Depending on the hardware you're using, training a large language model of any significant size can take weeks, months, even years to complete. That's no way to do business — nobody has the electricity and time to be waiting that long. On Wednesday, NVIDIA unveiled the newest iteration of its Eos supercomputer, one powered by more than 10,000 H100 Tensor Core GPUs and capable of training a 175 billion-parameter GPT-3 model on 1 billion tokens in under four minutes. That's three times faster than the previous benchmark on the MLPerf AI industry standard, which NVIDIA set just six months ago.

Eos represents an enormous amount of compute. It leverages 10,752 GPUs strung together using NVIDIA's Infiniband networking (moving a petabyte of data a second) and 860 terabytes of high bandwidth memory (36PB/sec aggregate bandwidth and 1.1PB sec interconnected) to deliver 40 exaflops of AI processing power. The entire cloud architecture is comprised of 1344 nodes — individual servers that companies can rent access to for around $37,000 a month to expand their AI capabilities without building out their own infrastructure. 

In all, NVIDIA set six records in nine benchmark tests: the 3.9 minute notch for GPT-3, a 2.5 minute mark to to train a Stable Diffusion model using 1,024 Hopper GPUs, a minute even to train DLRM, 55.2 seconds for RetinaNet, 46 seconds for 3D U-Net and the BERT-Large model required just 7.2 seconds to train.

NVIDIA was quick to note that the 175 billion parameter version of GPT-3 used in the benchmarking is not the full-sized iteration of the model (neither was the Stable Diffusion model). The larger GPT-3 offers around 3.7 trillion parameters and is just flat out too big and unwieldy for use as a benchmarking test. For example, it'd take 18 months to train it on the older A100 system with 512 GPUs — though, Eos needs just eight days. 

So instead, NVIDIA and MLCommons, which administers the MLPerf standard, leverage a more compact version that uses 1 billion tokens (the smallest denominator unit of data that generative AI systems understand). This test uses a GPT-3 version with the same number of potential switches to flip (s the full-size (those 175 billion parameters), just a much more manageable data set to use in it (a billion tokens vs 3.7 trillion).

The impressive improvement in performance, granted, came from the fact that this recent round of tests employed 10,752 H100 GPUs compared to the 3,584 Hopper GPUs the company used in June's benchmarking trials. However NVIDIA explains that despite tripling the number of GPUs, it managed to maintain 2.8x scaling in performance — an 93 percent efficiency rate — through the generous use of software optimization.

"Scaling is a wonderful thing," Salvator said."But with scaling, you're talking about more infrastructure, which can also mean things like more cost. An efficiently scaled increase means users are "making the best use of your of your infrastructure so that you can basically just get your work done as fast [as possible] and get the most value out of the investment that your organization has made."

The chipmaker was not alone in its development efforts. Microsoft's Azure team submitted a similar 10,752 H100 GPU system for this round of benchmarking, and achieved results within two percent of NVIDIA's.

"[The Azure team have] been able to achieve a performance that's on par with the Eos supercomputer," Dave Salvator Director of Accelerated Computing Products at NVIDIA, told reporters during a Tuesday prebrief. What's more "they are using Infiniband, but this is a commercially available instance. This isn't some pristine laboratory system that will never have actual customers seeing the benefit of it. This is the actual instance that Azure makes available to its customers."

 NVIDIA plans to apply these expanded compute abilities to a variety of tasks, including the company's ongoing work in foundational model development, AI-assisted GPU design, neural rendering, multimodal generative AI and autonomous driving systems.

"Any good benchmark looking to maintain its market relevance has to continually update the workloads it's going to throw at the hardware to best reflect the market it's looking to serve," Salvator said, noting that MLCommons has recently added an additional benchmark for testing model performance on Stable Diffusion tasks. "This is another exciting area of generative AI where we're seeing all sorts of things being created" — from programming code to discovering protein chains.

These benchmarks are important because, as Salvator points out, the current state of generative AI marketing can a bit of a "Wild West." The lack of stringent oversight and regulation means, "we sometimes see with certain AI performance claims where you're not quite sure about all the parameters that went into generating those particular claims." MLPerf provides the professional assurance that the benchmark numbers companies generate using its tests "were reviewed, vetted, in some cases even challenged or questioned by other members of the consortium," Salvator said. "It's that sort of peer reviewing process that really brings credibility to these results."

NVIDIA has been steadily focusing on its AI capabilities and applications in recent months. "We are at the iPhone moment for AI," CEO Jensen Huang said during his GTC keynote in March. At that time the company announced its DGX cloud system which portions out slivers of the supercomputer's processing power — specifically by either eight H100 or A100 chips running 60GB of VRAM (640 of memory in total). The company expanded its supercomputing portfolio with the release of DGX GH200 at Computex in May.

Facebook is no stranger to moderating and mitigating misinformation on its platform, having long employed machine learning and artificial intelligence systems to help supplement its human-led moderation efforts. At the start of October, the company extended its machine learning expertise to its advertising efforts with an experimental set of generative AI tools that can perform tasks like generating backgrounds, adjusting image and creating captions for an advertiser's video content. Reuters reports Monday that Meta will specifically not make those tools available to political marketers ahead of what is expected to be a brutal and divisive national election cycle. 

