The Westinghouse Contingency
3
*This essay was submitted to the* [*AI Progress Essay Contest*](https://www.metaculus.com/project/ai-fortified-essay-contest/)*, an initiative that focused on the timing and impact of transformative artificial intelligence. You can read the results of the contest and the winning essays* [*here*](https://www.metaculus.com/questions/11429/results-from-the-ai-progress-essay-contest/)*.* Leaders in artificial intelligence (AI) typically take one of three distinct approaches to talking about the future of the field, either alarmist (for example, Elon Musk), utopian (for example Sam Altman or Sundar Pichai) or messianic (for example, Geoffrey Hinton). While each of these approaches is very different, it is still the case that there is one belief central to all three narratives, namely, that artificial general intelligence (AGI) is inevitable. In late 2020, [Geoffrey Hinton made the bold claim](https://www.technologyreview.com/2020/11/03/1011616/ai-godfather-geoffrey-hinton-deep-learning-will-do-everything/) that “deep learning is going to be able to do everything.” Sure one might dismiss this as pure hubris. But there is no meaningful mainstream counter narrative to push back on such a clearly hyperbolic, self-serving claim. As such, this is an AI Safety risk that needs to be addressed. This will be the main focus of my essay. The conversation about “AI Progress” has become myopic. It is all built on one foundational yet potentially unwarranted belief, that we have been progressing down the optimal path toward achieving AGI since the emergence of deep learning. The Metaculus community apparently believes this to be the case, as evidenced by seventy (70) percent predicting AGI will be based on deep learning. The original goal of developing AGI was to achieve human-level reasoning or general intellectual ability. But the systems that have been developed are not capable of understanding syntax or forming semantic representations, and are incapable of making the same kind of inferences humans do. Oh, and also they cannot formulate or pursue long-term intentions or goals. When we anthropomorphize the current generation of AI based merely on statistics, we might make them more marketable, but these anthropomorphic analogies ignore the fact that the fundamental ways in which these AIs learn bears no resemblance to our own human acquisition of complex cognitive processes, such as we do in our language development. Consider the point Melanie Mitchell made [when she mused](https://www.quantamagazine.org/what-does-it-mean-for-ai-to-understand-20211216/) that ”understanding language requires understanding the world, and a machine exposed only to language cannot gain such an understanding.” The Metaculus community still thinks AGI is at least two decades away, but when considered alongside the above referenced forecast that AGI will be based on deep learning there are two clear implications. The first is that it is generally agreed upon that more data and greater computing power are the necessary and sufficient conditions for achieving AGI. The second implication is that generation after generation of AI researchers will merely do the work of scaling the existing deep learning paradigm until some unknown threshold is surpassed that will make it apparent AGI has been realized. In other words, there will be no additional paradigm changing breakthroughs or radical thinkers on the path to achieving AGI, just bigger data and computers. Many (if not most) mainstream claims about AGI are closer in epistemic status to *hopes* than *justified true beliefs*. And this makes sense given the personal investment of many in the field. There are few serious counters to Hinton’s claim that “deep learning can do everything”, and the few that do have the courage to speak up are typically either ignored or [ridiculed](http://www.cs.toronto.edu/~hinton/marcusquote.html) by leading figures in the AGI community. With each advancement the deep learning fanboys love to bask in the glory of their accomplishments while simultaneously making sure that their critics think that they have been proven wrong.  Let’s be clear that Altman is doing his part for the movement, and he does it really well. He has been very savvy in his approach to unveiling his company’s creations in ways that dictate the terms of the narrative. But under his singular leadership OpenAI has assumed a very different tone in their communications, particularly on the subject of AI Safety. This makes it apparent that Altman’s main aim and ability is marketing, and not the safe deployment of AIs. People seem to have forgotten about the abrupt shift in the way OpenAI talked about AI Safety between the release of GPT-2 and GPT-3, at which point AI Safety research devolved into two camps with competing philosophies. One group led by Altman has shifted the conversation away from the immediate risks of new systems to the long-term risks of superintelligent AGI that tries to compete with humans as the dominant “species” on Earth. The other, led by Jack Clark, [left OpenAI to start Anthropic](https://aibusiness.com/document.asp?doc_id=769951#:~:text=Anthropic%20was%20founded%20by%20OpenAI's,%2C%20co%2Dfounder%20of%20Skype.) to continue working on the immediate risks of new systems by trying to understand and interpret how deep learning does what it does. The current hype cycle began in February 2019 with [OpenAI’s clever unveiling of GPT-2](https://openai.com/blog/better-language-models/#sample8) which they (now famously) speculated to be the [“most dangerous language model of all time”](https://blog.floydhub.com/gpt2/) and subsequently judged to be (at least initially) far too dangerous to reveal publicly. It was a tantalizing suggestion and at the time, in large part because of OpenAI’s prior association with Elon Musk (a co-founder who has since severed all ties with the group), the claims were taken seriously (and at times even assumed seriously). This began a very important and perhaps [long overdue debate](https://www.theverge.com/2019/2/21/18234500/ai-ethics-debate-researchers-harmful-programs-openai) about [what exactly “too dangerous” means](https://slate.com/technology/2019/02/openai-gpt2-text-generating-algorithm-ai-dangerous.html), particularly when it comes to technology. How do we decide where to draw the line as we flirt with god-like power, or at least the illusion of it? Who gets to make that decision? The concerns raised at the time seemed reasonable in light of the radical advancements inherent in the technology they were releasing. The Metaculus community forecast for when a language generation model will yield approximately human-level output indicates that we may be on the precipice of an era when you won't be able to distinguish between humans and bots when communicating digitally. There are new things to be worried about in a world in which bots are indistinguishable from humans. Then suddenly in November 2019 OpenAI reversed course and decided to [release the full GPT-2 1.5 model](https://openai.com/blog/gpt-2-1-5b-release/) and datasets to the public. It was only eight months later when OpenAI [unveiled GPT-3 and the OpenAI API](https://openai.com/blog/openai-api/) to enable developers to embed their most powerful creation to date in their applications. Getting approved by OpenAI was apparently not entirely straightforward or democratic, with groups that successfully hacked the obscure process posting instructions for others to follow that amounted to an instruction that ‘if you know someone and are convincing enough they will let you cut the line’. [According to OpenAI](https://openai.com/blog/gpt-3-apps/), as of March 2021 more than 300 of their closest friends and family had integrated it into their applications, and then finally in November they eliminated the waitlist and [opened GPT-3 to the entire world](https://openai.com/blog/api-no-waitlist/). Now the OpenAI CEO unveils his company’s new AI breakthroughs with statements that indicate no sense of significant concern about safety implications.  Other leading figures in AI take tactically different but equally hyperbolic approaches to talking about the long-term implications of AGI. Sundar Pichai is not afraid of making high profile hyperbolic predictions about the future of AI. I guess when you are Elon Musk's peer, stealing the spotlight requires some degree of 'hypemanship'. But his company is also notoriously intolerant of anyone speaking up with an alternate narrative inside the company. When Google is met with legitimate questions and challenged on the ethics of their policies and actions they frequently turn to [bully tactics](https://www.wired.com/story/google-timnit-gebru-ai-what-really-happened/) to [silence their dissenters](https://techcrunch.com/2021/02/19/google-fires-top-ai-ethics-researcher-margaret-mitchell/).  Musk is generally very alarmist when speaking about AI (at least) publicly, but at Tesla he constantly pushes the limits with respect to his risk tolerance for deploying fully autonomous systems in high stakes use cases. In this way he at times resembles a bull in a china shop, moving fast and breaking things.  **What To Do Moving Forward?: The Westinghouse Contingency** The religious-like devotion among leading AI figures to the idea that deep learning (DL), big data and greater computing power is the obvious an inevitable recipe for achieving machine sentience and Artificial General Intelligence (AGI) is reminiscent of the story of the invention and development of electricity and its infrastructure and the competing ideas of Thomas Edison and Nikola Tesla for the future power infrastructure of the world. Tesla came to America idolizing Thomas Edison and quickly found himself working for Edison in Menlo Park building Direct Current (DC) motors. Tesla saw the obvious flaws in the DC-based system for electricity, which was so expensive and inefficient that each city block in New York required its own dedicated power plant. In other words, DC was a technology that was truly indistinguishable from magic, but also impossible to scale. If Edison’s closed system were to win out, electricity may never have made it beyond the city limits, and most of the world would have been left behind. Tesla had a vision for a new system called Alternating Current (AC) which Edison had tried but abandoned immediately and completely once he figured out how to make DC work. Yet, when he took it to Edison he was shocked when he was dismissed and his innovation shelved despite its obvious advantages. Edison’s self-certitude and competitive drive had been vital to overcoming tremendous adversity in the face of doubters early in his career, giving him the edge he needed in pursuit of his grandiose vision for a technological future. But at the height of his fame and power, he surrounded himself with nothing but sycophants willing to take orders. He was no longer willing to follow the best ideas wherever they took him, instead it was his way or the highway. These traits quickly became a liability. Edison’s obsession with beating Tesla and building a completely closed system based on his DC design led him to abandon all good sense. With the financial backing of JP Morgan, he set out to destroy Tesla and his financial backer George Westinghouse, no holds barred. Let’s remember that Edison was not fighting for survival, even if his DC design wasn’t chosen as the standard for the electrical infrastructure he still would be one of the greatest beneficiaries of the electrification of society as the inventor and monopolist of the light bulb. This point is particularly relevant to the current discussion of AI Progress and the dangers of myopic approaches to complex problems. The dismissiveness Fortunately, Tesla recognized the inherent flaws in Edison’s DC design and took it upon himself to design a more efficient and scalable alternative. We are lucky was also incapable of conforming out of mere reverence, the science was all that mattered, and he had the temerity to tell the old man he was wrong to dismiss his ideas and set out building his own version of the future. But Tesla’s genius wasn’t enough. His name and his alternative vision for the electrified future would probably have been lost to history were it not for George Westinghouse . "George Westinghouse was, in my opinion, the only man on this globe who could take my alternating-current system under the circumstances then existing and win the battle against prejudice and money power," Tesla wrote. "He was one of the world's true noblemen, of whom America may well be proud and to whom humanity owes an immense debt of gratitude." I propose that the current paradigm based on layers of neural networks, petabytes of data and unsustainable growth in computing power is analogous to Thomas Edison’s DC; impressive and praiseworthy indeed, but also unsustainable, at least at the scale required for AGI. If Nikola Tesla hadn’t been such a headstrong and fearlessly defiant young immigrant, willing to tell even his hero when he was wrong at great personal cost with no guarantee of We are long overdue for serious investments in an alternative vision for AI, what I will call the Westinghouse Contingency, in recognition of the courage and conviction of George Westinghouse to take a huge risk in financially backing the bold vision of a young Nikola Tesla, a vision which directly undermined that of the world’s most famous inventor and Tesla former employer Thomas Edison. The Westinghouse Contingency would be a concerted plan to develop alternative approaches to AI that hedge the massive bets currently made on a single technology. This could be thought of as akin to a Manhattan Project. But government run projects don’t have the benefit of a world class salesman behind them to give them a fighting chance in their battle against the incumbents. There are plenty of idle billionaires floating around, which one of them will answer the call to be this generation's George Westinghouse?