reduce geopolitical instability amidst fierce AI competition among states

ensure the benefits ofcutting-edge AI reach groups that are otherwise underserved by AI developmen

legitimate international access to advanced AI

a Frontier AICollaborative could take the form of an international private-public partnership that leverages existingtechnology and capacity in industry, for example by contracting access to or funding innovation inappropriate AI technology from frontier AI developers.

ligned countries may seek to form governance clubs, asthey have in other domains. This facilitates decision-making, but may make it harder to enlist othercountries later in the process

its effectiveness will depend on its membership, governance and standard-setting processes

Governance Organization should focus primarily on advanced AI systems thatpose significant global risks, but it will be difficult in practice to decide on the nature and sophisti-cation of AI tools that should be broadly available and uncontrolled versus the set of systems thatshould be subject to national or international governance

Automated (even AI-enabled) monitoring

some states may be especially reluctant to join due to fear of clandestinenoncompliance by other states

Other AI-specific incentives for participation include conditioning on participationaccess to AI technology (possibly from a Frontier AI Collaborative) or computing resources.22Statesmight also adopt import restrictions on AI from countries that are not certified by a GovernanceOrganization

while urgent risks may need to be addressed at first by smallergroups of frontier AI states, or aligned states with relevant expertise

The impact of a Governance Organization depends on states adoptingits standards and/or agreeing to monitoring.

standard setting (especially in an international and multistakeholder context) tends to be a slowprocess

detection and inspections of large data centers

self-reporting of compliance with international standards

Where states have incentives to undercut each other'sregulatory commitments, international institutions may be needed to support and incentivize bestpractices. That may require monitoring standards compliance.

therefore enable greater access to advanced AI

international standard setting wouldreduce cross-border frictions due to differing domestic regulatory regimes

uture regulations will limit access to powerful AI technologies in jurisdictions with inadequate AIgovernanc

he InternationalTelecommunication Union (ITU)

Advanced AI Governance Organization.

intergovernmental or multi-stakeholder

as less institutionalizedand politically authoritative scientific advisory panels on advanced AI

epresentation may trade off against a Commission’s ability to overcomescientific challenges and generate meaningful consensus

broad geographic representation in the main decisionmaking bodies, and a predominance of scientificexperts in working groups

he scientific understanding of the impacts of AI should ideally beseen as a universal good and not be politicized

foundational “Conceptual Framework

Commission might undertake activities that draw andfacilitate greater scientific attention, such as organizing conferences and workshops and publishingresearch agenda

o increase chances of success, aCommission should foreground scientific rigor and the selection of highly competent AI expertswho work at the cutting edge of technological development and who can continually interpret theever-changing technological and risk landscape

internationally representative group of experts

there is significant disagreement even among experts about the different opportunities andchallenges created by advanced AI

International consensus on the opportunities and risks from advanced AI

the Commission on Frontier AI could facilitatescientific consensus by convening experts to conduct rigorous and comprehensive assessments ofkey AI topics, such as interventions to unlock AI’s potential for sustainable development, the effectsof AI regulation on innovation, the distribution of benefits, and possible dual-use capabilities fromadvanced systems and how they ought to be managed

intergovernmental body

dangerous inputs: computing resources have been targeted by US, Japanese and Dutch export controlsthat prevent the sale of certain AI chips and semiconductor manufacturing equipment to Chin

norms and standards

standards

Establish guidelines

through education, infrastructure, andsupport of the local commercial ecosystem

develop frontier AI collectively or distribute and enable access

developing and/or enabling safe forms of access to AI.

ontrolling AIinputs

tobuild consensus on risksand how they can be mitigated,andset safety norms and standardsandsupport their implementationto help developers and regulators with responsible developmentand use. International efforts toconduct or support AI safety research

national regulation may be ineffective for managing the risks of AI even within states. States willinevitably be impacted by the development of such capabilities in other jurisdictions

building capacity to benefit from AI through education, infrastructure, andlocal commercial ecosystems

build consensus on AI opportunities

Standards

SO/IEC

whether these institutions should be new or evolutions of existing organizations, whether the con-ditions under which these institutions are likely to be most impactful will obtain,

A failure to coordinate or harmonizeregulation may also slow innovation.

It groups these functions into fourinstitutional models that exhibit internal synergies and have precedents in existingorganizations:

the Media in the Digital Age.

But if the superforecasters are so good at predictions, and experts have so much topic-specific knowledge, you might at least expect the two groups to influence each other’s beliefs.