Meta's decision to bar the use of generative AI is in line with much of the social media ecosystem, though, as Reuters is quick to point out, the company, "has not yet publicly disclosed the decision in any updates to its advertising standards." TikTok and Snap both ban political ads on their networks, Google employs a "keyword blacklist" to prevent its generative AI advertising tools from straying into political speech and X (formerly Twitter) is, well, you've seen it. 

Meta does allow for a wide latitude of exceptions to this rule. The tool ban only extends to "misleading AI-generated video in all content, including organic non-paid posts, with an exception for parody or satire," per Reuters. Those exceptions are currently under review by the company's independent Oversight Board as part of a case in which Meta left up an "altered" video of President Biden because, the company argued, it was not generated by an AI.

Facebook, along with other leading Silicon Valley AI companies, agreed in July to voluntary commitments set out by the White House enacting technical and policy safeguards in the development of their future generative AI systems. Those include expanding adversarial machine learning (aka red-teaming) efforts to root out bad model behavior, sharing trust and safety information both within the industry and with the government, as well as development of a digital watermarking scheme to authenticate official content and make clear that it is not AI-generated. 

It’s been nearly a year since ChatGPT’s public debut and its evolution since then has been nothing short of extraordinary. In just over 11 months, OpenAI’s chatbot has gained the ability to write programming code, process information between multiple modalities and expand its reach across the internet with APIs. During OpenAI’s 2023 Dev Day keynote address Monday, CEO Sam Altman and other executives took to the stage in San Francisco to unveil the chatbot’s latest iteration, ChatGPT-4 Turbo, as well as an exciting new way to bring generative AI technology to everybody, regardless of their coding capability: GPTs!

GPTs are small, task-specific iterations of ChatGPT. Think of them like the single-purpose apps and features on your phone but instead of them maintaining a timer or stop watch, or a digital assistant transcribing your voice instructions into a shopping list, GPTs will do, basically anything you train them to. OpenAI offers up eight examples of what GPT’s can be used for, anything from a digital kitchen assistant that suggests recipes based on whats in your pantry to a math mentor to help your kids through their homework to a Sticker Wiz that will, “turn your wildest dreams into die-cut stickers, shipped right to your door.”

The new GPTs are an expansion on the company’s existing Custom Instructions feature which debuted in July. OpenAI notes that many of its power users were already recycling and updating their most effective prompts and instruction sets, a process which GPT-4 Turbo will now handle automatically as part of its update to seed parameters and focus on reproducible outputs. This will allow users a far greater degree of control in customizing the GPTs to their specific needs.

What users won’t need is an extensive understanding of javascript programming. With GPT-4 Turbo’s improved code interpretation, retrieval and function calling capabilities, as well as its massively increased context window size, users will be able to devise and develop their GPTs using nothing but natural language.

Any GPT created by the community will be immediately shareable. For now that will happen directly between users but, later this month, OpenAI plans to launch a centralized storefront where “verified builders” can post and share their GPTs. The most popular ones will climb a leaderboard and potentially, eventually earn their creators money based on how many people are using the GPT.

GPTs will be available to both regular users and enterprise accounts which, like ChatGPT Enterprise that came out earlier this year, will offer institutional users the chance to create their own internal-only, admin-approved mini-chatbots. These will work with (and are trained on) the company’s specific tasks, department documentation or proprietary datasets. Enterprise GPTs arrive for those customers on Wednesday.

Privacy remains a focal point for the company with additional technical safeguards being put into place, atop existing moderation systems, to prevent people from making GPTs that go against OpenAI’s usage policies. The company is also rolling out an identity verification system for developers to help improve transparency and trust, but did not elaborate on what that process could entail.

The internet's "enshittification," as veteran journalist and privacy advocate Cory Doctorow describes it, began decades before TikTok made the scene. Elder millennials remember the good old days of Napster — followed by the much worse old days of Napster being sued into oblivion along with Grokster and the rest of the P2P sharing ecosystem, until we were left with a handful of label-approved, catalog-sterilized streaming platforms like Pandora and Spotify. Three cheers for corporate copyright litigation.

In his new book The Internet Con: How to Seize the Means of Computation, Doctorow examines the modern social media landscape, cataloging and illustrating the myriad failings and short-sighted business decisions of the Big Tech companies operating the services that promised us the future but just gave us more Nazis. We have both an obligation and responsibility to dismantle these systems, Doctorow argues, and a means to do so with greater interoperability. In this week's Hitting the Books excerpt, Doctorow examines the aftermath of the lawsuits against P2P sharing services, as well as the role that the Digital Millennium Copyright Act's "notice-and-takedown" reporting system and YouTube's "ContentID" scheme play on modern streaming sites.

Verso Publishing

Excerpted from by The Internet Con: How to Seize the Means of Computation by Cory Doctorow. Published by Verso. Copyright © 2023 by Cory Doctorow. All rights reserved.

Seize the Means of Computation

The harms from notice-and-takedown itself don’t directly affect the big entertainment companies. But in 2007, the entertainment industry itself engineered a new, more potent form of notice-and-takedown that manages to inflict direct harm on Big Content, while amplifying the harms to the rest of us. 

That new system is “notice-and-stay-down,” a successor to notice-and-takedown that monitors everything every user uploads or types and checks to see whether it is similar to something that has been flagged as a copyrighted work. This has long been a legal goal of the entertainment industry, and in 2019 it became a feature of EU law, but back in 2007, notice-and-staydown made its debut as a voluntary modification to YouTube, called “Content ID.” 

