TTI/Vanguard Newsletter: March 31, 2023
Thanks to all of you who joined us in person or over Zoom for the Data, AI, and Robotics conference. We hope you got as much out of it as we did—despite the challenges the weather threw our way. Speaker presentations are available to members logged into the conference webpage, and videos will be posted soon.
The early lineup for our June 21–23 conference in Montreal on Rebooting Computing is growing longer daily. Here’s the latest:
• Stephano Cetola, Director of Technical Programs, RISC-V: RISC-V—Why Silicon Hasn’t Been This Cool in Decades
• Jim Hendler, Professor and Director of the Future of Computing Institute, RPI: Engineering the Future of Computing
• Paul Mola, President and CEO, Roswell Biotechnology: Molecular Electronics: The Semiconductor Revolution to Digitize Biology
• Dirk Riehle, Professor, University of Erlangen: Developing a Corporate Open Source Strategy
• Armando Solar-Lezama, Massachusetts Institute of Technology: Implications of Large Language Models on Software Development
• Chenhui (Charles) Yuan, PhD Candidate, Massachusetts Institute of Technology: Building Abstractions for Quantum Programming
Register today, and be sure to arrive in time for a 1 pm departure on June 21 for a field trip to member organization CAE.
With reasonable confidence, your newsletterer pins the 2014 talk by Autodesk Research’s Carlos Olguin entitled “Programming Matter across Domains and Scales” as TTI/Vanguard’s first overt exposure to generative techniques, in that case a generative design tool for physical artifacts like chairs or airplane bulkheads. More recently, the various generations of GPT, DALL-E (Aditya Ramesh, Washington, D.C., Sep 2022), drug discovery tools Chroma (Molly Gibson) and Enki (Handol Kim), and their kin in all manner of domains are dominating conversations, as was the case in Cambridge earlier this month. So, here’s a question for our community: How long do you predict it will be until generative AI fades into the background as an essential substrate of life as we know it, just like e-commerce—which served as the theme of entire TTI/V conferences in the 1990s—is now simply “commerce”?
And, reflecting how quickly this field is rolling out, all manner of chatbots have emerged—or at least started to receive considerable airtime—since we met in Cambridge: Bing AI (rooted in GPT-4), Google’s Bard (rooted in LaMDA), Anthropic’s purposefully “helpful, honest, and harmless” Claude (rooted in the AnthropicLM v4-s3 Constitutional AI model), and Baidu’s Ernie Bot. Clearly the race is on.
Yet, what appears to be this week’s most popular activity among academics and AI startup founders alike? Signing onto Yoshua Bengio’s open letter calling for a pause on more-powerful-than-GPT-4 AI experiments for at least six months while humanity—most likely tilted toward its illuminati class—not only ponders the societal and ethical implications of continuing down the path of developing systems that are “human-competitive at general tasks,” but simultaneously develops and implements “a set of shared safety protocols for advanced AI design and development” and imposes well-constructed regulations on the industry. The stated goal: make “systems more accurate, safe, transparent, robust, aligned, trustworthy, and loyal.” (But aligned with what and loyal to whom?)
Of course, one of the dangers of generative AIs is the inability of humans to discern whether what they encounter is real or generated. While it is the custom of this newsletter to present articles about research rather than the research per se, this work about distinguishing between diffusion model-generated images and real images doesn’t yet seem to be written up, so we’re linking to the raw research. In it, the authors create a representation they dub the diffusion reconstruction error (DIRE) that inverts an input image and maps it to a noise vector then reconstructs an image from that noise vector through a denoising process. The computed difference between the input and output is the DIRE metric, which is universally much larger for real images than for generated images. This evaluation method outperforms previous generated-image detectors.
About five months ago we reported the seemingly successful early results of NASA’s Double Asteroid Redirection Test (DART) mission, executed by TTI/Vanguard member Johns Hopkins Applied Physics Laboratory, in which a spacecraft rammed an asteroid to test the viability of saving Earth from a future incoming space rock. Further analysis shows that DART, um, rocked it—although the deflection proved considerably greater than simulated, due to the exaggerated recoil effect from the scattering of ejected asteroid fragments, suggesting a direction for future refinements. (Dan Durda, Washington, D.C., May 2007) Click here to read the article.
Research over the past few years has shown the benefits of psychedelics-assisted therapy using LSD or psilocybin in alleviating symptoms of OCD, PTSD, addiction, and depression. Computational physicist and artist David Glowacki has developed a VR experience, dubbed Isness-D, to simulate a multiuser transcendental experience. Each participant self-represents as a glowing cloud in the VR environment. As participants collectively move together, they feel as though their cloud–bodies merge, egos dissolve, and connectedness comes to the fore. Isness-D is already being offered as a component of company wellness retreats; to patients, families, and caregivers facing terminal illness; and is being suggested for use in couples and family therapy. (Nolan Bushnell, San Francisco, Dec 2016) Click here to read the article.
Using ultrafast transient absorption spectroscopy, which probes living photosynthetic cells on the femtosecond timescale, researchers have discovered a previously unknown pathway in photosynthesis that enables electron extraction much earlier in the process. The hope is to improve the efficiency of light-induced energy production. Click here to read the article.
An international quartet of mathematicians have discovered a 2-D geometric shape that can successfully tile a surface without repetition. They have proven that the 13-sided aperiodic monotile dubbed “the hat” satisfies the criteria of the elusive einstein shape. Exciting, yes—but practical? The researchers suggest the most likely application of the hat is in the arts—or perhaps on your bathroom floor. Click here to read the article.
“If I had to pick one artist to tile my bathroom I would go with MC Escher.”—Demetri Martin
Regards,
Lew