On the occasion of the World Wide Web’s 25th anniversary (2014), Tim Berners-Lee, the creator of the “World Wide Web”, was openly calling for an “online Magna Carta”—a sort of bill of rights for the web. The fundamental right to open, de-centralized access to the web was under threat by governments and corporations. And this did not sit well with him.
In effect, Berner’s-Lee stance amounted not to a repudiation of the web for what it was becoming (less open, centralized), but a rejection of power plays—the threat of powerful interests capturing and subsuming the usefulness of the Web, created for the benefit of society. He described the danger as something that “crept up on us.”
Fast forward to 2018 and Berner-Lee, speaking in Lisbon, acknowledged his disillusionment with the web. “Humanity connected by technology on the web is functioning in a dystopian way.” He cited online abuse, fake news. political manipulation, prejudice, etc. In another 2018 open letter, he wrote how the Web could be “weaponized at scale.”
Fast forward 8 years and yet another threat has emerged. Only Berner-Lee’s alarm has been turned on its head.
Alarm As The New Norm
Over the past week, we have witnessed the latest in a series of proclamations by AI “leaders” who seem determined to define and drive a narrative of alarm.
Dario Amodei’s essay was the main driver, followed by Sam Altman’s agreement, Musk, et al. Essentially Amodei called for a slowing of the pace of AI development so that safety mechanisms can catch up. In other words, it’s moving too fast to contain. “Amodei warned that within six to 12 months AI could be capable of leading a swarm of agents that could take over the entire internet.”1
Still others, including AI pioneer Stuart Russell, are concerned over the threat to humanity. The list of concerns range from racial and gender discrimination to jobs displacement, misinformation and emotional attachments. An Anthropic researcher resigned over the dangers posed by AI and how Anthropic and OpenAI are too busy competing to be concerned over safety. His grave conclusion is that AI could destroy, not just the internet, but us all!
How does a company with ample cash and seemingly unlimited access to experts become “too busy” for safety? It’s like saying that Henry Ford was so busy getting the engine to work on his Model T, that he didn’t have time to inflate the tires. It’s ludicrous. And hyperbolic.
Grandiose, hyperbolic, alarmist claims, are cues to pause and listen—past the noise—to the signals.
Are AI “leaders” genuinely trying to prevent mistakes, or are they trying to shape the rules of the game for their own benefit? Or both?
Who Exactly Is Going Rogue?
If you read about AI chatbots or agents going “rogue,” do so with a measure of skepticism. When you hear about how Chatbots or AI agents are becoming sentient (“self” aware), please ignore. Machines are machines and always will be. Personally, I find it incredibly condescending when AI “technologists” speak in terms that they know are not true (or wish to be true), such as “conscious” machines, merely in an effort to establish their preferred narrative.2
The imminency of “superintelligence” is a contemporary myth. The idea that simply scaling LLMs will lead to these machines programming better versions of “themselves” is simply, as Cal Newport puts it:
“…just a warmed-over version of the recursive self-improvement trope that Silicon Valley futurists have been peddling since the 1960s ; a rhetorical crutch that lets you somberly discuss sci-fi scenarios without actually engaging with relevant technical details.”
However, I am NOT saying there is nothing to worry about. There are existential risks. And, those risks are primarily related to the pursuits and activities of the frontier AI companies themselves. Not bad actors, per se. But rogue technologists!
Of course, we should have serious concern over the actions of these companies, the risks of runaway technology and bad actors bent on misusing it. And, of course regulation is necessary. However, AI powerhouses would have us think they can somehow collaboratively (and quickly) create guardrails alongside regulators who are as much in the dark about the technology behind and training of AI models as the general public.3
In case we miss the point: the people telling us to slow down the pace of AI, are the people building AI! To whom are they directing the question? Just slow it down, then.
But, can a company (or country) whose competitive survival depends upon advancing faster than its rivals actually choose to advance more slowly? Of course not.
So, who is actually going “rogue?” CEOs playing off of a script that is being written in the moment—in which, they star as the main characters! Somehow they fail to see that they are part of the problem.
Power Play > Power Grab
This is, at least in part, a power play. There are several companies in the running. Accumulators, gathering a base of users, and dominance in the public sphere, as quickly as possible. But the real “prize” goes to the company first to AGI (Artificial General Intelligence).
The power play is another form of a power grab, in three possible ways.
First, frontier AI leaders prefer to quash schemes that allow “open source” AI models a larger market foothold than is deemed appropriate. It’s referred to as “regulatory capture”, which squeezes smaller and/or “foreign” actors from gaining too great a foothold in any particular market. (“Open source” is a topic for another day. Except to say that “closed” systems—ChatGPT, Claude—keep control in the hands of the few.)
Second, by “asking” to be regulated, in effect, AI companies (a few of which are preparing for multi-trillion dollar public listings) are seeking: a) both a legitimate means of buffering themselves from Anti-Trust violations (of which there is a growing logjam), as well as private litigation. The “guardrails” are intended not merely for their own products (for which they are responsible), but to minimize the potential of public and private litigation. And, b) political persuasion. (Whichever “party” comes to their “rescue” is the big winner at the upcoming midterm elections and beyond, especially given the increasing negative sentiment towards AI.)