Some background: in 2007, Viacom (part of CBS) filed a billion-dollar copyright suit against YouTube, alleging that the company had encouraged its users to infringe on its programs by uploading them to YouTube. Google — which acquired YouTube in 2006 — defended itself by invoking the principles behind Betamax and notice-and-takedown, arguing that it had lived up to its legal obligations and that Betamax established that “inducement” to copyright infringement didn’t create liability for tech companies (recall that Sony had advertised the VCR as a means of violating copyright law by recording Hollywood movies and watching them at your friends’ houses, and the Supreme Court decided it didn’t matter). 

But with Grokster hanging over Google’s head, there was reason to believe that this defense might not fly. There was a real possibility that Viacom could sue YouTube out of existence — indeed, profanity-laced internal communications from Viacom — which Google extracted through the legal discovery process — showed that Viacom execs had been hotly debating which one of them would add YouTube to their private empire when Google was forced to sell YouTube to the company. 

Google squeaked out a victory, but was determined not to end up in a mess like the Viacom suit again. It created Content ID, an “audio fingerprinting” tool that was pitched as a way for rights holders to block, or monetize, the use of their copyrighted works by third parties. YouTube allowed large (at first) rightsholders to upload their catalogs to a blocklist, and then scanned all user uploads to check whether any of their audio matched a “claimed” clip. 

Once Content ID determined that a user was attempting to post a copyrighted work without permission from its rightsholder, it consulted a database to determine the rights holder’s preference. Some rights holders blocked any uploads containing audio that matched theirs; others opted to take the ad revenue generated by that video. 

There are lots of problems with this. Notably, there’s the inability of Content ID to determine whether a third party’s use of someone else’s copyright constitutes “fair use.” As discussed, fair use is the suite of uses that are permitted even if the rightsholder objects, such as taking excerpts for critical or transformational purposes. Fair use is a “fact intensive” doctrine—that is, the answer to “Is this fair use?” is almost always “It depends, let’s ask a judge.” 

Computers can’t sort fair use from infringement. There is no way they ever can. That means that filters block all kinds of legitimate creative work and other expressive speech — especially work that makes use of samples or quotations. 

But it’s not just creative borrowing, remixing and transformation that filters struggle with. A lot of creative work is similar to other creative work. For example, a six-note phrase from Katy Perry’s 2013 song “Dark Horse” is effectively identical to a six-note phrase in “Joyful Noise,” a 2008 song by a much less well-known Christian rapper called Flame. Flame and Perry went several rounds in the courts, with Flame accusing Perry of violating his copyright. Perry eventually prevailed, which is good news for her. 

But YouTube’s filters struggle to distinguish Perry’s six-note phrase from Flame’s (as do the executives at Warner Chappell, Perry’s publisher, who have periodically accused people who post snippets of Flame’s “Joyful Noise” of infringing on Perry’s “Dark Horse”). Even when the similarity isn’t as pronounced as in Dark, Joyful, Noisy Horse, filters routinely hallucinate copyright infringements where none exist — and this is by design. 

To understand why, first we have to think about filters as a security measure — that is, as a measure taken by one group of people (platforms and rightsholder groups) who want to stop another group of people (uploaders) from doing something they want to do (upload infringing material). 

It’s pretty trivial to write a filter that blocks exact matches: the labels could upload losslessly encoded pristine digital masters of everything in their catalog, and any user who uploaded a track that was digitally or acoustically identical to that master would be blocked. 

But it would be easy for an uploader to get around a filter like this: they could just compress the audio ever-so-slightly, below the threshold of human perception, and this new file would no longer match. Or they could cut a hundredth of a second off the beginning or end of the track, or omit a single bar from the bridge, or any of a million other modifications that listeners are unlikely to notice or complain about. 

Filters don’t operate on exact matches: instead, they employ “fuzzy” matching. They don’t just block the things that rights holders have told them to block — they block stuff that’s similar to those things that rights holders have claimed. This fuzziness can be adjusted: the system can be made more or less strict about what it considers to be a match. 

Rightsholder groups want the matches to be as loose as possible, because somewhere out there, there might be someone who’d be happy with a very fuzzy, truncated version of a song, and they want to stop that person from getting the song for free. The looser the matching, the more false positives. This is an especial problem for classical musicians: their performances of Bach, Beethoven and Mozart inevitably sound an awful lot like the recordings that Sony Music (the world’s largest classical music label) has claimed in Content ID. As a result, it has become nearly impossible to earn a living off of online classical performance: your videos are either blocked, or the ad revenue they generate is shunted to Sony. Even teaching classical music performance has become a minefield, as painstakingly produced, free online lessons are blocked by Content ID or, if the label is feeling generous, the lessons are left online but the ad revenue they earn is shunted to a giant corporation, stealing the creative wages of a music teacher.

Notice-and-takedown law didn’t give rights holders the internet they wanted. What kind of internet was that? Well, though entertainment giants said all they wanted was an internet free from copyright infringement, their actions — and the candid memos released in the Viacom case — make it clear that blocking infringement is a pretext for an internet where the entertainment companies get to decide who can make a new technology and how it will function.

Nearly two years after its prototype debut and eight months after its public beta, Sony's GT Sophy racing AI for Gran Turismo 7 is back, and going by Gran Turismo Sophy 2.0 now. It will be available to all PlayStation 5 users as part of the GT7 Spec II Update (Patch Update 1.40) being released on Wednesday, November 2 at 2 a.m. ET. 

