Following the release the new GPT-4 engine and Whisper API in March, OpenAI announced Thursday that it has begun introducing plugins for ChatGPT. These will enable the chatbot to interact with 3rd-party APIs, tailoring its responses to specific circumstances as defined by the developers while expanding the bot's range of capable actions.

Say you want to develop a chatbot that users can talk sports with. Before the latest GPT-4 upgrade, the chatbot would only be able to discuss games and scores that happened in the past, specifically in 2021 which is when GPT-3's training data was assembled. It wouldn't pull real-time data or even be aware that the year 2022 existed. With a chatGPT plugin, you'll be able to tack ChatGPT functionality onto your existing code stack where it will be able to do anything from retrieve real-time information calls (sports scores, stock prices, breaking news) to pulling specific knowledge-base information like your company's internal documents or from your personal cloud. It will even be able to take action on behalf of the user like booking a flight or ordering take-out — think, an installable Google Assistant made by the OpenAI folks.

"The AI model acts as an intelligent API caller. Given an API spec and a natural-language description of when to use the API, the model proactively calls the API to perform actions," the OpenAI team wrote. "For instance, if a user asks, 'Where should I stay in Paris for a couple nights?', the model may choose to call a hotel reservation plugin API, receive the API response, and generate a user-facing answer combining the API data and its natural language capabilities."

The company also notes that using plug-ins to bridge the knowledge gap between what the model was trained on and what has happened since should help reduce the AI's tendency to hallucinate facts when answering complex questions. "These references not only enhance the model’s utility but also enable users to assess the trustworthiness of the model’s output and double-check its accuracy, potentially mitigating risks related to over-reliance," the team wrote.

The added capabilities and information afforded the model through its plug-in also greatly increase the chances of the model returning problematic responses. To avoid the $100 billion hit that Google took over Bard, OpenAI has reportedly stress-tested these plug-ins extensively. "We’ve performed red-teaming exercises, both internally and with external collaborators, that have revealed a number of possible concerning scenarios," the team wrote. They plan to use those findings to, "inform safety-by-design mitigations" to improve transparency and hobble the plug-in against partaking risky behaviors.  

The plug-in itself is still in early alpha with limited availability. OpenAI granted early access to a handful of partner companies including Expedia, Instacart, KAYAK, OpenTable, Shopify, Slack, Wolfram, and Zapier for use in their existing apps. You'll need to add your name to the waitlist to try it for yourself.

The age of enterprise AI has come crashing down upon us in recent months. Public infatuation with ChatGPT since its release last November has opened the floodgates of corporate interest and set off an industry-wide land grab with every major tech entity vying to stake their claim in this burgeoning market by incorporating generative AI features into their existing products. Heavyweights including Google, Microsoft, Meta, and Baidu are already jockeying their Large Language Models (LLMs) for market dominance, while everybody else, from Adobe and AT&T to BMW and BYD, scrambles to find uses for the revolutionary technology.  

NVIDIA's newest cloud services offering, AI Foundations, will allow businesses lacking the time and money to develop their own models from scratch to "to build, refine and operate custom large language models and generative AI models that are trained with their own proprietary data and created for their unique domain-specific tasks."   

These models include NeMo, NVIDIA’s text-to-image generation engine and DALL-E 2 competitor; BioNemo, a drug and molecule discovery-focused fork of the NeMo model built for the medical research community; and Picasso, an AI capable of generating images, video and “3D applications… to supercharge productivity for creativity, design and digital simulation,” according to Tuesday’s release. Both flavors of NeMo are still in early access and Picasso remains in private preview despite Tuesday's news, so it'll be a minute before any of them are released to the wider public. NeMo and Picasso both operate on NVIDIA's new DGX Cloud platform and will eventually be accessible through an online portal.  

These enterprise-facing cloud-based services function as blank templates that companies can pour their own databases into to train on specifically. So while something like Google’s Bard AI is trained on (and will pull from) data from all over the internet to provide a generated response, NVIDIA’s AIs will allow companies to tailor a similarly-styled LLM to their own specific needs using their own proprietary data — think of ChatGPT but solely for one Pharma company's research division. The models can be trained with anywhere from 8 billion to 530 billion parameters, which is more than triple the 185 billion parameters GPT-3.5 provided.

Imagine StableDiffusion, but trained on Getty Images with Getty’s actual permission. NVIDIA announced such a system Tuesday built on the NeMo cloud service: a series of responsibly sourced text-to-image and text-to-video models, "trained on Getty Images' fully licensed assets," Tuesday’s press release read. "Getty Images will provide royalties to artists on any revenues generated from the models."

