessor to ensure ultra-low latency performance for time-sensitive work.

Anyways, like and subscribe and check out these bonus clips. Thanks for watching!
[tr:trw].