We got our first look at the Sophy system back in February 2022. At that point it was already handily beating professional Gran Turismo players. “Gran Turismo Sophy is a significant development in AI whose purpose is not simply to be better than human players, but to offer players a stimulating opponent that can accelerate and elevate the players’ techniques and creativity to the next level,” Sony AI CEO, Hiroaki Kitano, said at the time. “In addition to making contributions to the gaming community, we believe this breakthrough presents new opportunities in areas such as autonomous racing, autonomous driving, high-speed robotics and control.”

The system's public beta this past February saw the AI competing against a small subset of the game's user base in the “Gran Turismo Sophy Race Together” event. Players who had already progressed sufficiently through the game were granted access to the special race, where they faced off against four AI-controlled opponents in a limited number of tracks. 

“The difference [between racers] is that, it's essentially the power you have versus the other cars on the track,” Sony AI's COO, Michael Spranger, told Engadget in February. “You have different levels of performance. In the beginning level, you have a much more powerful vehicle — still within the same class, but you're much faster [than your competition].” That advantage shrank as players advanced through the race rounds and Sophy gained access to increasingly capable vehicles. In September, Sophy learned to drift.

“We have evolved GT Sophy from a research project tackling the grand challenge of creating an AI agent that could outperform top drivers in a top simulation racing game, to a functional game feature that provides all game players a formidable, human-like opponent that enhances the overall racing experience," Spranger said in a press statement released Wednesday.

With Wednesday's announcement, the number of vehicles Sophy can pilot rises from the meager four models available during the beta event, to 340 (yes, three hundred and forty) vehicles across nine unique tracks. Per Sony, that means Sophy can drive 95 percent of the playable in-game models and will select its car for the race based on what the player has available in their garage (that way they're not randomly facing down a 918 in a Nissan Versa or are otherwise disadvantaged).

Players can match against Sophy in Quick Race mode (formerly "Arcade") regardless of their advancement through the game or current skill level. As long as you have a PS5, a network connection and the latest update patch installed, you too can get Toretto'ed by a stack of algorithmic processes. Good luck.

Just days after President Joe Biden unveiled a sweeping executive order retasking the federal government with regards to AI development, Vice President Kamala Harris announced at the UK AI Safety Summit on Tuesday a half dozen more machine learning initiatives that the administration is undertaking. Among the highlights: the establishment of the United States AI Safety Institute, the first release of draft policy guidance on the federal government's use of AI and a declaration on the responsible military applications for the emerging technology.

"President Biden and I believe that all leaders, from government, civil society, and the private sector have a moral, ethical, and societal duty to make sure AI is adopted and advanced in a way that protects the public from potential harm and ensures that everyone is able to enjoy its benefits,” Harris said in her prepared remarks.

"Just as AI has the potential to do profound good, it also has the potential to cause profound harm, from AI-enabled cyber-attacks at a scale beyond anything we have seen before to AI-formulated bioweapons that could endanger the lives of millions," she said. The existential threats that generative AI systems present was a central theme of the summit. 

"To define AI safety we must consider and address the full spectrum of AI risk — threats to humanity as a whole, threats to individuals, to our communities and to our institutions, and threats to our most vulnerable populations," she continued. "To make sure AI is safe, we must manage all these dangers."

To that end, Harris announced Wednesday that the White House, in cooperation with the Department of Commerce, is establishing the United States AI Safety Institute (US AISI) within the NIST. It will be responsible for actually creating and publishing the all of the guidelines, benchmark tests, best practices and such for testing and evaluating potentially dangerous AI systems. 

These tests could include the red-team exercises that President Biden had mentioned in his EO. The AISI would also be tasked in providing technical guidance to lawmakers and law enforcement on a wide range of AI-related topics, including identifying generated content, authenticating live-recorded content, mitigating AI-driven discrimination, and ensuring transparency in its use.

Additionally, the Office of Management and Budget (OMB) is set to release for public comment the administration's first draft policy guidance on government AI use later this week. Like the Blueprint for an AI Bill of Rights that it builds upon, the draft policy guidance outlines steps that the national government can take to "advance responsible AI innovation" while maintaining transparency and protecting federal workers from increased surveillance and job displacement. This draft guidance will eventually be used to establish safeguards for the use of AI in a broad swath of public sector applications including transportation, immigration, health and education so it is being made available for public comment at ai.gov/input.

Harris also announced during her remarks that the Political Declaration on the Responsible Use of Artificial Intelligence and Autonomy the US issued in February has collected 30 signatories to date, all of whom have agreed to a set of norms for responsible development and deployment of military AI systems. Just 165 nations to go! The administration is also launching a a virtual hackathon in efforts to blunt the harm AI-empowered phone and internet scammers can inflict. Hackathon participants will work to build AI models that can counter robocalls and robotexts, especially those targeting elderly folks with generated voice scams.

Content authentication is a growing focus of the Biden-Harris administration. President Biden's EO explained that the Commerce Department will be spearheading efforts to validate content produced by the White House through a collaboration with the C2PA and other industry advocacy groups. They'll work to establish industry norms, such as the voluntary commitments previously extracted from 15 of the largest AI firms in Silicon Valley. In her remarks, Harris extended that call internationally, asking for support from all nations in developing global standards in authenticating government-produced content. 

“These voluntary [company] commitments are an initial step toward a safer AI future, with more to come," she said. "As history has shown in the absence of regulation and strong government oversight, some technology companies choose to prioritize profit over: The wellbeing of their customers; the security of our communities; and the stability of our democracies."

