This week in AI - Can we rely on DeepMind's ethical practices?

This week in AI - Can we rely on DeepMind's ethical practices?

This week in the field of artificial intelligence (AI), DeepMind, the AI research and development lab owned by Google, published a paper introducing a framework for assessing the societal and ethical risks associated with AI systems.

The release of the paper, which emphasizes the involvement of AI developers, app developers, and a wide range of public stakeholders in the evaluation and auditing of AI, is strategically timed. It precedes the upcoming AI Safety Summit, a two-day event sponsored by the U.K. government that brings together international governments, leading AI companies, civil society groups, and research experts to discuss the management of risks posed by recent advancements in AI, including generative AI technologies like ChatGPT and Stable Diffusion. During the summit, the U.K. is planning to introduce a global advisory group on AI, modeled in part after the U.N.'s Intergovernmental Panel on Climate Change. This advisory group will consist of rotating academics tasked with providing regular reports on cutting-edge AI developments and their associated risks.

DeepMind is making its perspective known ahead of the policy discussions at the summit. The research lab makes some sensible suggestions, such as the need to examine AI systems at the point of human interaction and consider how these systems are utilized and integrated into society.

To evaluate DeepMind's proposals effectively, it is helpful to examine the transparency of its parent company, Google. A recent study conducted by Stanford researchers evaluated and ranked ten major AI models based on their degree of openness. One of Google's flagship text-analyzing AI models, PaLM 2, received a lackluster score of 40% when assessed on criteria including disclosure of training data sources, information about hardware usage, labor involved in training, and other relevant details. While DeepMind did not directly develop PaLM 2, the lab itself has not consistently demonstrated transparency regarding its own models. The fact that its parent company falls short on transparency measures indicates a lack of top-down pressure for DeepMind to improve in this area.

On the other hand, DeepMind seems to be taking steps to address perceptions of being secretive about its models and their inner workings. Several months ago, DeepMind, together with OpenAI and Anthropic, committed to providing the U.K. government with early or priority access to its AI models to support research on evaluation and safety.

However, one might question the motivations behind these actions. DeepMind is not known for philanthropy, as it earns hundreds of millions of dollars annually through licensing agreements with Google teams. The lab's next significant ethics challenge might come in the form of Gemini, its upcoming AI chatbot, which DeepMind CEO Demis Hassabis has claimed will rival OpenAI's ChatGPT in terms of capabilities. If DeepMind intends to be taken seriously in the field of AI ethics, it must transparently and thoroughly outline the weaknesses and limitations of Gemini, not just its strengths. Observers will undoubtedly be closely monitoring how this unfolds in the coming months.

In addition to the DeepMind update, here are some other noteworthy AI stories from the past few days:

1. Microsoft study finds flaws in GPT-4: A scientific paper affiliated with Microsoft examined the "trustworthiness" and toxicity of large language models, including OpenAI's GPT-4. The researchers discovered that an earlier version of GPT-4 can be more easily prompted to generate toxic or biased text compared to other models. This raises concerns.

2. ChatGPT gains web searching and DALL-E 3: OpenAI has officially launched the web browsing feature for ChatGPT, following a beta release after several months of hiatus. Furthermore, OpenAI has transitioned DALL-E 3 into beta, marking the latest version of their text-to-image generation system.

3. Competing models against GPT-4V: OpenAI is preparing to release GPT-4V, a variant of GPT-4 with image understanding abilities. However, two open-source alternatives, LLaVA-1.5 and Fuyu-8B from well-funded startup Adept, have already emerged. These models may not match GPT-4V in capability, but they come close and have the benefit of being freely available for use.

4. Can AI play Pokémon?: Over the past few years, software engineer Peter Whidden from Seattle has trained a reinforcement learning algorithm to play the first game in the Pokémon series. While it currently only reaches Cerulean City, Whidden is confident that the algorithm will continue to improve.

5. Ai-powered language tutor: Google aims to challenge Duolingo with its new Google Search feature, specifically designed to enhance English speaking skills. Rolling out on Android devices in select countries over the next few days, this interactive feature offers language learners the opportunity to practice speaking and translation involving English.

6. Amazon expands its fleet of warehouse robots: This week, at a special event, Amazon unveiled its plans to test Agility's bipedal robot, Digit, within its facilities. However, it remains uncertain whether Amazon will fully deploy Digit to its warehouse operations, which already rely on over 750,000 robot systems, according to Brian's analysis.