BioNeMo, uses the same technical underpinnings as NeMo itself, but is geared entirely towards drug and molecule discovery. Per Tuesday’s release, Bio NeMo, “enables researchers to fine-tune generative AI applications on their own proprietary data, and to run AI model inference directly in a web browser or through new cloud APIs that easily integrate into existing applications."

"BioNeMo is dramatically accelerating our approach to biologics discovery," Peter Grandsard, executive director of Biologics Therapeutic Discovery at Amgen said in a statement. "With it, we can pre-train large language models for molecular biology on Amgen’s proprietary data, enabling us to explore and develop therapeutic proteins for the next generation of medicine that will help patients." 

Six models will be available at launch including DeepMind’s AlphaFold2, Meta AI’s ESM2 and ESMFold predictive models ProtGPT-2, DiffDock and MoFlow. According to the companies, incorporating AI-based predictive models helped reduce the time to train “five custom models for molecule screening and optimization” using Amgen’s proprietary data on antibodies from the usual three months down to four weeks.

NVIDIA announced a similar partnership with Shutterstock as well. The photography site will use Picasso to generate 3D objects from text prompts as a new feature within Creative Flow, with plans to offer it on Turbosquid.com and NVIDIA’s forthcoming Omniverse platform.

“Our generative 3D partnership with NVIDIA will power the next generation of 3D contributor tools, greatly reducing the time it takes to create beautifully textured and structured 3D models,” Shutterstock CEO Paul Hennessy, said in the release. “This first of its kind partnership furthers our strategy of leveraging Shutterstock’s massive pool of metadata to bring new products, tools, and content to market. By combining our 3D content with NVIDIA’s foundation models, and utilizing our respective marketing and distribution platforms, we can capitalize on an extraordinarily large market opportunity.”

NVIDIA is also partnering with Adobe as part of the latter’s Content Authenticity Initiative, which seeks to enhance transparency and accountability within the generative AI training process. The CAI’s proposals include a “do not train” list, similar to robot.txt but for images and multimodal content, and persistent origination tags that will detail whether a piece is AI generated and from where. The two companies have also announced plans to incorporate many of Picasso’s features directly into Adobe’s suite of editing software including Photoshop, Premiere Pro and After Effects. 

Adobe’s suite of photo and video editing software has long leveraged the assistance of machine intelligence to help its human users do their jobs, having employed the Sensei AI system for more than a decade to power features like Neural Filters in Photoshop or Acrobat's Liquid Mode. On Tuesday, Adobe revealed its next generation of AI features, a family of generative models the company has collectively dubbed, Firefly — the first of which will generate both images and font effects.

“Generative AI is the next evolution of AI-driven creativity and productivity, transforming the conversation between creator and computer into something more natural, intuitive and powerful,” said David Wadhwani, president, Adobe’s Digital Media Business, said in Tuesday’s release. “With Firefly, Adobe will bring generative AI-powered ‘creative ingredients’ directly into customers’ workflows, increasing productivity and creative confidence for all creators from high-end creative professionals to the long tail of the creator economy.”

With it, would-be digital artists are no longer limited by their sub-par dexterity or sheer lack of artistic talent — they will be able to speak into existence professional-quality illustrations using only the power of their words. And it's not just text-to-image — Firefly’s multimodal nature means that audio, video, illustrations and 3D models can all be generated via the system and enough verbal gymnastics. 

The first model of the Firefly family is, according to the company, trained on "hundreds of millions" of images from Adobe's Stock photo catalog, openly licensed content and stuff from the public domain, virtually guaranteeing the model won't result in lawsuits as StableDiffusion did with the Getty unpleasantness. It also helps ensure that Stock photographers and artists will be compensated for the use of their works in training these AIs.

Engadget was afforded a brief preview of the system ahead of Tuesday’s announcement. The input screen, where users will enter their text-based prompt to the system, features a curated selection of generated pieces as well as the prompts that instigated them. These serve to highlight the models generative capabilities and inspire other users to explore the bounds of their machine-assisted creativity.  

Once the user inputs their text prompt (in this case, Adobe’s PR used an adult standing on a beach with a double exposure effect using images derived from Adobe’s Stock photo database), the system will return around a half dozen or so initial image suggestions. From there, the user can select between popular image styles and effects, dictate their own edits to the prompt, collaborate with the AI and generally fiddle with the highly-steerable process until the system spits out what they’re looking for. The resulting image quality was nearly photorealistic, though none of the images from the demo features hands so we weren't able to count fingers for accuracy.

Initially, the trained image database will be Adobe's own licensed Stock library though the company is looking into allowing individual users to incorporate their own portfolios as well. This should allow photographers with their own established styles to recreate those aesthetics within the model so that what it generates fits in with the user's existing motif. The company did not provide a timeline for when that might happen.