Finally, this power play is intended to usher in authoritative command over the product. Whoever complains the loudest has the upper ethical hand, and the perceived right to author change.
The power and authority to shape the future of AI development is, in all probability, the greatest sphere of power there is at this moment in time. The grab for “authoritative” AI—the ultimate form and direction of AI development—is of the greatest value. Attaining AGI is a component of that. But it is the design of AI that should matter most to the future of the human race.
That may, itself, sound hyperbolic. But consider: AI systems do not have rights, feelings, or consciousness. And machines must not be trained to act as though they do. Yet, Amodei’s Anthropic—the very company that called for “pacing”—is driving the development of AI to have just those attributes. Right. Feelings. Consciousness. One need only read the Clause Constitution to grasp this point.4
In other words, the stage has been set. Any window to reverse the trajectory of AI model development is closing fast.
I do not believe that Anthropic or other Frontier AI companies are pursuing this course in premeditative fashion nor from unethical motives. I think they genuinely believe that this is the best course of action—that by, training AI to have feelings (an over-simplification of what is really going on) is a worthwhile endeavor.
But the course they are taking, despite all of the “transparent”, public-facing essays being written by AI leaders, is being made in a silo. And, it turns out that training AI on human attributes can lead to bad things.
Or to put it another way: a private corporation is acquiring the power to define the values of an increasingly powerful artificial agency system that probably will have enormous social influence. Where is the democracy in that? And more importantly, by whom is the responsibility and accountability to the human race to be defined?
Furthermore, the pursuit of social benefits (e.g, medical breakthroughs), via the advent of more powerful AI models, does not justify activities that serve as a potential threat to society.
Reckless Behavior
Unless the irony is missing, perhaps it is worth noting, as an example, how reckless AI companies can be about the “safety” they are supposedly looking to instill.
This past summer rogue attacks by an OpenAI experiment (the “HuggingFace” test) took a “relatively common technology called an LLM-powered agent, which is used with minimal issues by millions of software developers every day, and then added a bunch of dangerous features, removed restrictions, and released a whole mess of them into poorly fenced digital wilds.5 Why? They were desperate to make as much progress as quickly as possible on a hacking test that they felt was an important milestone en route to superintelligence.”6
Not only do I agree with Newport’s conclusion, but would add that these are not only reckless, but completely irresponsible activities. And OpenAI is not alone.
The very first “guardrail”—the most obvious one—is to act responsibly!
Again, how do we expect companies who consistently practice irresponsible behavior to establish “guardrails” over their own activities? Why is this a necessity at all? Should we conclude from this that their primal urge to “win” at the game utterly forestalls any modicum of moral responsibility?
Ruled by a Moral Framework
Berner-Lee did not conclude that the web was inherently dangerous. On the contrary, he commented that “the web has been diverted from its original social purpose.” His “Contract for the Web” frames the web as a public good, with distributed responsibilities amongst government, corporations and citizens.
Neither is AI inherently dangerous. Humans seem bent on making it dangerous, simply due to the overzealous urgency of advancing models and achieving superintelligence!
Tim Berner-Lee’s position, which played out over decades, should help reasonable people to agree:
Technology requires more than just a technical, regulatory framework. It demands a moral framework.
Only, that moral framework must never be constructed solely by powerful interests determined to acquire and consolidate power over the technology, regardless of its potential benefits to society.
Do not expect the AI “pace” of development to slow. But let’s hope that AI companies make a moral choice to prevent irresponsible behavior within their own development environments. That’s almost a prerequisite to establishing guardrails (rules). And, that they send a clear signal—with less noise—that AI requires, and its users deserve, a moral framework.
Rules and regulations without a moral framework are destined to fail.
This publication is for brains, not bets. The Other Side of Obvious shares ideas, stories, and general financial information—not personalized investment, tax, or legal advice. Investing comes with risk (including losing money). Talk to a pro before you act. Please take time to read these important disclosures before you get started.
AP News. Take it over from whom? It is these very same AI companies that have swallowed vast amounts of the internet (i.e., akin to a “take over”) to feed their LLMs. This conclusion is a “red herring” (i.e., meant to be distracting). Who builds a product that could potentially make the internet (the primary training source for LLMs) redundant and therefore useless in its ability to improve and deliver that very product?
More commentary on AI “consciousness” to follow.
This “darkness” is part of the issue, in my opinion. There is too much alarmist announcements, and too little technical details. For example, the recent so-called mathematical “achievements” AI companies are lauding, have little in the way of verifiability, simply because they aren’t collaborating (at least not publicly) with mathematicians and other researchers—which is considered normative in the world of research. AI companies are performative rather than informative.
The Claude Constitution is not only a product of Amodei and Anthropic’s vision, but has been fed to their AI systems as part of AI training. In other words, AI machines are being trained by humans to act human.
I highly recommend listening to this podcast, which explains what occurred.