Google and Microsoft Square Off in an AI-Powered Search Engine Fight: This week, Google and Microsoft announced plans to roll out new AI tools that could dramatically reshape both companies and the search engine business as a whole. On Monday, Google CEO Sundar Pichai unveiled Bard — a “conversational AI service” that, like OpenAI’s ChatGPT, appears capable of responding to human queries and synthesizing information. Unlike ChatGPT, it seems that Bard — which is based on a “lightweight” version of Google’s LaMDA language model — will be connected to the internet, allowing it to draw on new and up-to-date information. The day after Pichai’s announcement, Microsoft announced its own AI-powered changes to its Bing search engine and Edge web browser. Microsoft reached a multibillion-dollar investment deal with OpenAI in January, and according to the company, a “new, next-generation” language model developed by the research lab is under the hood of the “new Bing” (unconfirmed rumors say that model is the long-awaited GPT-4). For now, Google’s Bard is still not available to the public, and users have to jump through some hoops to use the new Bing. But it seems likely that, for better or worse, language model-powered tools will become a major part of the search engine business.
DeepMind Researchers Explore Ways to Speed Up and Scale Up AI Models: While the heavyweight fight between Google and Microsoft takes center stage, recent research from the Alphabet-owned DeepMind seems to offer a glimpse into the near future of AI development:
In a paper released in January, researchers from DeepMind introduced a reinforcement learning-trained agent that can adapt to new problems, learns from first-person demonstrations, and shows signs of using hypothesis-driven trial-and-error approaches to unsolved problems. As the paper explains, the RL agent — which was trained on a “vast, smooth, and diverse task space” — adapted to previously unseen situations on a timeline comparable with human players. Furthermore, the paper’s finding that performance scales with the size of the agent, its memory, and its training data could indicate a path for future research. As CSET Non-Resident Research Fellow Jack Clark noted in his newsletter, similar findings on scaling and performance contributed to increased interest in (and funding for) language models.
In another paper published last week, DeepMind researchers proposed a more efficient way to sample from large language models. Using a smaller, faster model (in this case, a 4 billion parameter one) to generate a “draft” output, the researchers then used a larger, more sophisticated “target” model (the 70 billion parameter Chinchilla) to score the draft output, rather than generate its own output from scratch. According to the paper, this method of “speculative sampling” produced a 2–2.5 times decoding speedup compared to the target model working on its own. Because the proposed method does not require changes to the target model, theoretically it should be relatively easy to implement with models like those powering ChatGPT, Bard and other popular LLMs.
Expert Take: “While it is encouraging to see the Netherlands and Japan agreeing to join the United States on these controls, U.S. officials should not make a habit of using unilateral and extraterritorial controls to get allies on board with our policies. Continuing to do so will only sour our relationship with countries that we need on our side to effectively compete with China. This is also the first time that this approach has worked. No U.S. allies have gone along with Huawei controls or adopted any of our Entity List controls. Although from the outside we may not have seen the same levels of pressure applied by the Biden administration to Japan and the Netherlands in this scenario, there has long been an understanding that multilateral controls are generally more effective long-term.” — Emily S. Weinstein, Research Fellow
NIST Releases Its AI Risk Management Framework: The National Institute of Standards and Technology released the first official version of its AI Risk Management Framework and a companion AI RMF Playbook, both of which are meant to help developers anticipate and manage the risks unique to AI systems. Created at the direction of Congress as part of the FY2021 NDAA, the voluntary framework outlines key characteristics of trustworthy AI systems and describes four functions that NIST says organizations can employ to address AI risk. The accompanying playbook includes suggested actions, supplementary reading, and other resources for organizations that want to implement these four functions. NIST’s AI RMF comes at an important moment — the appeal of generative AI tools has both spurred the development of new AI systems and increased concerns about their potential for misuse. But it remains to be seen whether voluntary frameworks such as the AI RMF and the Office of Science and Technology Policy’s “Blueprint for an AI Bill of Rights” will be enough to mitigate potential risks. NIST appears determined to keep an eye on the impact of the AI RMF and adapt accordingly — a note at the beginning of the document says the agency plans to review the “content and usefulness” of the RMF regularly and update it when warranted, potentially including industry-specific recommendations (the playbook, meanwhile, will be updated frequently).
The United States and India Team Up On Critical and Emerging Tech: Last week, officials from the United States and India held the inaugural meeting of the U.S.-India initiative on Critical and Emerging Technology. The iCET initiative was announced last year by President Biden and Indian Prime Minister Narendra Modi; last week’s meetings — led by U.S. National Security Advisor Jake Sullivan and his Indian counterpart, Ajit Doval — offered more details on the initiative meant to “elevate and expand” technological and defense industrial cooperation between the two countries. In comments to the press, Sullivan acknowledged the initiative’s relationship to the United States’ broader strategic competition with China, calling it a “big foundational piece of an overall strategy to put the entire democratic world in the Indo-Pacific in a position of strength.” Among other things, the two governments announced a partnership between the U.S. National Science Foundation and Indian science agencies meant to increase scientific collaboration (including on AI research), launched an “Innovation Bridge” to connect defense startup companies in the United States and India, and discussed plans to build up India’s semiconductor industry. The next iCET meeting will take place in New Delhi later this year.
The Pentagon’s AI Office Is Tasked With “Reinvigorating” DOD Experiments: Earlier this month, the Pentagon concluded the fifth in a series of Global Information Dominance Experiments (GIDE V), joint exercises meant to test the military’s data sharing and integration, evaluate its use of AI systems, and provide insight into the implementation of Joint All-Domain Command and Control, the DOD’s plan to connect and coordinate its services’ sensors in a single network. The previous four iterations of GIDE had been run by NORAD and Northern Command, but GIDE V was led by the Pentagon’s Chief Digital and Artificial Intelligence Office in partnership with the Joint Chiefs of Staff. In a press release, the DOD said that responsibility had been shifted to the CDAO in the hopes of “reintroducing and reinvigorating” the experiments. The Pentagon plans to run three more GIDE exercises over the course of 2023. In comments to Breaking Defense, GIDE Mission Commander Col. Matt Strohmeyer said the coming exercises will involve “increasingly complex demonstrations” with the goal of “stress-testing current systems and processes.” A date has not yet been set for GIDE VI.
In Translation CSET’s translations of significant foreign language documents on AI
Necessary cookies are absolutely essential for the website to function properly. This category only includes cookies that ensures basic functionalities and security features of the website. These cookies do not store any personal information.
Any cookies that may not be particularly necessary for the website to function and is used specifically to collect user personal data via analytics, ads, other embedded contents are termed as non-necessary cookies. It is mandatory to procure user consent prior to running these cookies on your website.