|As powerful as general-purpose text-generating language models are proving to be, baking in domain specificity improves their output within a given realm. A case in point: BloombergGPT, which is trained on Bloomberg’s wealth of financial data and is being used as a textual interface for users who seek the full power of Bloomberg Query Language (BQL) but lack the coding expertise. In Montreal in June, Bloomberg’s Gideon Mann will share how his team built this and its implications. Here’s more information about the conference and how to register for it.
It’s been eight years since the Be My Eyes app first appeared. Using it, a blind person is connected to a sighted volunteer who looks at the view from the blind user’s phone camera and answers a question that person has: Am I at the right address? What is the SPF of the sunscreen in this tube? Is this the black or the red wire? These days AI tools can—and do—answer such questions, but the volunteer base of Be My Eyes hopes to retain a role, as it provides them with the opportunity to help others. We shouldn’t forget the benefits of human-to-human contact, even when mediated by an app. Click here to read the article.
But sometimes giving a problem over to automation can be just the ticket. Toward this end, a developer created the GPT-4-based tool Wolverine that can intervene with when a Python program crashes. It introduces a fix, reruns the script, and iterates until no bugs remain. Perhaps best of all, it offers an explanation for the human to mull over. (Armando Solar-Lezama, Washington, D.C., Sep 2018, who will be with us in Montreal this June; David Gunning, Brooklyn, Jun 2018) Click here to read the article.
Japanese researchers have developed a simple pipette-based method to construct layered hydrogels in a cube-like structure to serve as a scaffold for growing complex, 3-D organoids for use in drug evaluation and as a stepping stone for lab-grown artificial organs. (Michael McAlpine, Washington, D.C., Sep 2018; UCLA field trip, Jun 2022) Click here to read the article.
In a biomimetic thrust, Carnegie Mellon researchers are moving away from overreliance on vision sensors for legged robots navigating unfamiliar terrain. They’ve added a tail for balance and internal sensors for proprioception to let their robots right themselves when stumbling. The strategy merges proprioception and motion planning, giving the robots the ability to mimic the natural habits of felines to walk a path that follows their own footprints. (Sangbae Kim, Cambridge, Mar 2023; Los Angeles, Feb 2011; David Hu, Pittsburgh, Jun 2019) Click here to read the article.
A team of physicists spanning a trio of pan-New York State institutions and Aalto University in Finland have demonstrated the potential to form novel quantum devices by combining the Klein tunneling property of graphene and the Andreev reflection property of high temperature superconductors into Andrev–Klein electronic transport—a phenomenon that has been theoretically predicted but only now observed experimentally. The low-power property of graphene makes this material a promising contributor, in combination with well-proven superconductors, in future development of both qubit and topological quantum computers. (John Preskill, virtual conference, Sep 2020; Chris Monroe, Washington, D.C., Sep 2019; Rodney Van Meter, San Francisco, Dec 2014, and Tokyo regional meeting, Mar 2017) Click here to read the article.
The TTI/Vanguard community is better attuned to the Silicon Valley–AI landscape than most, yet with all the current hype it can be hard to keep track of which mover or shaker belongs to each of the various AI camps. This Washington Post article offers a roadmap.
With generative AIs taking center stage, and the incorporation of not only expert systems—as per BloombergGPT, above—adding to its utility in specific domains, this general mode of AI does not supplant the need for logical reasoning in important endeavors like science. Based on the principle of symbolic regression to devine equations to fit real-world data, “AI-Descartes” has successfully rediscovered Kepler’s third law of planetary motion and Irving Langmuir’s 1932 Nobel Prize-winning equations governing the adsorption of gas on a solid surface. Once the system comes up with candidate equations—something other science-directed AIs can do—it then uses logic to determine which one best fits existing scientific theory. Click here to read the article.
‘The scientific method actually correctly uses the most direct evidence as the most reliable, because that’s the way you are least likely to get led astray into dead ends and to misunderstand your data.”—Aubrey de Grey