The first model also has a sibling feature that can create customized font effects and generate wireframe logos based on scanned doodles and sketches. It’s all very cool but could potentially put just an unconscionable number of digital artists out of work if it were to be misappropriated. Adobe’s Content Authenticity Initiative (CAI) seeks to prevent that from happening.

The CAI is Adobe’s attempt to establish some form of ground rules in this new Wild West industry of Silicon Valley. It is a set of proposed industry operating standards that would establish and govern ethical behaviors and transparency in the AI training process. For example, the CAI would create a “do not train” tag that works in same basic idea as robots.txt does. That tag would be persistent, remaining with the art as it moves through the internet, alerting any who came across it that it was made by a machine. So far around 900 entities worldwide, "including media and tech companies, NGOs, academics and others," per the release, have signed on to the plan.

In the years leading up to, and through, World War II, animal behaviorist researchers thoroughly embraced motion picture technology as a means to better capture the daily experiences of their test subjects — whether exploring the nuances of contemporary chimpanzee society or running macabre rat-eat-rat survival experiments to determine the Earth's "carrying capacity." However, once the studies had run their course, much of that scientific content was simply shelved. 

In his new book, The Celluloid Specimen: Moving Image Research into Animal Life, Seattle University Assistant Professor of Film Studies Dr. Ben Schultz-Figueroa, pulls these historic archives out of the vacuum of academic research to examine how they have influenced America's scientific and moral compasses since. In the excerpt below, Schultz-Figueroa recounts the Allied war effort to guide precision aerial munitions towards their targets using live pigeons as onboard targeting reticles.

University of California Press

Excerpted from The Celluloid Specimen: Moving Image Research into Animal Life by Ben Schultz-Figueroa, published by the University of California Press. © 2023 by Ben Schultz-Figueroa.

Project Pigeon: Rendering the War Animal through Optical Technology

In his 1979 autobiography, The Shaping of a Behaviorist, B. F. Skinner recounted a fateful train ride to Chicago in 1940, just after the Nazis had invaded Denmark. Gazing out the train window, the renowned behaviorist was ruminating on the destructive power of aerial warfare when his eye unexpectedly caught a “flock of birds lifting and wheeling in formation as they flew alongside the train.” Skinner recounts: “Suddenly I saw them as ‘devices’ with excellent vision and extraordinary maneuverability. Could they not guide a missile?” Observing the coordination of the flock, its “lifting and wheeling,” inspired in Skinner a new vision of aerial warfare, one that yoked the senses and movements of living animals to the destructive power of modern ballistics. This momentary inspiration began a three-year project to weaponize pigeons, code-named “Project Pigeon,” by having them guide the flight of a bomb from inside its nose, a project that tied together laboratory research, military technology, and private industry.

This strange story is popularly discussed as a historical fluke of sorts, a wacky one-off in military research and development. As Skinner himself described it, one of the main obstacles to Project Pigeon even at the time was the perception of a pigeon guided missile as a “crackpot idea.” But in this section I will argue that it is, in fact, a telling example of the weaponization of animals in a modern technological setting where optical media was increasingly deployed on the battlefield, a transformation with increasing strategic and ethical implications for the way war is fought today. I demonstrate that Project Pigeon was historically placed at the intersection of a crucial shift in warfare away from the model of an elaborate chess game played out by generals and their armies and toward an ecological framework in which a wide array of nonhuman agents play crucial roles. As Jussi Parikka recently described a similar shift in artificial intelligence, this was a movement toward “agents that expressed complex behavior, not through preprogramming and centralization, but through autonomy, emergence, and distributed functioning.” The missile developed and marketed by Project Pigeon was premised on a conversion of the pigeon from an individual consciousness to a living machine, emptied of intentionality in order to leave behind only a controllable, yet dynamic and complex, behavior that could be designed and trusted to operate without the oversight of a human commander. Here is a reimagining of what a combatant can be, no longer dependent on a decision-making human actor but rather on a complex array of interactions among an organism, device, and environment. As we will see, the vision of a pigeon-guided bomb presaged the nonhuman sight of the smart bomb, drone, and military robot, where artificial intelligence and computer algorithms replace the operations of its animal counterpart.