But superforecasters and AI experts seemed to hold very different views of how societies might respond to small-scale damage caused by AI . Superforecasters tended to think that would prompt heavy scrutiny and regulation to head off bigger problems later. Domain experts, by contrast, tended to think that commercial and geopolitical incentives might outweigh worries about safety, even after real harm had been caused.

Such people share a few characteristics, such as careful, numerical thinking and an awareness of the cognitive biases that might lead them astray.

The emergence of modern, powerful machine-learning models dates to the early years of the 2010s. And the field is still developing quickly. That leaves much less data on which to base predictions.

If humans used AI to help design more potent bioweapons, for instance, it would have contributed fundamentally, albeit indirectly, to the disaster.

One reason for AI ’s strong showing, says Dan Maryland, a superforecaster who participated in the study, is that it acts as a “force multiplier” on other risks like nuclear weapons

The median superforecaster reckoned there was a 2.1% chance of an AI -caused catastrophe, and a 0.38% chance of an AI -caused extinction, by the end of the century. AI experts, by contrast, assigned the two events a 12% and 3% chance, respectively.

A “catastrophe” was defined as something that killed a mere 10% of the humans in the world, or around 800m people. (The second world war, by way of comparison, is estimated to have killed about 3% of the world’s population of 2bn at the time.) An “extinction”, on the other hand, was defined as an event that wiped out everyone with the possible exception of, at most, 5,000 lucky (or unlucky) souls.

On the one hand were subject-matter, or “domain”, experts in nuclear war, bio-weapons, AI and even extinction itself. On the other were a group of “superforecasters”—general-purpose prognosticators with a proven record of making accurate predictions on all sorts of topics, from election results to the outbreak of wars.

These days, worries about “existential risks”—those that pose a threat to humanity as a species, rather than to individuals—are not confined to military scientists. Nuclear war; nuclear winter; plagues (whether natural, like covid-19, or engineered); asteroid strikes and more could all wipe out most or all of the human race. The newest doomsday threat is artificial intelligence ( ai ). In May a group of luminaries in the field signed a one-sentence open letter stating: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

India insists data must be stored locally: to give its law-enforcement agencies easy access, to protect against foreign snooping and as a way to boost investment in the tech sector.

most governments lack policymakers with relevant technical expertise and most digital issues cut across different domains, extending beyond the traditional remit of trade negotiators.

In 2019 Abe Shinzo, the late Japanese prime minister, proposed the concept of Data Free Flow with Trust. That rather nebulous idea is materialising as a set of global norms to counter digital protectionism. As Matthew Goodman of CSIS, a think-tank in Washington, puts it: “It’s about the un-China approach to data governance.”

“There’s a vacuum in terms of rules, norms and agreements that govern digital trade,” laments Nigel Cory of the Information Technology and Innovation Foundation, a research institute in Washington .

The Philippines and Guam have emerged as attractive substitutes.

Hong Kong was traditionally one of three major data hubs in Asia, with Japan and Singapore.

Intra-Asia data flows make up over 50% of the region’s bandwidth, up from 47% in 2018, while the share going to America and Canada has dipped from 40% to 34% over the same period.

he most congested cable route in Asia is also its most contested: the South China Sea is the “main street” of submarine cables, especially between Japan, Singapore and Hong Kong, notes Murai Jun, a Japanese internet pioneer.

“Customers are asking more about the security of cables and routes,” says Uchiyama Kazuaki of NTT World Engineering Marine Corporation, the firm that owns the Kizuna.

Aside from a heavy state hand in China’s cable industry, such infrastructure tends to be privately financed and owned. A small handful of companies dominate the production and installation of cables; big tech firms are their main users.

While in the past constructing internet infrastructure tended to be a “collaborative effort” between countries and between firms, in recent years its enabling environment has soured amid growing friction between China and America.

Asia saw international bandwidth usage grow by 39% in 2022, compared to the global average of 36%, according to TeleGeography, a research firm.

migrant and refuge

Our continued investment in innovation and our specific regulatory environment is what helped us lead the world in critical tech industries like quantum computing.

The average American is already starting to see the benefits of AI technology in accessibility, efficiency, and reduction of human error.

The resulting generative AI models need not be trained from scratch but can build upon open-source generative AI that has used lawfully sourced content.

Vendor and customer contracts can include AI-related language added to confidentiality provisions in order to bar receiving parties from inputting confidential information of the information-disclosing parties into text prompts of AI tools.

they should demand terms of service from generative AI platforms that confirm proper licensure of the training data that feed their AI.

Developing these audit trails would assure companies are prepared if (or, more likely, when) customers start including demands for them in contracts as a form of insurance that the vendor’s works aren’t willfully, or unintentionally, derivative without authorization

would increase transparency about the works included in the training data.

has announced that artists will be able to opt out of the next generation of the image generator.

