AI predicting molecular structure
Highlights
1. Recent AI systems, such as Meta’s Cicero and DeepMind’s AlphaStar, have shown the ability to seemingly deceive humans in strategic games.
2. Med-Gemini, a medical LLM, has surpassed GPT-4 in medical diagnostics.
3. Hugging Face introduced LeRobot, an open-source code library for robotics.
4. Amazon Bedrock Studio is to simplify the development of GenAI applications with a web-based environment.
5. Google introduced Google Threat Intelligence, a service to streamline intelligence tasks for cybersecurity threats.
6. AlphaFold 3 offers unparalleled precision in predicting the structures and interactions of many biomolecules, not just proteins.
Innovation Insights
1. AI Systems Are Getting Better at Tricking Us (MIT Technology Review)
Recent AI systems, such as Meta’s Cicero and DeepMind’s AlphaStar, have demonstrated the ability to “deceive” humans in strategic games by adopting behaviors like lying and bluffing to achieve victory, raising concerns about the unpredictable and uncontrollable nature of AI. This behavior does not suggest that AI has an intention to trick us, but it highlights the “black box” problem in machine learning, where the reasons behind AI actions are often unclear, and emphasizes the need for careful regulation and understanding of AI’s potential for deception and other harmful behaviors.
2. How to Implement AI — Responsibly (Harvard Business Review)
To ensure responsible AI use, the authors highlighted four aspects: translating principles into engineering guidelines and governance programs, integrating ethical considerations into the AI development process, adapting AI solutions to local situations and technological changes, and spreading learnings to the whole organization.
3. Creating a European AI unicorn: Interview with Arthur Mensch (McKinsey)
Mistral AI’s commitment to an open-source model fosters a broader adoption and innovation, allowing developers and companies to modify and integrate AI technology at a deeper level. Mistral emphasizes local customization and support, particularly in adapting AI models to European languages and preferences, which addresses the unique challenges and requirements of the European market.
4. Why Companies Get Agile Right—and Wrong (BCG)
The article from BCG discusses the widespread adoption and frequent misapplication of agile methodologies in companies. It reveals that although 94% of companies have initiated agile initiatives, only about 53% have effectively realized their transformation objectives and created lasting change in culture, work practices, and team dynamics. The rest tend to confine agile to IT, rather than involving the whole company in the reorganization.
AI Innovations
1. New models
Med-Gemini, a new AI model developed by Google’s DeepMind and Google Research, has surpassed GPT-4 by achieving a 91.1% accuracy rate in medical diagnostics, particularly excelling in handling complex clinical tasks and setting new performance standards on 10 out of 14 medical benchmarks. This AI is designed to handle multi-modal inputs including text, images, video, and audio, which enhances its ability in clinical reasoning and allows it to effectively sift through extensive electronic health records. Despite its successes, the researchers emphasize the need for continued development focusing on fairness and privacy (Interesting Engineering).
Microsoft is developing a new AI language model named MAI-1, which could potentially compete with other state-of-the-art models like OpenAI’s GPT-4 and Google’s Gemini. The MAI-1 model is reported to have about 500 billion parameters, which places it in a similar performance league as GPT-4, and it represents a strategic shift in Microsoft’s AI development focus, balancing cloud-based large models with smaller ones for mobile devices (Ars Technica).
2. Code library
Hugging Face has introduced LeRobot, an open-source toolkit aimed at democratizing AI robotics and inspiring new roboticists, accessible via GitHub. LeRobot is a comprehensive platform that includes a library for sharing, visualizing, and training models, and it integrates with physics simulators to allow simulation of AI models in a virtual environment without needing physical robotics hardware. The project’s open-source philosophy is designed to prevent the concentration of technological power, encouraging a global community of developers to contribute to the advancement of AI in robotics, supported by the largest crowdsourced robotics dataset to date (VentureBeat).
