AI news as at February 26, 2025

[smfadmin (OP)]

India to kickstart its AI program

Hot on the heels of DeepSeek, India has announced plans to develop its own AI model, aiming to compete with industry giants in the United States and China. The country’s first indigenous AI foundational model will be ready within 10 months and is expected to be cost-effective — while global AI models charge upwards of US$2.50 per hour for computation, India’s AI system is projected to cost less than $1.15 per hour after a 40% government subsidy. The country currently ranks fourth in AI innovation, according to Stanford University’s Global AI Index, though it faces limits on data, skilled professionals and clear regulatory frameworks.


AI maths model rises on the podium

AI problem-solver AlphaGeometry 2 has reached the lev
el of a gold medallist at the International Mathematical Olympiad, a prestigious competition that sets tough maths problems for gifted high-school students. That’s an upgrade from the Google DeepMind tool’s silver-level performance last year. It’s not your typical AI system (ever try asking ChatGPT to do simple arithmetic?). AlphaGeometry 2 contains both a specialized language model and a ‘neuro-symbolic’ system — one that does not train by learning from data, but that has abstract reasoning coded in by humans.


Scientists use AI to design life-like enzymes

Researchers have leveraged AI to design brand-new enzymes that can go through multi-step reactions, a key feature of natural enzymes.

They began with a tool called RFdiffusion which generates new enzyme structures from scratch. They then created a deep neural network, called PLACER, to refine these designs by modelling the locations of atoms in the enzyme and the molecules it binds to throughout each step of the reaction.

The new enzymes were 60,000 times better at speeding up the reaction than ones previously designed to work in a similar way.

The latest model is based on 128,000 genomes, including those of humans and other animals, plants and other eukaryotic organisms.

These genomes encompass a total of 9.3 trillion DNA letters.

The biggest ever AI model for biology just dropped, trained on 128,000 genomes spanning the tree of life, from humans to single-celled bacteria and archaea. It can write whole chromosomes and small genomes from scratch. It can also make sense of existing DNA, including hard-to-interpret ‘non-coding’ gene variants that are linked to disease.


Google’s co-scientist joins the AI agent race

Another day, another potentially transformative AI research agent. Google’s new co-scientist tool pulls from its Gemini large language models to search academic literature, connect to databases, and even access Google’s AlphaFold system. So what do the scientists who tried it out have to say? Cell biologist Steven O’Reilly tested it to identify methods for treating liver fibrosis, and says, “there is nothing new here”. But microbiologist José Penadés was shocked when the tool proposed the very same hypothesis contained in his unpublished discovery. “I sent an email to Google saying, you have access to my computer. Is that right? Because otherwise I can’t believe what I’m reading here,” he said. (The AI had, however, been fed a paper containing an earlier, related, hypothesis from his group.)

AI for video games levels up

When a Microsoft team asked video-game developers what they wanted in AI tools, they said engines that can generate gameplay sequences consistent with the rules and physics of the game.

Microsoft’s new WHAM system does just that, after training on video frames and controller inputs from 500,000 anonymized game sessions of the multiplayer game Bleeding Edge. That’s equal to seven years of continuous play.

Developers can drag and drop imagery right into the WHAM prototype and watch it appear in the generated game world.


Your guide to the latest AI models

With new impressive AI tools dropping constantly, it can be hard to keep up. In this guide, Nature spoke to researchers about which tools to use when (with many adding the perennial caveat that all models make mistakes, so use their outputs with care).

● o3-mini (the reasoner):

OpenAI’s speedy model relies on a ‘chain of thought’ process to simulate human reasoning. That makes it good at technical tasks, such as solving coding issues, reformatting data, or picking apart unfamiliar concepts in a new mathematical proof.

● DeepSeek (the all-rounder):

DeepSeek-R1 is available through an API at a fraction of the cost of its competitors and is open weight, meaning that although its training data have not been released, anyone can download the underlying model and tailor it to their specific research project. Its fortes are maths problems and writing code, though it’s also good at tasks such as generating hypotheses.

● Llama (the workhorse):

Llama has long been a go-to LLM for the research community. Researchers have built on Llama models to make LLMs that predict materials’ crystal structure, as well as to simulate the outputs of a quantum computer, as it’s well-suited for adapting to the specialized language of scientific projects.

● Claude (the coder):

In Silicon Valley, many people swear by Claude 3.5 Sonnet for writing code. Its models also get plaudits for their writing style, too.

● Olmo (the really open one):

Researchers who want to understand what’s going on under the hood of an LLM will need something even more transparent than the models offered by Llama and DeepSeek. The top-performing model that comes with the algorithm’s training data plus the code used to train the models is OLMo 2.

-- Quote of the day --

“The AI revolution extends far beyond structure prediction, with diverse machine learning models now driving various molecular design tasks and significantly elevating the drug development process.”

Chemist Qi Hao predicts that, with proper interdisciplinary collaboration and well-trained datasets, AI has the potential to revolutionize drug development.

That's all, folks -- Chip !