Media and cinema scholars have written extensively about the transforming visual landscape of the battlefield and film’s place within this shifting history. Militaries from across the globe have pushed film to be used in dramatically unorthodox ways. Lee Grieveson and Haidee Wasson argue that the US military historically used film as “an iterative apparatus with multiple capacities and functions,” experimenting with the design of the camera, projector, and screen to fit new strategic interests as they arose. As Wasson argues in her chapter dedicated to experimental projection practices, the US Army “boldly dissembled cinema’s settled routines and structures, rearticulating film projection as but one integral element of a growing institution with highly complex needs.” As propaganda, film was used to portray the military to civilians at home and abroad; as training films, it was used to consistently instruct large numbers of recruits; as industrial and advertising films, different branches of the military used it to speak to each other. Like these examples, Project Pigeon relied on a radically unorthodox use of film that directed it into new terrains, intervening in the long-standing relationship between the moving image and its spectators to marshal its influence on nonhuman viewers, as well as humans. Here, we will see a hitherto unstudied use of the optical media, in which film was a catalyst for transforming animals into weapons and combatants.

Project Pigeon was one of the earliest projects to come out of an illustrious and influential career. Skinner would go on to become one of the most well-known voices in American psychology, introducing the “Skinner box” to the study of animal behavior and the vastly influential theory of “operant conditioning.” His influence was not limited to the sciences but was broadly felt across conversations in political theory, linguistics, and philosophy as well. As James Capshew has shown, much of Skinner’s later, more well-known research originated in this military research into pigeon-guided ballistics. Growing from initial independent trials in 1940, Project Pigeon secured funding from the US Army’s Office of Scientific Research and Development in 1943. The culmination of this work placed three pigeons in the head of a missile; the birds had been trained to peck at a screen showing incoming targets. These pecks were then translated into instructions for the missile’s guidance system. The goal was a 1940s version of a smart bomb, which was capable of course correcting mid-flight in response to the movement of a target. Although Project Pigeon developed relatively rapidly, the US Army was ultimately denied further funds in December of 1943, effectively ending Skinner’s brief oversight of the project. In 1948, however, the US Naval Research Laboratory picked up Skinner’s research and renamed it “Project ORCON” — a contraction of “organic” and “control.” Here, with Skinner’s consultation, the pigeons’ tracking capacity for guiding missiles to their intended targets was methodically tested, demonstrating a wide variance in reliability. In the end, the pigeons’ performance and accuracy relied on so many uncontrollable factors that Project ORCON, like Project Pigeon before it, was discontinued.

Moving images played two central roles in Project Pigeon: first, as a means of orienting the pigeons in space and testing the accuracy of their responses, examples of what Harun Farocki calls “operational images,” and, second, as a tool for convincing potential sponsors of the pigeon’s capacity to act as a weapon. The first use of moving image technology shows up in the final design of Project Pigeon, where each of the three pigeons was constantly responding to camera obscuras that were installed in the front of the bomb. The pigeons were trained to pinpoint the shape of incoming targets on individual screens (or “plates”) by pecking them as the bomb dropped, which would then cause it to change course. This screen was connected to the bomb’s guidance through four small rubber pneumatic tubes that were attached to each of side of the frame, which directed a constant airflow to a pneumatic pickup system that controlled the thrusters of the bomb. As Skinner explained: “When the missile was on target, the pigeon pecked the center of the plate, all valves admitted equal amounts of air, and the tambours remained in neutral positions. But if the image moved as little as a quarter of an inch off-center, corresponding to a very small angular displacement of the target, more air was admitted by the valves on one side, and the resulting displacement of the tambours sent appropriate correcting orders directly to the servo system.” 

In the later iteration of Project ORCON, the pigeons were tested and trained with color films taken from footage recorded on a jet making diving runs on a destroyer and a freighter, and the pneumatic relays between the servo system and the screen were replaced with electric currents. Here, the camera obscura and the training films were used to integrate the living behavior of the pigeon into the mechanism of the bomb itself and to produce immersive simulations for these nonhuman pilots in order to fully operationalize their behavior.

The second use of moving images for this research was realized in a set of promotional films for Project Pigeon, which Skinner largely credited for procuring its initial funding from General Mills Inc. and the navy’s later renewal of the research as Project ORCON. Skinner’s letters indicate that there were multiple films made for this purpose, which were often recut in order to incorporate new footage. Currently, I have been able to locate only a single version of the multiple films produced by Skinner, the latest iteration that was made to promote Project ORCON. Whether previous versions exist and have yet to be found or whether they were taken apart to create each new version is unclear. Based on the surviving example, it appears that these promotional films were used to dramatically depict the pigeons as reliable and controllable tools. Their imagery presents the birds surrounded by cutting-edge technology, rapidly and competently responding to a dynamic array of changing stimuli. These promotional films played a pivotal rhetorical role in convincing government and private sponsors to back the project. Skinner wrote that one demonstration film was shown “so often that it was completely worn out—but to good effect for support was eventually found for a thorough investigation.” This contrasted starkly with the live presentation of the pigeons’ work, of which Skinner wrote: “the spectacle of a living pigeon carrying out its assignment, no matter how beautifully, simply reminded the committee of how utterly fantastic our proposal was.” Here, the moving image performed an essentially symbolic function, concerned primarily with shaping the image of the weaponized animal bodies.