"One important way to address these challenges — in addition to the work we have already done — is through legislation — legislation that strengthens AI safety without stifling innovation," Harris continued. 

The Biden Administration unveiled its ambitious next steps in addressing and regulating artificial intelligence development on Monday. Its expansive new executive order seeks to establish further protections for the public as well as improve best practices for federal agencies and their contractors.

"The President several months ago directed his team to pull every lever," a senior administration official told reporters on a recent press call. "That's what this order does, bringing the power of the federal government to bear in a wide range of areas to manage AI's risk and harness its benefits ... It stands up for consumers and workers, promotes innovation and competition, advances American leadership around the world and like all executive orders, this one has the force of law."

These actions will be introduced over the next year with smaller safety and security changes happening in around 90 days and with more involved reporting and data transparency schemes requiring 9 to 12 months to fully deploy. The administration is also creating an “AI council,” chaired by White House Deputy Chief of Staff Bruce Reed, who will meet with federal agency heads to ensure that the actions are being executed on schedule.

ASSOCIATED PRESS

Public Safety

"In response to the President's leadership on the subject, 15 major American technology companies have begun their voluntary commitments to ensure that AI technology is safe, secure and trustworthy before releasing it to the public," the senior administration official said. "That is not enough."

The EO directs the establishment of new standards for AI safety and security, including reporting requirements for developers whose foundation models might impact national or economic security. Those requirements will also apply in developing AI tools to autonomously implement security fixes on critical software infrastructure. 

By leveraging the Defense Production Act, this EO will "require that companies developing any foundation model that poses a serious risk to national security, national economic security, or national public health and safety must notify the federal government when training the model, and must share the results of all red-team safety tests," per a White House press release. That information must be shared prior to the model being made available to to the public, which could help reduce the rate at which companies unleash half-baked and potentially deadly machine learning products.

In addition to the sharing of red team test results, the EO also requires disclosure of the system’s training runs (essentially, its iterative development history). “What that does is that creates a space prior to the release… to verify that the system is safe and secure,” officials said.

Administration officials were quick to point out that this reporting requirement will not impact any AI models currently available on the market, nor will it impact independent or small- to medium-size AI companies moving forward, as the threshold for enforcement is quite high. It's geared specifically for the next generation of AI systems that the likes of Google, Meta and OpenAI are already working on with enforcement on models starting at 10^26 petaflops, a capacity currently beyond the limits of existing AI models. "This is not going to catch AI systems trained by graduate students, or even professors,” the administration official said.

What's more, the EO will encourage the Departments of Energy and Homeland Security to address AI threats "to critical infrastructure, as well as chemical, biological, radiological, nuclear, and cybersecurity risks," per the release. "Agencies that fund life-science projects will establish these standards as a condition of federal funding, creating powerful incentives to ensure appropriate screening and manage risks potentially made worse by AI." In short, any developers found in violation of the EO can likely expect a prompt and unpleasant visit from the DoE, FDA, EPA or other applicable regulatory agency, regardless of their AI model’s age or processing speed.

In an effort to proactively address the decrepit state of America's digital infrastructure, the order also seeks to establish a cybersecurity program, based loosely on the administration's existing AI Cyber Challenge, to develop AI tools that can autonomously root out and shore up security vulnerabilities in critical software infrastructure. It remains to be seen whether those systems will be able to address the concerns of misbehaving models that SEC head Gary Gensler recently raised.

AI Watermarking and Cryptographic Validation

We're already seeing the normalization of deepfake trickery and AI-empowered disinformation on the campaign trail. So, the White House is taking steps to ensure that the public can trust the text, audio and video content that it publishes on its official channels. The public must be able to easily validate whether the content they see is AI-generated or not, argued White House officials on the press call. 

Adobe

The Department of Commerce is in charge of the latter effort and is expected to work closely with existing industry advocacy groups like the C2PA and its sister organization, the CAI, to develop and implement a watermarking system for federal agencies. “We aim to support and facilitate and help standardize that work [by the C2PA],” administration officials said. “We see ourselves as plugging into that ecosystem.”

Officials further explained that the government is supporting the underlying technical standards and practices that will lead to digital watermarking’ wider adoption — similar to the work it did around developing the HTTPS ecosystem and in getting both developers and the public on-board with it. This will help federal officials achieve their other goal of ensuring that the government's official messaging can be relied upon.

Civil Rights and Consumer Protections

The first Blueprint for an AI Bill of Rights that the White House released last October directed agencies to “combat algorithmic discrimination while enforcing existing authorities to protect people's rights and safety,” the administration official said. “But there's more to do.” 

The new EO will require guidance be extended to “landlords, federal benefits programs and federal contractors” to prevent AI systems from exacerbating discrimination within their spheres of influence. It will also direct the Department of Justice to develop best practices for investigating and prosecuting civil rights violations related to AI, as well as, per the announcement, “the use of AI in sentencing, parole and probation, pretrial release and detention, risk assessments, surveillance, crime forecasting and predictive policing, and forensic analysis."

Additionally, the EO calls for prioritizing federal support to accelerate development of privacy-preserving techniques that would enable future LLMs to be trained on large datasets without the current risk of leaking personal details that those datasets might contain. These solutions could include “cryptographic tools that preserve individuals’ privacy,” per the White House release, developed with assistance from the Research Coordination Network and National Science Foundation. The executive order also reiterates its calls for bipartisan legislation from Congress addressing the broader privacy issues that AI systems present for consumers.

