This one is exciting! In the 1980’s the discovery of lithium cobalt oxide led to the development of lithium batteries which now power our mobile devices and a whole plethora of other electronic machines. Discovering new materials is incredibly rare and time consuming, usually as a result of trial and error or the computational screening of millions of potential candidates.
Microsoft Research has just released MatterGen,an AI tool that generates “materials with desired chemistry, mechanical, electronic, or magnetic properties, as well as combinations of different constraints”. This is a huge scientific development, but one that was largely ignored by mainstream media. The implications of this technology could be massive especially in terms of making use of more efficient energy – it could also be used to absorb CO2 from the atmosphere. It generates “proposed structures by adjusting the positions, elements, and periodic lattice from a random structure”.
Trained on 1.5 trillion DNA and RNA sequences, using a 7-billion-parameter transformer model (a machine learning model for processing sequences), designed to monitor pandemics, detect pathogens, and discover emerging health threats. This new model was developed by researchers at University of South Carolina, Prime Intellect, and the Nucleic Acid Observatory, and it analyses large collections of human wastewater samples in order to “establish the full distribution of genomic information present across the human microbiome”. This model will facilitate global pathogen detection and minimise pandemic risks in an era of rapid biological advancement.
The National Institutes of Health has granted $4 million to Case Western Reserve University, University Hospitals, and Houston Methodist to create an AI model for predicting heart failure and other cardiovascular risks using CT scans. The AI will analyse calcium-scoring CT scans along with clinical data to offer more precise risk predictions.
This initiative is looking to enhance cardiovascular disease prevention by identifying at-risk individuals early on, allowing for personalised treatments. The project aims to integrate AI into clinical practices, providing a cost-effective, non-invasive tool for assessing cardiovascular health.
Despite what will go down as a very bad month for the technology company in terms of their financial losses, the month started off with a bang. At the CES conference in Las Vegas, CEO Jensen Huang presented multiple game changing technological advances and a huge leap towards physical AI:
NVIDIA’s Cosmos is a world foundation model designed to understand and simulate the physical world. You can get a bit existential over-analysing this new development, because what Cosmos does is it essentially creates a digital replica of the real world. Jensen Huang even used the term “multiverse generation”, so every single possible physical outcome of a situation can be generated by this model’s system.
Why do we need this? Firstly, it’s genuinely amazing. Secondly, and more importantly, for physical AI to thrive there needs to be training data, which is difficult and expensive to create. So, instead of going out and putting robots (autonomous cars included) in different scenarios to gather data, this can now be generated and transferred into a robot’s ‘brain’. At the end of presenting Cosmos; Huang went on to say “the ChatGPT for robotics is just around the corner”.
You can now have your own AI supercomputer at home. One of these machines can run a 200 billion parameter model, which to us normal human beings doesn’t really mean all too much. However, to data scientists and AI engineers this is the holy grail. This accessibility represents a major advancement in making AI training more available to individuals and smaller organisations, which means it will aid in the advancement of AI at an even faster rate.
“A new world of AI computer graphics”. A graphics card like never before seen, it uses neural networks to generate over 90% of the pixels in a video game – in real time. Which means that we are essentially shifting towards video games being “created” right before our eyes. Ray tracing, which simulates light paths for realistic graphics, has now taken an enormous leap through AI, resulting in more beautiful and lifelike images.
I would recommend watching all of Jenson Huang’s presentation at CES, not only for a full run-down of NVIDIA’s releases but also insights into the current state of AI and what the future could potentially hold.
No monthly review is possible without mentioning OpenAI and while they too were shaken by DeepSeek, with Sam Altman calling it “legit invigorating to have a new competitor”, they have also had a busy month.
The much anticipated release of OpenAI’s venture into the agentic future is finally here. ChatGPT can now act as an AI agent and autonomously take control of your web browser. This means it can type, click, and scroll without any human interaction. Ask it to find you a dog-friendly restaurant in Bonnevoie, it will then scour the web and do so. Give it your salary slips and it will fill out your tax forms. Most tasks that you can do on your browser, it can do, and probably better than you...
