The framework will also outline when severe risks should be considered "unacceptable" unless they can be fully mitigated, and what measures the companies will take to ensure risk thresholds are not exceeded. In the most extreme cases, if mitigating measures cannot keep the risk below the threshold, these companies also commit to "stopping the development or deployment of models and systems." In defining risk thresholds, they will consider the views of credible participants, including their own governments, and plan to publish by the 2025 French AI Action Summit.

Scientific Event

Scientists collectively question the openness of AlphaFold 3, and Nature responds on May 8, local time, Nature online published a collaborative study by Google's Deepmind team and the AI drug development team Isomorphic Labs, showcasing the new protein prediction model — AlphaFold 3. However, this release has led to many scientists questioning and criticizing the AlphaFold team and Nature. The issue is that when AlphaFold 2 was released, all researchers had access to the complete underlying code. But the "pseudo-code" that came with the AlphaFold 3 study only described the functions and working principles of the code. Many researchers expressed their disappointment with the limited open-source availability of the code.

On May 14, several researchers jointly published an open letter to Nature, pointing out that the open-source code and models of AlphaFold 2, through continuous expansion and adjustment, had made subsequent research and benchmark testing possible, but the code restrictions of AlphaFold 3 also limited subsequent verification and reproducibility work, which does not align with the principles of scientific progress; and Nature's choice to publish such a study lacking open-source data does not facilitate peer review. The open letter eventually collected signatures from over 500 scientists in support.

In response, DeepMind's Vice President for Research Pushmeet Kohli expressed on Platform X that the team is committed to releasing the AF3 model (including weights) for academic use within six months. Once the code is released, Nature will update the already published paper. On May 22, Nature published a response to these questions and criticisms. In this editorial, Nature stated that its editorial policy is designed to support open science but acknowledged that these policies may have limits, such as the lack of reporting standards for data across all disciplines or the technical infrastructure required for data public storage, or the need to protect confidentiality, as well as situations where data retention is necessary for security, legal, or other reasons, hence there may be an option to publish limited code.

Nature hopes that journals can better and more openly collaborate with the private sector and scientists, but achieving this goal requires a process and the participation and dialogue of all stakeholders, so they hope to solicit opinions on how to ensure the openness of all parties in the research ecosystem.

Microbiology

For the first time, bacteria have been found to convert exogenous RNA into their own genes, with bacteria using their own immune systems to fend off viral invasions. Among them, the defense-related reverse transcriptase (DRT) system generated by bacteria uses DNA synthesis, providing a form of immune strategy.

Recent scientific research discovered a unique immune mechanism in Klebsiella pneumoniae, which achieves de novo gene synthesis in an unprecedented rolling-circle retrotranscription manner through non-coding RNA (ncRNA). The research team detected abnormally long DNA sequences in the bacteria, which consisted of many identical repetitive segments, each matching a short segment of mysterious RNA.

Research has found that reverse transcriptase in Klebsiella pneumoniae can loop around the RNA sequence multiple times, copying the RNA molecule into DNA multiple times. This class of enzyme, known as DRT2 enzyme, can generate tandem cDNA repetitive sequences by template jumping on ncRNA, eventually facilitating the formation of long double-stranded DNA. Notably, these tandem repetitive DNA fragments can encode a protein sequence known as an open reading frame (ORF). The research team named this sequence “neo,” which, due to the lack of a stop codon, will continuously proceed once expression begins. When the bacterium is subjected to viral infection, the Neo protein is induced and leads to the cell stopping growth and division. This study showcases for the first time a pathway by which bacteria may acquire entirely new genes through the reverse transcription process.

In the field of optogenetics, researchers at Massachusetts Institute of Technology (MIT) have successfully used light-control techniques to regulate muscle contractions more precisely. This is a major advancement compared to traditional functional electrical stimulation (FES) methods and is more consistent with the body's natural muscle control mechanisms, representing a significant step forward in helping paralyzed or amputee patients recover limb function. They compared the muscle strength produced under the influence of FES and optogenetic technology in experimental mice and found that the muscle contractions induced by optogenetic stimulation were smooth and gradual and could reduce muscle fatigue.

Based on experimental data, the research team developed a mathematical model of optogenetic muscle control and designed a closed-loop controller to calculate the light intensity needed to produce specific muscle strength. The study showed that muscles stimulated with optogenetic techniques could maintain over an hour of active, non-fatigued performance, while muscles stimulated by FES could only last for 15 minutes. To apply this technology to humans, scientists are working on solving the method of delivering light-sensitive proteins into human tissues without causing an immune response, and developing new types of sensors and light sources.