Stable Diffusion, Midjourney and others have created their models based on the LAION-5B dataset, which contains almost six billion tagged images compiled from scraping the web indiscriminately, and is known to include substantial number of copyrighted creations.

Customers of AI tools should ask providers whether their models were trained with any protected content

There’s also the risk of accidentally sharing confidential trade secrets or business information by inputting data into generative AI tools.

to become unequivocally “transformative,”

Google successfully defended itself against a lawsuit by arguing that transformative use allowed for the scraping of text from books to create its search engine, and for the time being, this decision remains precedential.

without the owner’s permission “for purposes such as criticism (including satire), comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research,”

the interpretation of the fair use doctrine,

the bounds of what is a “derivative work” under intellectual property laws

Getty, an image licensing service, filed a lawsuit against the creators of Stable Diffusion alleging the improper use of its photos, both violating copyright and trademark rights it has in its watermarked photograph collection.

If a court finds that the AI’s works are unauthorized and derivative, substantial infringement penalties can apply.

Andersen v. Stability AI et al., three artists formed a class to sue multiple generative AI platforms on the basis of the AI using their original works without license to train their AI in their styles

how the laws on the books should be applied

does copyright, patent, trademark infringement apply to AI creations?

This process comes with legal risks, including intellectual property infringement

to copyright infringement, ownership of AI-generated works, and unlicensed content in training data.

With humans and AI working to their respective strengths, they can transform unknown unknowns into known unknowns, opening the door to breakthrough thinking: logical and conceptual leaps that neither could make without the other.

our brains work in a reductive manner; we generate ideas and then explain them to other people.

By focusing on areas where the human brain and machines complement one another.

the technology is fundamentally backward-looking, trained on yesterday’s data — and the future might not look anything like the past

How might you use those opportunities to throw people off balance so they’ll generate questions that reach beyond what they intellectually know to be right, what makes them emotionally comfortable, and what they are accustomed to saying and doing?

Increased question velocity, variety, and especially novelty give facilitate recognizing where you’re intellectually wrong, and becoming emotionally uncomfortable and behaviorally quiet — the very conditions that, we’ve found, tend to produce game-changing lines of inquiry.

AI can take really obscure variables and make novel connections.

sift through much more data, and connect more dots,

“category jumping” questions — the gold standard of innovative inquiry

uncover patterns and correlations in large volumes of data — connections that humans can easily miss without the technology

more questions don’t necessarily amount to better questions, which means you’ll still need to exercise human judgment in deciding how to proceed.

we found that 79% of respondents asked more questions, 18% asked the same amount, and 3% asked fewer.

humans can start exploring the power of more context-dependent and nuanced questions

to reveal deeply buried patterns in the data

we’ve defined “artificial intelligence” broadly to include machine learning, deep learning, robotics, and the recent explosion of generative AI.)

design-thinking sessions

to help people become more inquisitive, creative problem-solvers on the job.

can help people ask smarter questions, which in turn makes them better problem solvers and breakthrough innovators.

from identification to ideation.

Paired with “soft” inquiry-related skills such as critical thinking, innovation, active learning, complex problem solving, creativity, originality, and initiative, this technology can further our understanding of an increasingly complex world

it can help people ask better questions and be more innovative.

still view AI rather narrowly, as a tool that alleviates the costs and inefficiencies of repetitive human labor and increasing organizations’ capacity to produce, process, and analyze piles and piles of data

AI increases question velocity, question variety, and question novelty.

“Cables are an enormous lever of power,” Wicker said. “If you can’t control these networks directly, you want a company you can trust to control them.”

n 1997, AT&T sold its cable-laying operation, including a fleet of ships, to Tyco International, a security company based in New Jersey. In 2018, Tyco sold the cable unit, by this time dubbed TE SubCom, for $325 million to Cerberus, the New York private equity firm.

That project, known as the Oman Australia Cable, was spearheaded by SUBCO, a Brisbane-based subsea cable investment company owned by Australian entrepreneur Bevan Slattery.

“Silicon Valley is waking up to the reality that it has to pick a side,”

Microsoft – whose President Brad Smith said in 2017 that the tech sector needed to be a “neutral digital Switzerland” – announced in May that it had discovered Chinese state-sponsored hackers targeting U.S. critical infrastructure, a rare example of a big tech firm calling out Beijing for espionage.

America’s SubCom, Japan’s NEC Corporation, France’s Alcatel Submarine Networks and China’s HMN Tech.