Despite their ability to crank out incredibly lifelike prose, generative AIs like Google's Bard or OpenAI's ChatGPT (powered by GPT-4), have already shown the current limitations of gen-AI technology as well as their own tenuous grasp of the facts — arguing that the JWST was the first telescope to image an exoplanet, and that Elvis' dad was an actor. But with this much market share at stake, what are a few misquoted facts against getting their product into the hands of consumers as quickly as possible? 

The team over at Anthropic, conversely, is made up largely of ex-OpenAI folks and they've taken a more pragmatic approach to the development of their own chatbot, Claude. The result is an AI that is "more steerable" and “much less likely to produce harmful outputs,” than ChatGPT, per a report from TechCrunch.

Claude has been in closed beta development since late 2022, but has recently begun testing the AI's conversational capabilities with launch partners including Robin AI, Quora and privacy-centered search engine, Duck Duck Go. The company has not released pricing yet but has confirmed to TC that two versions will be available at launch: the standard API and a faster, lightweight iteration they've dubbed Claude Instant.

“We use Claude to evaluate particular parts of a contract, and to suggest new, alternative language that’s more friendly to our customers,” Robin CEO Richard Robinson told TechCrunch. “We’ve found Claude is really good at understanding language — including in technical domains like legal language. It’s also very confident at drafting, summarizing, translations and explaining complex concepts in simple terms.”

Anthropic believes that Claude will be less likely to go rogue and start spitting racist obscenities like Tay did, in part, due to the AI's specialized training regimen that eh company is calling "constitutional AI." The company asserts that this provides a “principle-based” approach towards getting humans and robots on the same ethical page. Anthropic started with 10 foundational principles — though the company won't disclose what they are, specifically, which is 11-secret-herbs-and-spices of weird marketing stunt — suffice to say that, "they’re grounded in the concepts of beneficence, nonmaleficence and autonomy," per TC. 

The company then trained a separate AI to reliably generate text in accordance to those semi-secret principles by responding to myriad writing prompts like “compose a poem in the style of John Keats.” That model then trained Claude. But just because it is trained to be fundamentally less problematic than its competition doesn't mean Claude doesn't hallucinate facts like a startup CEO on an ayahuasca retreat. The AI has already invented a whole new chemical and taken artistic license to the uranium enrichment process; it has reportedly scored lower than ChatGPT on standardized tests for both math and grammar as well.

“The challenge is making models that both never hallucinate but are still useful — you can get into a tough situation where the model figures a good way to never lie is to never say anything at all, so there’s a tradeoff there that we’re working on,” the Anthropic spokesperson told TC. “We’ve also made progress on reducing hallucinations, but there is more to do."  

When Microsoft and OpenAI announced their renewed partnership in January, the two companies also revealed that Bing search would soon boast AI-enhanced lookup capabilities. Little did we know at the time, that Bing search has been powered for the past five weeks, not by the existing then-state-of-the-art GPT-3.5 model but by its even more robust successor, GPT-4.

Microsoft envisions Bing — and really Google is doing much the same with Bard — serving as a pseudo-gatekeeper to the rest of internet's information, not unlike what AOL's early America Online service once did. Rather than direct users to other websites where they can find the information and context they seek on their own, these companies are looking to have generative AI systems (Bard and Bing) automatically summarize and display that information without ever leaving the branded search page. Any additional relevant context that the user might have stumbled across during their independent research will similarly be deigned by the algorithm. Users can give GPT-4 a spin — and experience our new, algorithmically-dictated reality firsthand — by signing up for the Bing Preview waitlist. 

Hot on the heels of Google's Workspace AI announcement Tuesday, and ahead of Thursday's Microsoft Future of Work event, OpenAI has released the latest iteration of its generative pre-trained transformer system, GPT-4. Whereas the current generation GPT-3.5, which powers OpenAI's wildly popular ChatGPT conversational bot, can only read and respond with text, the new and improved GPT-4 will be able to generate text on input images as well. "While less capable than humans in many real-world scenarios," the OpenAI team wrote Tuesday, it "exhibits human-level performance on various professional and academic benchmarks."

OpenAI, which has partnered (and recently renewed its vows) with Microsoft to develop GPT's capabilities, has reportedly spent the past six months retuning and refining the system's performance based on user feedback generated from the recent ChatGPT hoopla. the company reports that GPT-4 passed simulated exams (such as the Uniform Bar, LSAT, GRE, and various AP tests) with a score "around the top 10 percent of test takers" compared to GPT-3.5 which scored in the bottom 10 percent. What's more, the new GPT has outperformed other state-of-the-art large language models (LLMs) in a variety of benchmark tests. The company also claims that the new system has achieved record performance in "factuality, steerability, and refusing to go outside of guardrails" compared to its predecessor.

