Key ideas in AI: reinforcement learning
This Week in AI
Following the shockwave that the AI model DeepSeek sent through the global AI scene, particularly US stock markets who dominate AI research and development, there have been several developments in both frontier AI research as well as ethics and novel developments.
OpenAI unveils Deep Research
On February 2nd OpenAI unveiled the latest tool in their AI suite, called Deep Research. Deep Research is a tool that scours the internet to research a question provided by the user. It can spend extended time looking through many online sources, performing mathematical calculations and running code to provide the user with a thoroughly researched and formatted response to their question, often in the form of a report or briefing. In their livestream, they demonstrated several business related use cases, such as project scope research, market assessment, scientific research, and more personal uses such as researching products and reviews.
They also provided several benchmarks, the most notable of which is its score in Humanity’s Last Exam, designed to test highly specific and specialised doctorate-level questions across all fields of research. OpenAI’s previous top model, o3, scored 13% on its own. However, the Deep Research tool was able to achieve a score of 26.6% which is more than double the previous.
This has huge implications particularly for business because it means many sophisticated tasks like research and recommendations can begin to be automated. It has the potential to save countless valuable hours. Further, since it provides its sources as information, its initial research can be quickly supported and verified to ensure that its response is factual. This eliminates a major argument many use for dismissing AI products, that they make mistakes as it is difficult to trust factually, since these responses now contain extensive verifiable research, and if they do make a mistake they can be corrected.
https://openai.com/index/introducing-deep-research/
Synthesising new materials with MatterGen
Microsoft Research released a paper in January explaining their new architecture for generating inorganic materials fast. The model works using a diffusion architecture, which means taking a completely random sample and morphing it into a candidate material, similar to how image generators work. Trained on 608,000 stable materials from the Materials Project, up to 30% of its generated structures are stable, novel and unique. This is extremely important because it means that researchers can more reliably test these materials in practice and consider them for use in other applications. Automation of stable and novel materials is an extremely valuable technology because it may mean a dramatic increase in breakthroughs for batteries, devices, engineering and medicine.
New Vatican document provides comprehensive AI warnings
On January 28th, the Vatican City released Antiqua et Nova, a document providing ethical guidelines and warnings to the use of AI within society. This follows a dramatic increase in AI spending by the tech giants like OpenAI, in addition to the introduction of DeepSeek, a Chinese competitor to Western state-of-the-art models.
The document details several key risks, such as the ethical concerns of removing human moral judgement from warfare decision-making, the threat of AI to privacy and individual sanctity, threat to education and critical thinking, and the major risk to the environment that AI poses through energy consumption, despite its potential to help develop climate solutions.
The essay promotes understanding that AI is on the trajectory to being more than an automation tool, and calls for new frameworks, ethical oversight and regulation for new AI tools to ensure that it truly aligns with human authenticity, morals and dignity.