7. Advances in simulation technology: Alongside Nvidia's recent demonstration of leveraging an LLM (Language Model) to write reinforcement learning code in order to improve the performance of an AI-driven robot, Meta introduced Habitat 3.0. This latest iteration of Meta's dataset facilitates training AI agents in realistic indoor environments, allowing for the potential inclusion of human avatars in virtual reality settings.

8. China's tech giants invest in OpenAI competitor: Zhipu AI, a China-based startup specializing in generative AI models that directly contend with OpenAI's offerings, recently secured ¥2.5 billion ($340 million) in total funding for this year. This announcement coincides with escalating geopolitical tensions between the United States and China, with no indication of resolution in the near future.

9. US restricts China's access to AI chips: In response to growing geopolitical tensions, the Biden administration introduced a series of measures aiming to curb Beijing's military aspirations. Among these measures is an additional restriction on Nvidia's shipment of AI chips to China. The new rules will impact A800 and H800, two AI chips explicitly designed by Nvidia for continued shipment to China.

10. AI-generated pop song renditions become internet sensations: Amanda explores an intriguing trend involving TikTok accounts that employ AI technology to create renditions of popular rock songs from the '90s and '00s, performed by animated characters such as Homer Simpson. While seemingly lighthearted and entertaining, Amanda delves into the underlying dark implications associated with this practice.

More Opportunities for Machine Learning

Machine learning models have revolutionized advances in the biological sciences. Examples such as AlphaFold and RoseTTAFold have demonstrated how the right AI model can trivialize complex problems like protein folding. Building on these successes, David Baker and his team have expanded the prediction process beyond the structure of amino acid chains. Understanding the interactions of proteins with other compounds and elements is crucial for comprehending their actual shape and activity. Thus, the development of RoseTTAFold All-Atom represents a significant step forward in simulating biological systems.

In addition, integrating visual AI into laboratory work and educational tools offers immense possibilities. The SmartEM project, a collaboration between MIT and Harvard, incorporates computer vision and ML control systems into a scanning electron microscope. This intelligence allows the microscope to intelligently examine specimens, focusing on areas of interest or clarity while avoiding irrelevant regions. Furthermore, it enables smart labeling of resulting images.

Exploring the applications of AI and advanced technologies in archaeology continues to captivate the imagination. Lidar has unveiled ancient Mayan cities and highways, while leveraging AI to decipher incomplete texts from ancient Greece bridges historical gaps. A remarkable achievement is the reconstruction of a scroll believed to have been destroyed during the Pompeii volcanic eruption.

Luke Farritor, a computer science student at the University of Nebraska–Lincoln, trained a machine learning model to enhance scans of charred, rolled-up papyrus. This innovative approach made invisible patterns visible to the naked eye, demonstrating potential for valuable academic contributions. Despite revealing only the word "purple" so far, even this limited insight has astonished papyrologists.

Moreover, AI has made significant strides in the realm of academic research and publication. While AI lacks the ability to ascertain truth and factual accuracy, it can discern contextual cues of high-quality Wikipedia articles and citations. Leveraging these cues, AI can explore alternative sources on the web, providing valuable support to articles lacking citations or facing editorial uncertainty.

Even in the realm of higher mathematics, language models can be fine-tuned to achieve surprising results. Llemma, an open model trained on mathematical proofs and papers, boasts the capability to solve complex problems. While not the first of its kind, Llemma's success and improved efficiency demonstrate the competitiveness of "open" models. Ensuring that certain domains are not monopolized by private models, the replication of their capabilities within the open-source community holds considerable value.

However, there are aspects of AI research that raise concerns. Meta's progress in academic work related to reading minds demands attention. Although findings presented in the paper "Brain decoding: Toward real-time reconstruction of visual perception" may insinuate mind-reading, the reality is less direct. By studying high-frequency brain scans while individuals view specific images, researchers can then generate near real-time reconstructions of their thought processes. Nonetheless, generative AI likely plays a role in creating visual representations that may not directly correspond to the original scans.

The question of whether we should use AI to read people's minds, if ever possible, remains open. DeepMind's insight into this matter may shed light on the ethical considerations involved. Finally, an aspirational project at LAION called Multilingual Contrastive Learning for Audio Representation Acquisition (CLARA) strives to enhance language models' understanding of the subtleties in human speech. Detecting sarcasm, lies, and emotional states from sub-verbal cues like tone and pronunciation poses a challenge for machines. CLARA employs a library of audio and text in multiple languages to identify emotional states and other non-verbal cues related to speech understanding.