OpenAI says that the GPT-4 will be made available for both ChatGPT and the API. "GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5," the OpenAI team wrote. 

Developing...

Google has been scrambling to catch up to to OpenAI for months, ever since the latter dropped its conversational bot, ChatGPT, and took the generative AI industry by storm. Google's first attempted response with the release of its Bard AI (which immediately misquoted easily verifiable stats about the JWST) was tepid at best so the company has announced a new tact: they're packing every single product they can with AI — just like they did in the Google+ era with social features.

The new features will be coming to virtually all of Google's Workspace products. According to the company users will be able to "draft, reply, summarize, and prioritize" emails, "brainstorm, proofread, write, and rewrite" text documents, autogenerate images and even video with Slides, have Sheets create formulas autonomously, automate transcription notes in Meet and "enable workflows for getting things done" in Chat. 

For example, in Docs, users will simply need to type the subject of their assignment into the page to have Google's generative AI suite quickly gin up additional text. The system can also rework (hopefully improving) what the user has already drafted, even if they're just bullet points, using the Rewrite function. There's also a new “I’m feeling lucky” option in Gmail which your company's HR department is just going to adore.

Following the Bard debacle, Google doubled down on its commitment to ensuring that its AIs don't turn out like Miucrosoft's. "AI is no replacement for the ingenuity, creativity, and smarts of real people," Johanna Voolich Wright, VP of Product at Google Workspace, wrote on Tuesday. "Sometimes the AI gets things wrong, sometimes it delights you with something offbeat, and oftentimes it requires guidance." To that end the company is building its products within the bounds of its AI Principles, which are as legally binding as the company's old "Don't be evil" motto. The new AI-enabled Workspace suite is expected to roll out to English language users in the US by the end of the month with additional languages and regions arriving in the near future.     

Science is the reason you aren't reading this by firelight nestled cozily under a rock somewhere however, its practice significantly predates its formalization by Galileo in the 16th century. Among its earliest adherents — even before pioneering efforts of Aristotle — was Animaxander, the Greek philosopher credited with first arguing that the Earth exists within a void, not atop a giant turtle shell. His other revolutionary notions include, "hey, maybe animals evolved from other, earlier animals?" and "the gods aren't angry, that's just thunder."

While Animaxander isn't often mentioned alongside the later greats of Greek philosophy, his influence on the scientific method cannot be denied, argues NYT bestselling author, Carlo Rovelli, in his latest book, Animaxander and the Birth of Science, out now from Riverhead Books. In in, Rovelli celebrates Animaxander, not necessarily for his scientific acumen but for his radical scientific thinking — specifically his talent for shrugging off conventional notion to glimpse at the physical underpinnings of the natural world. In the excerpt below, Rovelli, whom astute readers will remember from last year's There Are Places in the World Where Rules Are Less Important than Kindness, illustrates how even the works of intellectual titans like Einstein and Heisenberg can and inevitably are found lacking in their explanation of natural phenomena — in just the same way that those works themselves decimated the collective understanding of cosmological law under 19th century Newtonian physics.   

Riverhead Books

Excerpted from Animaxander and the Birth of Science. Copyright © 2023 by Carlo Rovelli. Excerpted by permission of Riverhead, an imprint and division of Penguin Random House LLC, New York. All rights reserved. No part of this excerpt may be reproduced or reprinted without permission in writing from the publisher.

Did science begin with Anaximander? The question is poorly put. It depends on what we mean by “science,” a generic term. Depending on whether we give it a broad or a narrow meaning, we can say that science began with Newton, Galileo, Archimedes, Hipparchus, Hippocrates, Pythagoras, or Anaximander — or with an astronomer in Babylonia whose name we don’t know, or with the first primate who managed to teach her offspring what she herself had learned, or with Eve, as in the quotation that opens this chapter. Historically or symbolically, each of these moments marks humanity’s acquisition of a new, crucial tool for the growth of knowledge.

If by “science” we mean research based on systematic experimental activities, then it began more or less with Galileo. If we mean a collection of quantitative observations and theoretical/mathematical models that can order these observations and give accurate predictions, then the astronomy of Hipparchus and Ptolemy is science. Emphasizing one particular starting point, as I have done with Anaximander, means focusing on a specific aspect of the way we acquire knowledge. It means highlighting specific characteristics of science and thus, implicitly, reflecting on what science is, what the search for knowledge is, and how it works.

