Here’s the broadest question I have for our global and ongoing debate on Generative AI, triggered by the rollout over the last six months of ChatGPT and its many cousins:
How do we move this discussion from debate over walls, to deliberation on filters?
Walls? Filters? It’s an abstract question. I get it. But work with me on this. Later today I’m joining that annual global incubator of “Ideas Worth Spreading”, the main TED conference in Vancouver that always draws some of the world’s finest minds to explore shifts in technology, art, and culture.
This year’s theme is “POSSIBILITY”. Let that word sink in for a second or two. Possibility.
No word captures the quintessence of the social, economic, and cultural transformation on which we are embarked as effectively. We are at the edge, to borrow a concept from the sage Kevin Kelly, a regular at TED, of a planetary “phase transition”. The classic phase transition we all learned about in high school physics is that which happens with the molecule H₂O. Below 32 degrees fahrenheit, it is a solid: ice. Above it, H₂O is a liquid: water. And at 212 degrees Fahrenheit (assuming you are at sea level), the “phase” of water transitions to a vapor.
A million quantum leaps beyond this high school example, and we are at the scale and scope of today’s pending transformation of reality. Data-driven, digital technologies are bringing about this planetary shift, of course. But the “Ur force” of the phase transition, as Kelly wrote in his 2016 book The Inevitable, is the arrival of artificial intelligence.
I’ll get to the relevance of walls and moats. And I really want to share my thoughts on the critical alternatives of filters, screens, and windows. But first let me preface with what is driving my thoughts here as I, like the rest of humanity, sip from the torrent of news and information on AI. And, in my role as the CEO of data.world, I’m closer to it than most.
I hardly need to repeat the stats or speed with which Generative AI, including the iterations of OpenAI’s ChatGPT, Codex, and DALL·E, its image-creating cousins like Midjourney and Stable Diffusion, or its many emerging rivals such as Google’s Bard, Perplexity AI, and so many others that have appeared in our lives. Its warp speed diffusion is ironic in the technology realm where entrepreneurs such as myself more often worry about the barriers – cultural, economic, even psychological – to the “S-Curve” spread of innovation. How many thousands of product discussions have turned on the hopes for “early adopters”? Remember the 150 years it took the British navy to adopt Captain James Lancaster’s successful 1601 use of citric acid (lime juice in his case) to prevent the disease scurvy that killed more sailors than warfare or accidents?
Not this time around!
Get ready for programming on steroids, ‘prompt engineers’, and lawyers
It’s hard for me to skip over the many accounts of the impact of this technology just in the first few months, of engineers boosting their code-writing output ten-fold such as described in Christopher Lochhead’s podcast on this. AI-assisted programming, in AI geek speak, is the first step toward “recursive self-improvement”, with daunting implications. We are witnessing this first hand at my company data.world. I’m eager to talk about “prompt engineering”, a job category that didn’t even exist a few months ago. There’s so much more to be said, and unfortunately litigated, on the intellectual property implications of technologies that can suck up and repurpose everything from ancient scripture, to last week’s city council meeting, to Johannes Vermeer’s 1665 masterpiece “Girl With A Pearl Earring”.
I’d love to share some of the summaries my daughter Rachel, who is joining me at TED for her first time there, has written using these tools. Or the beautiful images and homework hacks my son Levi has produced. I’m bursting to share the many discussions we are having about the quality, integrity, and governance of the underlying “training data” so casually referenced in passing in every ChatGPT essay, article, or blog.
For our team at data.world – where our data catalog is becoming the most widely deployed data tool of all time at our customers – it’s akin to watching a debate on the impact of the transport sector’s greenhouse gas contributions with virtually no discussion of the difference between the viscous bunker fuel in cargo ships, the refined petroleum used by most cars, the aviation turbine fuel used by jets, or the role of wind power enabling EVs. This is why we’re all keen to share our newly launched AI Lab.
Is the torrent of analyses and commentary on AI lacking? You bet. But most urgently in need of attention on our critical path with these world-changing technologies is the maypole dance of would-be prophets of apocalypse pivoting around the now-famous open letter issued in late March. It calls “on all AI labs to immediately pause for at least six months the training of AI systems more powerful than GPT-4… If such a pause cannot be enacted quickly, governments should step in and institute a moratorium”. The italics around the part that is really bad are mine. When I last checked, the number of signers had grown to over 26,000.
To be sure, fine minds are on the list, including that of its initiator, MIT physicist Max Tedmark whose argument for a moratorium is vivid and compelling. We are at the most important fork in the road in the history of the human species, and we could get it very wrong, he argues. So it’s not without pause that I argue that this call to hit the brakes, and wall off this technology until we get a better handle on its trajectory, misjudges reality.
The real issue we’ve yet to confront is analog regulation in the digital age
Don’t misunderstand. I’m not in some dogmatic anti-regulation camp. From the brave FDA regulator whose call on thalidomide in 1960 spared Americans the thousands of birth defects suffered by European children, to February’s near miss between two planes at my hometown airport in Austin, there’s no shortage of evidence for the role in modern life of government regulation.