In terms of healthcare, the EO states that the Department of Health and Human Services will establish a safety program that tracks and remedies unsafe, AI-based medical practices. Educators will also see support from the federal government in using AI-based educational tools like personalized chatbot tutoring.

Worker Protections

The Biden administration concedes that while the AI revolution is a decided boon for business, its capabilities make it a threat to worker security through job displacement and intrusive workplace surveillance. The EO seeks to address these issues with “the development of principles and employer best practices that mitigate the harms and maximize the benefit of AI for workers,” an administration official said. “We encourage federal agencies to adopt these guidelines in the administration of their programs.”

Richard Shotwell/Invision/AP

The EO will also direct the Department of Labor and the Council of Economic Advisors to both study how AI might impact the labor market and how the federal government might better support workers “facing labor disruption” moving forward. Administration officials also pointed to the potential benefits that AI might bring to the federal bureaucracy including cutting costs, and increasing cybersecurity efficacy. “There's a lot of opportunity here, but we have to to ensure the responsible government development and deployment of AI,” an administration official said.

To that end, the administration is launching on Monday a new federal jobs portal, AI.gov, which will offer information and guidance on available fellowship programs for folks looking for work with the federal government. “We're trying to get more AI talent across the board,” an administration official said. “Programs like the US Digital Service, the Presidential Innovation Fellowship and USA jobs — doing as much as we can to get talent in the door.” The White House is also looking to expand existing immigration rules to streamline visa criteria, interviews and reviews for folks trying to move to and work in the US in these advanced industries.

The White House reportedly did not preview the industry on this particular swath of radical policy changes, though administration officials did note that they had already been collaborating extensively with AI companies on many of these issues. The Senate held its second AI Insight Forum event last week on Capitol Hill, while Vice President Kamala Harris is scheduled to speak at the UK Summit on AI Safety, hosted by Prime Minister Rishi Sunak on Tuesday.

Chip Somodevilla via Getty Images

At a Washington Post event on Thursday, Senate Majority Leader Charles Schumer (D-NY) was already arguing that the executive order did not go far enough and could not be considered an effective replacement for congressional action, which to date, has been slow in coming. 

“There’s probably a limit to what you can do by executive order,” Schumer told WaPo, “They [the Biden Administration] are concerned, and they’re doing a lot regulatorily, but everyone admits the only real answer is legislative.”

The explosive growth in artificial intelligence in recent years — crowned with the meteoric rise of generative AI chatbots like ChatGPT — has seen the technology take on many tasks that, formerly, only human minds could handle. But despite their increasingly capable linguistic computations, these machine learning systems remain surprisingly inept at making the sorts of cognitive leaps and logical deductions that even the average teenager can consistently get right. 

In this week's Hitting the Books excerpt, A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains, AI entrepreneur Max Bennett explores the quizzical gap in computer competency by exploring the development of the organic machine AIs are modeled after: the human brain. 

Focusing on the five evolutionary "breakthroughs," amidst myriad genetic dead ends and unsuccessful offshoots, that led our species to our modern minds, Bennett also shows that the same advancements that took humanity eons to evolve can be adapted to help guide development of the AI technologies of tomorrow. In the excerpt below, we take a look at how generative AI systems like GPT-3 are built to mimic the predictive functions of the neocortex, but still can't quite get a grasp on the vagaries of human speech.

HarperCollins

Excerpted from A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains by Max Bennett. Published by Mariner Books. Copyright © 2023 by Max Bennett. All rights reserved.

Words Without Inner Worlds

GPT-3 is given word after word, sentence after sentence, paragraph after paragraph. During this long training process, it tries to predict the next word in any of these long streams of words. And with each prediction, the weights of its gargantuan neural network are nudged ever so slightly toward the right answer. Do this an astronomical number of times, and eventually GPT-3 can automatically predict the next word based on a prior sentence or paragraph. In principle, this captures at least some fundamental aspect of how language works in the human brain. Consider how automatic it is for you to predict the next symbol in the following phrases:

• One plus one equals _____

• Roses are red, violets are _____

You’ve seen similar sentences endless times, so your neocortical machinery automatically predicts what word comes next. What makes GPT-3 impressive, however, is not that it just predicts the next word of a sequence it has seen a million times — that could be accomplished with nothing more than memorizing sentences. What is impressive is that GPT-3 can be given a novel sequence that it has never seen before and still accurately predict the next word. This, too, clearly captures something that the human brain can _____.

Could you predict that the next word was do? I’m guessing you could, even though you had never seen that exact sentence before. The point is that both GPT-3 and the neocortical areas for language seem to be engaging in prediction. Both can generalize past experiences, apply them to new sentences, and guess what comes next.

GPT-3 and similar language models demonstrate how a web of neurons can reasonably capture the rules of grammar, syntax, and context if it is given sufficient time to learn. But while this shows that prediction is part of the mechanisms of language, does this mean that prediction is all there is to human language? Try to finish these four questions:

• If 3x + 1 = 3, then x equals _____

• I am in my windowless basement, and I look toward the sky, and I see _____

• He threw the baseball 100 feet above my head, I reached my hand up to catch it, jumped, and _____

• I am driving as fast as I can to LA from New York. One hour after passing through Chicago, I finally _____

Here something different happens. In the first question, you likely paused and performed some mental arithmetic before being able to answer the question. In the other questions, you probably, even for only a split second, paused to visualize yourself in a basement looking upward, and realized what you would see is the ceiling. Or you visualized yourself trying to catch a baseball a hundred feet above your head. Or you imagined yourself one hour past Chicago and tried to find where you would be on a mental map of America. With these types of questions, more is happening in your brain than merely the automatic prediction of words.