First, Washington needs SubCom to expand the Navy’s undersea cable network so that it can better coordinate military operations and enhance surveillance on China’s expanding fleet of submarines and warships, the people said. Second, the Biden administration wants SubCom to build more commercial subsea internet cables controlled by U.S. companies, a strategy aimed at ensuring that America remains the primary custodian of the internet, according to the two industry officials.

Subsea cables are vulnerable to sabotage and espionage, and Beijing and Washington have accused each other of tapping cables to spy on data or carry out cyberattacks.

SubCom’s journey from Cold War experiment to global cable constructor and now a shadowy player in the U.S.-China tech war is detailed in this story for the first time.

This dual role has made SubCom increasingly valuable to Washington as global internet infrastructure – from undersea cables to data centers and 5G mobile networks – risks fracturing into two systems, one backed by the United States, the other controlled by China.

SubCom is the exclusive undersea cable contractor to the U.S. military, laying a web of internet and surveillance cables across the ocean floor

The CS Dependable is owned by SubCom, a small-town New Jersey cable manufacturer that’s playing an outsized role in a race between the United States and China to control advanced military and digital technologies that could decide which country emerges as the world’s preeminent superpower.

The long-term vision of enabling any employee — and customers as well — to easily access important knowledge within and outside of a company to enhance productivity and innovation is a powerful draw. Generative AI appears to be the technology that is finally making it possible.

Any company that commits to embedding its own knowledge into a generative AI system should be prepared to revise its approach to the issue frequently over the next several years.

either though training or policies — include:

Generative AI capabilities, including awareness of context and history, generating new content by aggregating or combining knowledge from various sources, and data-driven predictions, can provide powerful support for knowledge work.

To realize opportunities and manage potential risks of generative AI applications to knowledge management, companies need to develop a culture of transparency and accountability that would make generative AI-based knowledge management systems successful.

User prompts into publicly-available LLMs are used to train future versions of the system, so some companies (Samsung, for example) have feared propagation of confidential and private information and banned LLM use by employees. However, most companies’ efforts to tune LLMs with domain-specific content are performed on private instances of the models that are not accessible to public users, so this should not be a problem. In addition, some generative AI systems such as ChatGPT allow users to turn off the collection of chat histories, which can address confidentiality issues even on public systems.

the company’s attorneys helped create a series of “pre-prompts” that tell the generative AI system what types of questions it should answer and those it should politely avoid.

many legal observers feel that “fair use” provisions of copyright law will apply to them, although this hasn’t been tested in the courts (and not all countries have such provisions in their copyright laws).

For example, for BloombergGPT, which is intended for answering financial and investing questions, the system was evaluated on public dataset financial tasks, named entity recognition, sentiment analysis ability, and a set of reasoning and general natural language processing tasks.

The good news is that companies who have tuned their LLMs on domain-specific information have found that hallucinations are less of a problem than out-of-the-box LLMs, at least if there are no extended dialogues or non-business prompts.

if content authors are aware of how to create effective documents.

Most companies that do not have well-curated content will find it challenging to do so for just this purpose.

Morgan Stanley, for example, has a group of 20 or so knowledge managers in the Philippines who are constantly scoring documents along multiple criteria; these determine the suitability for incorporation into the GPT-4 system.

The alternative is to create vector embeddings — arrays of numeric values produced from the text by another pre-trained machine learning model (Morgan Stanley uses one from OpenAI called Ada).

The prompt-trained system is operated in a private cloud that is only accessible to Morgan Stanley employees.

Morgan Stanley, for example, used prompt tuning to train OpenAI’s GPT-4 model using a carefully curated set of 100,000 documents with important investing, general business, and investment process knowledge.

It requires considerable data science expertise; the scientific paper for the Google project, for example, had 31 co-authors.

with Google’s general PaLM2 LLM

to add specific domain content to a system that is already trained on general knowledge and language-based interaction.

Bloomberg’s data scientists employed 700 tokens, or about 350 billion words, 50 billion parameters, and 1.3 million hours of graphics processing unit time.

some of the same factors that made knowledge management difficult in the past are still present.

the heyday of the “knowledge management” movement in the 1990s and early 2000s, but most companies found the technology of the time inadequate for the task

a generative AI-based system for customer support led to increased productivity of customer support agents and improved retention, while leading to higher positive feedback on the part of customers.

can’t respond to prompts or questions regarding proprietary content or knowledge.

to express complex ideas in articulate language

knowledge within organizations is typically generated and captured across various sources and forms, including individual minds, processes, policies, reports, operational transactions, discussion boards, and online chats and meetings.

through effective and agile creation, management, application, recombination, and deployment of knowledge assets and know-how.