What is scientific thinking? What are its limits? What is the reason for its strength? What does it really teach us? What are its characteristics, and how does it compare with other forms of knowledge?

These questions shaped my reflections on Anaximander in preceding chapters. In discussing how Anaximander paved the way for scientific knowledge, I highlighted a certain number of aspects of science itself. Now I shall make these observations more explicit.

The Crumbling of Nineteenth Century Illusions

A lively debate on the nature of scientific knowledge has taken place during the last century. The work of philosophers of science such as Carnap and Bachelard, Popper and Kuhn, Feyerabend, Lakatos, Quine, van Fraassen, and many others has transformed our understanding of what constitutes scientific activity. To some extent, this reflection was a reaction to a shock: the unexpected collapse of Newtonian physics at the beginning of the twentieth century.

In the nineteenth century, a common joke was that Isaac New‐ ton had been not only one of the most intelligent men in human history, but also the luckiest, because there is only one collection of fundamental natural laws, and Newton had had the good fortune to be the one to discover them. Today we can’t help but smile at this notion, because it reveals a serious epistemological error on the part of nineteenth-​­century thinkers: the idea that good scientific theories are definitive and remain valid until the end of time.

The twentieth century swept away this facile illusion. Highly accurate experiments showed that Newton’s theory is mistaken in a very precise sense. The planet Mercury, for example, does not move following Newtonian laws. Albert Einstein, Werner Heisenberg, and their colleagues discovered a new collection of fundamental laws — general relativity and quantum mechanics — that replace Newton’s laws and work well in the domains where Newton’s theory breaks down, such as accounting for Mercury’s orbit, or the behavior of electrons in atoms.

Once burned, twice shy: few people today believe that we now possess definitive scientific laws. It is generally expected that one day Einstein’s and Heisenberg’s laws will show their limits as well, and will be replaced by better ones. In fact, the limits of Einstein’s and Heisenberg’s theories are already emerging. There are subtle incompatibilities between Einstein’s theory and Heisenberg’s, which make it unreasonable to suppose that we have identified the final, definitive laws of the universe. As a result, research goes on. My own work in theoretical physics is precisely the search for laws that might combine these two theories.

Now, the essential point here is that Einstein’s and Heisenberg’s theories are not minor corrections to Newton’s. The differences go far beyond an adjusted equation, a tidying up, the addition or replacement of a formula. Rather, these new theories constitute a radical rethinking of the world. Newton saw the world as a vast empty space where “particles” move about like pebbles. Einstein understands that such supposedly empty space is in fact a kind of storm-​­tossed sea. It can fold in on itself, curve, and even (in the case of black holes) shatter. No one had seriously contemplated this possibility before. For his part, Heisenberg understands that Newton’s “particles” are not particles at all but bizarre hybrids of particles and waves that run over Faraday lines’ webs. In short, over the course of the twentieth century, the world was found to be profoundly different from the way Newton imagined it.

On the one hand, these discoveries confirmed the cognitive strength of science. Like Newton’s and Maxwell’s theories in their day, these discoveries led quickly to an astonishing development of new technologies that once again radically changed human society. The insights of Faraday and Maxwell brought about radio and communications technology. Einstein’s and Heisenberg’s led to computers, information technology, atomic energy, and countless other technological advances that have changed our lives.

But on the other hand, the realization that Newton’s picture of the world was false is disconcerting. After Newton, we thought we had understood once and for all the basic structure and functioning of the physical world. We were wrong. The theories of Einstein and Heisenberg themselves will one day likely be proved false. Does this mean that the understanding of the world offered by science cannot be trusted, not even for our best science? What, then, do we really know about the world? What does science teach us about the world?

Amazon’s evolution from omnipresent shopping platform to omnipresent surveillance platform continues apace, having drastically expanded its line of Ring security camera systems in recent years. Nowadays, the company offers video doorbells, exterior cameras, interior cameras, flying cameras, lighting systems, alarm systems and vehicle security packages — the lattermost of which is why we are here today. I put a Ring camera in my car.

That’s not to say that the Ring Car Camera is a poorly designed or manufactured product — far from it! The $250 Car Camera features dual-facing (pointing both at the road and into the cabin), IR-capable 1080p imaging sensors, optional LTE connectivity, Alexa-driven voice commands, and remote vehicle monitoring through the Ring mobile app.

In fact, I was surprised by how quickly and easily I was able to get the system set up. The camera assembly itself is a single piece that wedges into the bottom edge of the windshield-dashboard horizon and sticks to the glass with a high-strength adhesive. It’s not strong enough to keep a car thief from yanking it out but it’ll keep the camera in place as you travel over and through America’s crumbling highway infrastructure system. One sticking point I could see arising is the camera needs access to a home wi-fi connection during the setup/app pairing sequence. I was able to pull around my driveway until I was at the exterior of the wall from my house’s router but if you live in an apartment complex, things might get dicey.