Here’s another weirdly intimate, and awkward video of them releasing yet another product:
OpenAI also developed a language model for protein engineering, which is capable of improving the reprogramming of cells into stem cells. In collaboration with Retro Biosciences, OpenAI’s model, GPT-4b, suggested modifications to Yamanaka factors, proteins used in reprogramming cells, making them over 50 times more effective.
This is OpenAI’s first venture into biological data and represents a potential breakthrough in protein engineering. The model has produced better results than human scientists in many cases, once again showcasing AI’s potential to contribute to scientific discoveries. However, the model is still in a demonstration phase and not available for public use.
This collaboration, alongside Microsoft, will utilise NVIDIA’s Venado supercomputer at Los Alamos National Laboratory – which is insane in itself, considering it was the site where the atomic bomb was developed – benefiting researchers across multiple labs. The project focuses on accelerating scientific breakthroughs in fields like materials science, energy, cyber-security, and disease treatment. It also aims to enhance national security through improved threat detection and nuclear security. OpenAI’s commitment to AGI’s safe and beneficial use aligns with U.S. government goals. This partnership marks a significant step in leveraging AI to advance science, security, and technological leadership. The race to become the dominant superpower in the AI race and the first to reach AGI has well and truly heated up.
Sam Altman also wrote another blog post that reflects on OpenAI’s successes and more importantly that they are now shifting their attention to not Artifical General Intelligence but Artifical Super intelligence.
You surely saw the news, the introduction of a new Chinese Large Language Model that performs at a similar level to the leading LLM’s in the world but for a fraction of the price. This caused Wall Street’s tech stocks to take a huge hit. While OpenAI has received billions in investment, DeepSeek were able to produce their product at the mere price of $5.6 million.
How were they able to do this? Some people suspect that, despite the US blocking chips being exported to China, China have managed to acquire more chips that they were supposed to. CEO of Scale AI, Alexandr Wang, claims that DeepSeek has about 50,000 H100s, a chip that was not meant to be in China (news report excerpt on the Twitter link below).
The model is impressive and it will undoubtedly force its US competitors to make their products cheaper and better. The race to AGI has now taken a twist and everybody wants to be first.
Okay, if you’ve made it this far then you are definitely intrigued by the capabilities and future of AI, great! So, this bit is all a bit more technical (but potentially very impactful) and in order to understand this you need to understand what a transformer is.
All Large Language Models (LLMs) are based on transformers and they are AI models that process sequences of data, like text, by considering all parts of the sequence simultaneously using a mechanism called self-attention. This allows it to understand the relationships between different elements in the data (that was ChatGPT’s explanation, very good). A new research paper by Google Research proposes a new architecture for transformers (which were first developed by Google).
If you prompt an LLM with a longer, more complex prompt then it will take more time to understand the prompt and will likely perform significantly worse than with shorter prompts. Now – you can look at Titans like a development of transformers, whereby Google is trying to mimic the human-brain function in order to process more complex prompts. Humans have short-term-, long-term-, and meta-memory (memory that helps improve learning and adaptation), and the architecture they have works in a similar fashion. It will also enable LLMs to process information at a much faster rate.
If companies like OpenAI, Anthropic, and all the others, apply this new architecture to their LLMs then the results that follow may end up aiding in multiple new developments and discoveries.
Too many people are oblivious to the power of Artifical intelligence. We have existed as a civilisation for around 6,000 years and in that time period, not once has something existed that was more intelligent than human beings. We are currently in the process of doing this, creating a higher power, a tool that will unlock methods, inventions, and items that our brains cannot fathom as of yet.
Agents will likely enter the workforce within the next 6 months, then next year there is a chance of the creation of Artificial General Intelligence, the next frontier after that is Artificial Super intelligence. We will reach this status and humankind will change forever. The level of scientific discovery will increase exponentially and people need to start accepting this.
That is why you should probably stay up to date with all things Artificial Intelligence, and you can read the other monthly reviews here.