We have to be very, very careful about ensuring that it doesn’t come across as AI surveillance,

introducing AI ethicists

While there are laws being whipped up around how employers should ethically and responsibly implement AI around, for example, hiring, so as to ensure job posts aren’t discriminatory in any way (like NYC’s Law 144), there are still relatively few guardrails.

more traditional kind, which detects patterns from data and provides predictive analysis, which has been used for the last few years by some businesses, but isn’t yet mainstream.

they need to invest in reskilling their workforces

A bit like the customer relationship management software that airlines use, to create a more personalized travel experience.

“I’m a big believer in human interaction and connection. I don’t think that will go away. Automation and speed will help us do different things, focusing on more strategic, creative endeavors. But it will still be on us to understand our people on a human level.”

The theory goes, that if people are saving time on more administrative, tedious aspects of their work, they’ll have more brain capacity for creative and strategic thinking.

improve the meaning and purpose of work.

“Knowing that AI will change many roles in the workforce, CEOs should encourage people to embrace experimentation with the technology and communicate the upskilling opportunities for them to work in tandem with the AI, rather than be replaced by it,”

“AI is not a technology that can be given to a CTO or a similar position, it is a tool that must be understood by the CEO and the entire management team.”

four “trust-building” processes: improve the governance of AI systems and processes; confirm AI-driven decisions are interpretable and easily explainable; monitor and report on AI model performance and protect AI systems from cyber threats and manipulations, per a 2023 Trust survey from management consultancy PwC.

to road test AI.

but leaders still need to be able to frame the issues, make the smart calls and ensure they’re well executed.”

the human element of interpreting data and asking AI the right questions to make sound judgments, will be crucial.

Most (94%) business leaders agree that AI is critical for success, per a 2022 Deloitte report.

we must not dismiss legitimate concerns regarding the alignment problem

We should prioritize targeted regulation for specific applications — recognizing that each domain comes with its own set of ethical dilemmas and policy challenges.

shift in the allocation of National Science Foundation (NSF) funding toward responsible AI research

addressing data privacy reform is essential.

we implement retention and reproducibility requirements for AI research

the government must limit abuse across applications using existing laws, such as those governing data privacy and discrimination

Proposals for centralized government-backed projects underestimate the sheer diversity of opinions among AI researchers.

while the Manhattan Project’s ultimate goal was relatively singular — design and build the atomic bomb — AI safety encompasses numerous ambiguities ranging from the meaning of concepts like “mechanistic interpretability” to “value alignment.

to address the “alignment problem,” the fear that powerful AI models might not act in a way we ask of them; or to address mechanistic interpretability, the ability to understand the function of each neuron in a neural network.

gets mired in unprecedented alarmism rather than focusing on addressing these more proximate — and much more likely — challenges.

rogue actors using large AI models to dismantle cybersecurity around critical infrastructure; political parties using disinformation at scale to destabilize fragile democratic governments; domestic terrorists using these models to learn how to build homemade weapons; and dictatorial regimes using them to surveil their populations or build dystopian social credit systems, among others.

distracts from the grounded conversations necessary to develop well-informed policies around AI governance.

The technology is neither inherently good nor evil; in contrast, philosophers, ethicists, and even the pope have argued that the same could not necessarily be said about nuclear weapons, because their mere possession is an inherent threat to kill millions of people.

These types of common exaggerations ultimately detract from effective policymaking aimed at addressing both immediate risks and potential catastrophic threats posed by certain AI technologies.

Bundling occurs when a company offers multiple products together as a single package.

Experience has also shown how firms can use “open first, closed later” tactics in ways that undermine long-term competition.

The open-source innovation explosion in image generation, coupled with new developments in optimizations, made it possible for nearly anyone to develop, iterate on, and deploy the models using smaller datasets and lower-cost consumer hardware

Additionally, some markets for specialized chips are—or could be, without appropriate competition policies and antitrust enforcement—highly concentrated.

the fine-tuning technique LoRa (Low-Rank Adaptation) can enable a developer to fine-tune a model to perform a specific task using consumer grade GPUs

Access to computational resources is a third key input in generative AI markets.

Companies that can acquire both the engineering experience and professional talent necessary to build and package the final generative AI product or service will be better positioned to gain market share.

Developing a generative model requires a significant engineering and research workforce with particular—and relatively rare—skillsets, as well as a deep understanding of machine learning, natural language processing, and computer vision.

This may be particularly true in specialized domains or domains where data is more highly regulated, such as healthcare or finance

First, more established companies may benefit from access to data collected from their users over many years—especially if the incumbents also own digital platforms that amass large amounts of data.