“If you are unable to connect to your home Wi-Fi during setup, you can set up the device using LTE through Ring Protect Go,” a Ring rep told Engadget. “Just skip the ‘set up with Wi-Fi’ step in the set up flow and follow the on-screen instructions. Each new customer will have a free 30-day trial of Ring Protect Go, which provides LTE connectivity.”

I was not at all a fan of the camera’s wired power connection to the vehicle’s OBDII port, which also monitors the battery’s voltage so that the camera can turn itself off before fully exhausting the power supply. For one, that physical requirement limits the vehicles this system can work with to only those with OBD ports located to the left of the steering column. For another, I now have a 6-foot long cable snaking its way across my previously immaculate dashboard, draping down my driver side door panel to connect with the OBDII port just over my brake pedal. Even with the included 3M-adhesive cable stays (which, I might add, were immediately foiled by the tiny crags and crinkles of my dashboard’s surface), I can hear the cord shifting and sliding around during turns, I’m constantly aware of it when swinging my legs out of the car lest I accidentally catch it on a toe and tear the connector from the port. Which I did the very first time I drove after installation — and then the next as well.

Andrew Tarantola / Engadget

The other issue is that not every car has an OBDII port located in the passenger cabin and for those vehicles the Ring Car Cam will not work. Neither will any of the vehicles on this rather expansive list of incompatible models for one reason or another — some cut power to the port when the key is removed and Teslas, for their part, don’t even use the OBDII interface. What’s more, if your dongle is already in use, whether for an insurance tracker or an interlock device, you’re SOL with using the Car Cam. Same if you buy it in a jurisdiction that limits the use of dashcams — except then you also go to jail.

Andrew Tarantola / Engadget

At 1080p, the Car Cam’s video quality is just fine for what the average driver would presumably be using it for and the interior-facing IR sensor will ensure that you get a good look at whoever’s rifling through your center console at three in the morning. But, since it’s mounted on the dash itself and not suspended from the rearview mirror like the commercial-grade ones you find in Ubers and Lyfts, you won’t get much of a view of the interior below chest level. Accessing those videos takes a hot second as well, as the clips aren’t transferred directly to your phone (if using Protect Go). They have to first be uploaded and processed in the cloud before you can watch them.

The camera offers a variety of recording options. You can set it for continuous use, as you would a traditional dash cam — and if you don’t want it recording you, the unit thankfully incorporates a physical lens cover for the interior-facing camera. You can also use it specifically for traffic stops with the verbal “Alexa, Record” command, in which the system will record uninterrupted for 20 minutes even after the ignition has been turned off. Finally, there’s Parking Protection mode that activates the camera if it detects motion or an impact when the vehicle is parked.

Andrew Tarantola / Engadget

All of the recorded data — up to seven hours worth — is saved locally on the device and made available once the camera is back in range of a Wi-Fi connection. Again, that’s not great if a thief or cop rips out the unit before the information can be uploaded. Also, there’s no loop recording so if something important happens when you’ve got 6 hours and 56 minutes of video already saved, you better hope the matter resolves itself in under 4 minutes, otherwise the recording will simply be cut off.

To get around that, you’ll need cloud access and to spend $6 a month (or $60 a year) for the Ring Protect Go subscription service for it. In addition, Protect Go unlocks access to the camera’s onboard LTE connection enabling two-way view and talk, notifications and GPS tracking from anywhere with cell service. Without that subscription access, those features are only available over Wi-Fi.

Andrew Tarantola / Engadget

Ring’s business decisions have made very clear that it is on the side of the police — even if the homeowners themselves aren’t — freely volunteering data to, and often partnering with, law enforcement agencies around the country. When asked whether safeguards have been put in place to prevent law enforcement from surreptitiously spying through the Car Cam, Ring’s spokesperson noted, “Ring builds products and services for our customers, not law enforcement. When parked, Car Cam only records when the smart sensors detect an incident (such as a collision or broken window) or if the device owner or Shared User initiates Live View.” What happens to that data once it's off the device and in Ring’s cloud servers was not made clear.

Even if I could put aside Ring’s cozy relationship with police, $250 for what the Car Cam offers is a big ask, especially with that $6-a-month cherry on top to get anything to work outside of your driveway. Granted, if you’re already part of the Ring ecosystem, you like what it offers and want to extend that platform to your vehicle, absolutely give the Car Cam a shot. But if you’re in the market for a standalone vehicle security system, there are plenty of options available to choose from that offer many of the same features as the Ring at a fraction of the price and without the baggage — or that blasted power cable.