Rather, the issue we’ve only begun to confront is analog regulation in the digital age. As authors Chris Wiggins and Matthew L. Jones point out in their marvelous new book How Data Happened – A History from the Age of Reason to the Age of Algorithms, ours is an age where at a congressional hearing…, “one confused senator asked (Mark) Zuckerberg if Facebook was ‘the same thing’ as Twitter, and another senator asked Zuckerberg to clarify if the firm sells ads – their primary business.”
What could possibly go wrong? We’ll soon find out.
In response to the open letter, just last week the Biden administration took the initial steps to ramp up and re-tool its still vague “AI Bill of Rights”, first unveiled in October. The Europeans are further ahead, as they have been on data privacy. A draft of the EU AI Act was first unveiled two years ago, and like the “GDPR” privacy regs, it is more nuanced than anything discussed on this side of the Atlantic. But following the moratorium call, the heavily-lobbied rules are again up for grabs amid a demand for much stricter regulation of yet-to-be-defined systems “which may become high risk”. Not to be outdone, China also responded last week, issuing draft regulation days ago which – surprise, surprise – cover just about every dimension of Generative AI imaginable, including the imperative for all training data to reflect the “core values of socialism”.
In sum, I do share worry of the risks and disruptions entailed in this “phase transition”. There are many. On balance, however, I’m more concerned with what happens if we don’t thoughtfully move ahead on a technology that promises so much – from accelerated mitigation of climate change, to drug development, to access to education and legal services, to what I believe will be an unprecedented entrepreneurial boom… and so much more.
If that’s not enough, consider how a moratorium will deepen our existing failure to keep pace with China, our biggest rival. “In nearly every major aspect of computing, from hardware manufacturing to algorithm design to advanced computing systems, American leadership is waning or already gone,” wrote a trio of scientists in a riveting series last February for the Georgetown Public Policy Review.
Lastly, a moratorium is simply unenforceable. While this view has many adherents including Bill Gates, the best argument was inadvertently made as I began working on this post and news dropped that Amazon will soon purvey IKEA-style, do-it-yourself AI kits based in the cloud servers of AWS. The online retailer’s “Bedrock” will give customers “easy access to foundation models (FMs)”, which they can customize with their own data and deploy in their apps. There goes your pause, folks.
So, I rest my case on the peril and meaninglessness of a moratorium. But that doesn’t mean we shrug and remain blind to real and potential dangers, that realistically include not just jobs disruption but exponential disinformation, and a widened digital divide of many dimensions.
Let’s shape AI with transparency, sandboxes, and a commitment to public benefit
Heading into TED where I know I’ll learn vastly more, I have three ideas on my mind.
First, let’s “build this in public” as OpenAI CEO’s Sam Altman urged in conversation with podcaster and new Austin resident Lex Fridman: “We believe it is important for the world to get access to this early to shape the way it’s going to be developed, to help us find the good things, the bad things.” This approach will certainly be ours at data.world and I’m eager to hear what OpenAI President Greg Brockman will say at TED. When Brockman spoke two years ago, I immediately reached out to him to get data.world early access to OpenAI’s GPT-# to advance our industry. A public example that I heard about today on the Hard Fork podcast was the emergent behavior of generative agents in “Smallville” - it’s fascinating and the derivatives of SimCity will never be the same. The applications for entertainment, including gaming and experiential TV, from just this one example are virtually unlimited.
Second, I hope we can have a broad discussion on the emerging concept of “sandbox regulation”. A concept developed in Britain, initially for financial regulation, it’s an alternative to the “no until we say yes” model of analog regulation. Essentially, it’s a limited time and space parameter for innovators and regulators to jointly test and define the rules before they are made permanent. While not a “sandbox” play per se, the science in the fast lane “emergency use authorization” for pandemic vaccines – that took us from lab to a shot in the arm in nine months – is a good example.
Third, let’s consider the public benefit “B Corp” model by which companies legally obligate themselves not just to shareholders, but to stakeholders – the community, the environment, and society. Our company data.world has been a B Corp since our public launch on July 11, 2016. All companies developing AI should consider embracing this model.
We don’t need walls and moats to lock up innovation until declared ready. We do need filters and screens to sort out what works and what doesn’t, and windows on the process for all to look in. As my friend and fellow Austinite, the polymath Bryon Reese wrote in the seminal work on AI, Stories, Dice, and Rocks that Think, we are as Hernan Cortes in 1519, burning his ships in the New World so his soldiers could not return.
“We too have burned the ships. We have built a world without an undo button,” Byron wrote, “We cannot go back.”
So, let’s go forward! I look forward to learning more and discussing this profound shift at TED.
Header Image: generated by Levi Hurt in Midjourney. Levi’s entrepreneurial dream is to build a Dyson Swarm, as he discussed on the Deep Future podcast.