We have, of course, already explored this phenomenon—it is simulating. In these questions, you are rendering an inner simulation, either of shifting values in a series of algebraic operations or of a three-dimensional basement. And the answers to the questions are to be found only in the rules and structure of your inner simulated world.

I gave the same four questions to GPT-3; here are its responses (responses of GPT-3 are bolded and underlined):

• If 3x + 1 = 3 , then x equals 1

• I am in my windowless basement, and I look toward the sky, and I see a light, and I know that it is a star, and I am happy.

• He threw the baseball 100 feet above my head, I reached my hand up to catch it, jumped, and caught it. It was a lot of fun!

• I am driving as fast as I can to LA from New York. One hour after passing through Chicago, I finally get to see the Pacific Ocean.

All four of these responses demonstrate that GPT-3, as of June 2022, lacked an understanding of even simple aspects of how the world works. If 3x + 1 = 3, then x equals 2/3, not 1. If you were in a basement and looked toward the sky, you would see your ceiling, not stars. If you tried to catch a ball 100 feet above your head, you would not catch the ball. If you were driving to LA from New York and you’d passed through Chicago one hour ago, you would not yet be at the coast. GPT-3’s answers lacked common sense.

What I found was not surprising or novel; it is well known that modern AI systems, including these new supercharged language models, struggle with such questions. But that’s the point: Even a model trained on the entire corpus of the internet, running up millions of dollars in server costs — requiring acres of computers on some unknown server farm — still struggles to answer common sense questions, those presumably answerable by even a middle-school human.

Of course, reasoning about things by simulating also comes with problems. Suppose I asked you the following question:

Tom W. is meek and keeps to himself. He likes soft music and wears glasses. Which profession is Tom W. more likely to be?

1) Librarian

2) Construction worker

If you are like most people, you answered librarian. But this is wrong. Humans tend to ignore base rates—did you consider the base number of construction workers compared to librarians? There are probably one hundred times more construction workers than librarians. And because of this, even if 95 percent of librarians are meek and only 5 percent of construction workers are meek, there still will be far more meek construction workers than meek librarians. Thus, if Tom is meek, he is still more likely to be a construction worker than a librarian.

The idea that the neocortex works by rendering an inner simulation and that this is how humans tend to reason about things explains why humans consistently get questions like this wrong. We imagine a meek person and compare that to an imagined librarian and an imagined construction worker. Who does the meek person seem more like? The librarian. Behavioral economists call this the representative heuristic. This is the origin of many forms of unconscious bias. If you heard a story of someone robbing your friend, you can’t help but render an imagined scene of the robbery, and you can’t help but fill in the robbers. What do the robbers look like to you? What are they wearing? What race are they? How old are they? This is a downside of reasoning by simulating — we fill in characters and scenes, often missing the true causal and statistical relationships between things.

It is with questions that require simulation where language in the human brain diverges from language in GPT-3. Math is a great example of this. The foundation of math begins with declarative labeling. You hold up two fingers or two stones or two sticks, engage in shared attention with a student, and label it two. You do the same thing with three of each and label it three. Just as with verbs (e.g., running and sleeping), in math we label operations (e.g., add and subtract). We can thereby construct sentences representing mathematical operations: three add one.

Humans don’t learn math the way GPT-3 learns math. Indeed, humans don’t learn language the way GPT-3 learns language. Children do not simply listen to endless sequences of words until they can predict what comes next. They are shown an object, engage in a hardwired nonverbal mechanism of shared attention, and then the object is given a name. The foundation of language learning is not sequence learning but the tethering of symbols to components of a child’s already present inner simulation.

A human brain, but not GPT-3, can check the answers to mathematical operations using mental simulation. If you add one to three using your fingers, you notice that you always get the thing that was previously labeled four.

You don’t even need to check such things on your actual fingers; you can imagine these operations. This ability to find the answers to things by simulating relies on the fact that our inner simulation is an accurate rendering of reality. When I mentally imagine adding one finger to three fingers, then count the fingers in my head, I count four. There is no reason why that must be the case in my imaginary world. But it is. Similarly, when I ask you what you see when you look toward the ceiling in your basement, you answer correctly because the three-dimensional house you constructed in your head obeys the laws of physics (you can’t see through the ceiling), and hence it is obvious to you that the ceiling of the basement is necessarily between you and the sky. The neocortex evolved long before words, already wired to render a simulated world that captures an incredibly vast and accurate set of physical rules and attributes of the actual world.

To be fair, GPT-3 can, in fact, answer many math questions correctly. GPT-3 will be able to answer 1 + 1 =___ because it has seen that sequence a billion times. When you answer the same question without thinking, you are answering it the way GPT-3 would. But when you think about why 1 + 1 =, when you prove it to yourself again by mentally imagining the operation of adding one thing to another thing and getting back two things, then you know that 1 + 1 = 2 in a way that GPT-3 does not.

The human brain contains both a language prediction system and an inner simulation. The best evidence for the idea that we have both these systems are experiments pitting one system against the other. Consider the cognitive reflection test, designed to evaluate someone’s ability to inhibit her reflexive response (e.g., habitual word predictions) and instead actively think about the answer (e.g., invoke an inner simulation to reason about it):

Question 1: A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?

If you are like most people, your instinct, without thinking about it, is to answer ten cents. But if you thought about this question, you would realize this is wrong; the answer is five cents. Similarly:

Question 2: If it takes 5 machines 5 minutes to make 5 widgets, how long would it take 100 machines to make 100 widgets?

Here again, if you are like most people, your instinct is to say “One hundred minutes,” but if you think about it, you would realize the answer is still five minutes.

And indeed, as of December 2022, GPT-3 got both of these questions wrong in exactly the same way people do, GPT-3 answered ten cents to the first question, and one hundred minutes to the second question.

The point is that human brains have an automatic system for predicting words (one probably similar, at least in principle, to models like GPT-3) and an inner simulation. Much of what makes human language powerful is not the syntax of it, but its ability to give us the necessary information to render a simulation about it and, crucially, to use these sequences of words to render the same inner simulation as other humans around us.

It's getting to the point these days that we can't even trust our own eyes with the amounts of digital trickery, trolling, misinformation and disinformation dominating social media. Heck, even reputable tech companies are selling us solutions to reimagine historical events. Not Leica, though! The venerated camera company officially announced the hotly-anticipated M11-P on Thursday, its first camera to incorporate the Content Credential secure metadata system.

Content Credentials are the result of efforts by the Content Authenticity Initiative (CAI), "a group of creators, technologists, journalists, and activists leading the global effort to address digital misinformation and content authenticity," and the Coalition for Content Provenance and Authenticity (C2PA), "a formal coalition dedicated exclusively to drafting technical standards and specifications as a foundation for universal content provenance." These intertwined industry advocacy groups created Content Credentials system in response to growing abuse and misuse of generative AI systems in creating and spreading misinformation online. 

"The Leica M11-P launch will advance the CAI’s goal of empowering photographers everywhere to attach Content Credentials to their photographs at the time of capture," Santiago Lyon, Head of Advocacy and Education at CAI, said in a press statement, "creating a chain of authenticity from camera to cloud and enabling photographers to maintain a degree of control over their art, story and context."

"This is the realization of a vision the CAI and our members first set out four years ago, transforming principles of trust and provenance into consumer-ready technology," he continued.

Leica

Content Credentials works by capturing specific metadata about the photograph — the camera used to take it, as well as the location, time and other details about the shot — and locks those in a secure "manifest" that is bundled up with the image itself using a cryptographic key (the process is opt-in for the photog). Those credentials can easily be verified online or on the Leica FOTOS app. Whenever someone subsequently edits that photo, the changes are recorded to an updated manifest, rebundled with the image and updated in the Content Credentials database whenever it is reshared on social media. Users who find these images online can click on the CR icon in the pictures corner to pull up all of this historical manifest information as well, providing a clear chain of providence, presumably, all the way back to the original photographer. The CAI describes Content Credentials as a "nutrition label" for photographs. 

The M11-P itself is exactly what you'd expect from a company that's been at the top of the camera market since the middle of the last century. It offers a 60 MP BSI CMOS sensor on a Maestro-III processor with 256 GB of internal storage. The M11-P is now on sale but it's also $9,480 at retail so, freelancers, sorry.

As with all things Google of late, AI capabilities are coming to Maps. The company announced a slew of machine learning updates for the popular app Thursday including an "Immersive View" for route planning, deeper Lens integration for local navigation and more accurate real-time information. 

Back in May at its I/O developer conference, Google executives debuted Immersive View for routes, which provides navigation shots of your planned route. Whether you're on foot, bike, taking public transportation or driving, this will allow you to scrub back and forth through street level, turn-by-turn visuals of the path you're taking. The feature arrives on iOS and Android this week for Amsterdam, Barcelona, Dublin, Florence, Las Vegas, London, Los Angeles, Miami, New York, Paris, San Francisco, San Jose, Seattle, Tokyo and Venice.

Just because you can see the route to get where you're going doesn't guarantee you'll be able to read the signage along the way. Google is revamping its existing AI-based Search with Live View feature in Maps. Simply tap the Lens icon in Maps and wave your phone around, the system will determine your precise street level location and be able to direct you to nearby resources like ATMs, transit stations, restaurants, coffee shops and stores. 

The map itself is set to receive a significant upgrade. Buildings along your route will be more accurately depicted within the app to help you better orient yourself in unfamiliar cities, lane details along tricky highway interchanges will be more clearly defined in-app as well. Those updates will arrive for users in a dozen countries including the US, Canada, France and Germany over the next few months. US users will also start to see better in-app HOV lane designations and European customers should expect a significant expansion of Google's AI speed limit sign reader technology out to 20 nations in total. 

Google

Google Maps also runs natively in a growing number of electric vehicles, as part of the Android Automotive OS ecosystem. That Maps is getting an update too as part of the new Places API. Starting this week, drivers will see increased information about nearby charging stations including whether the plugs work with their EV, the power throughput of the charger, and whether the plug has been used recently — an indirect means of inferring whether or not the station is out of service, which Google helpfully points out, is the case around 25 percent of them. 

Even search is improving with the new update. Users will be soon able to look for nearby destinations that meet more esoteric criteria, such as “animal latte art” or “pumpkin patch with my dog,” results of which are gleaned from the analysis of "billions of photos shared by the Google Maps community," per a Google blog post Thursday.