The general atmosphere around AI has become much more alarmist over the past couple of months. With news of rogue AI agents escaping into the wild to wreak havoc or Anthropic claiming they have blocked several attempts to develop biological weapons, we are now squarely in the “This is just too big” era of AI. This is the same space occupied by climate change, a problem so big and complicated, with so many vested interests preferring the status quo, that it seems impossible or at least improbable that there will be any action to satisfy the urgent need for regulation or protection or oversight or, really, anything that minimizes the risk to humanity and the world at large.
Opinions on the ethical and financial value of AI are wide-ranging, touching on everything from NIMBYism and global markets to ethics and the human condition. Worries about the governance and environmental impacts of data centers are the current political hot button in our country. The dialog around job destruction has subsided for the moment as some evidence points to AI creating new jobs in the short term – though I fear the long term will be more dire for employment. The moral and philosophical debates on its impact on art and humanity continue and they are amongst the most important of all discussions but perhaps too esoteric for regular dinner table conversations. And the bullish excitement about rocketing stock market values combined with an anxious, hope-for-the-best and don’t-bother-planning-for-the-worst hand-wringing about a potential bubble is an emotional roller coaster for anyone with a retirement fund.
It’s an all-encompassing topic and the future keeps arriving faster than we can handle. With AI frontier labs churning out more and more advanced, and more and more energy-hungry, models with more and more concerning consequences, I can’t help but wonder what is their purpose? Anthropic states on their home page that “[a]t Anthropic, we build AI to serve humanity’s long-term well-being.” Their purpose: “We believe AI will have a vast impact on the world. Anthropic is dedicated to building systems that people can rely on and generating research about the opportunities and risks of AI.” Meanwhile, Open AI “is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.” Despite both prioritizing the idea of research, they are also companies, with revenue and operating costs and margins and so forth. They are both planning IPOs within the next six months, which will bring even greater scrutiny into their financial performance and lead to a reckoning of their balance between markets and mission.
I see the frontier labs continued drive to build more advanced models as bad business practice. Adoption of AI across public sector and commercial organizations remains relatively slow, with military and intelligence organizations being the most avid. Moreover, the current state of models is more than good enough to meet almost all of the use cases that companies would benefit from. Putting it in terms of Anthropic’s Claude models, which are the ones I most use, Sonnet is sufficient for business use with Opus, let alone Fable, adding little additional value. So the frontier labs are essentially spending huge amounts of capital on models that have limited practical value for their enterprise and business customers, organizations whose continued adoption of (and payment for) AI is financially existential to the continued existence of the labs.
How will the markets respond? Can the frontier labs, once public, balance the pressure for increasing revenue with their ambitious appetites, and outsized costs, for cutting-edge research? Are the advanced models flooding the market with too much supply for something that has little demand? Perhaps I lack the necessary imagination but I consider myself fairly fluent in AI usage, certainly more so than most of my peers in the corporate world, and I find Opus to be rather disappointing. If and when I am in a position to procure technology for an organization, I know that Sonnet is fully sufficient for almost every use case and Opus would be a buy for perhaps a small number of individuals to experiment with – certainly not needed for the enterprise, especially at 2.5x the cost. Businesses can derive immense value from Sonnet and they should be using it (or other similar models from other labs). Opus? Pass.
Of course, many companies have been able to successfully combine business and research in all sorts of industries, like pharmaceuticals and manufacturing. I have worked most of my career at IBM and, across its various incarnations, it has a legacy that continues to this day of running an IBM Research division, which is currently exploring the cutting edge of quantum. The difference is in the stakes. The amount of money that is tied up in the frontier labs means they must be financially successful and because of their operating costs the target for financial success is a huge number. They have already become too big to fail and that is before becoming public. The consequences of failure for other companies is implosion; for the AI labs, it is explosion, a domino effect with an outsized impact on the world economy.
There have been so many eloquent and well-reasoned arguments to slow down the pace of AI development that focus on the societal, geopolitical, and environmental consequences. I tend to agree with most of them. But, counterintuitively, I also think we should slow down the pace of AI development for economic reasons as well. The pace is too fast even in our short-term, quarter-to-quarter world, making it increasingly difficult for the business world to effectively adopt and deploy the technology. We are so inured to the capitalistic drive for growth, growth, growth but this is a classic example of the need to walk before running. If you don’t allow customers to walk first they may just throw their hands up and say no mas – and that’s not a good thing for the margins these companies must maintain.
One of my favorite ideas I’ve read is to split these companies up into a profit-making enterprise and a research arm, with the latter being absorbed by the academic community. Of course, with our universities under unparalleled attack by myopic political forces, this idea is likely just a pipe dream. I still find it most compelling. The profit-making side could focus on tweaking models and building tools for their customers, and bringing more advanced models to market more thoughtfully, in the same way that technology companies release upgrades and new versions. And the research side would reside within a world that is more accountable and more collaborative, distributing control of the future of AI whereas in the current state it is concentrated in the hands of a very few and increasingly unstable men.
Some day, activist investors will likely force a spinoff of the research arms of the labs anyway, so why not get ahead of it with a win-win proposition of benefits for business through more focus and a more controlled and transparent development for AI, where it can actually benefit humanity. The alternative, absent any serious regulation or geopolitical compromise, is a light speed trip to financial disaster or a world where the oligarchs end up making use of those underground bunkers they have been building…
– – –
Moments after my initial draft of this post, I read two interesting and relevant articles. The first covers how an OpenAI model solved one of the Millenium Prize Problems (in mathematics). This report covers a lot of interesting angles. Two thoughts I want to draw out as they relate to what I wrote above. First, if you are a CIO, your reaction would likely be that’s cool… but so what? Is a model powerful enough to solve the Navier-Stokes model really what you need to improve your business process and go-to-market capabilities?
Second, the article notes that the estimated cost for solving the problem was $15M. This is an example of the tension between business and research. This may be $15M well spent for the advancement of knowledge (though many of the mathematicians in the article make excellent refutations to that claim) but it is $15M poorly spent for business. A traditional McKinsey consultant would slash this line more quickly than underfunded school districts axe their arts programs.
The second covers an essay published by Dario Amodei, the CEO of Anthropic, urging an AI slowdown. This report (behind a paywall but other sources widely available I’m sure) summarizes his points. According to the summary (the essay itself is 3800 words long, twice as long as this post), his main argument is safety, one of the most important and compelling arguments. But one can’t help but wonder whether Dario Amodei and his company may be the worst actors of all. We hear Anthropic continually publish warnings about the dangers of AI and make the most frequent calls for action on regulation, but they always externalize the need instead of actually doing anything themselves. Once widely seen as an admirable counterweight to the psychopathy of Elon Musk and emotional instability of Sam Altman, Anthropic has proven itself to be just as ego-driven as their competitors, throwing in fistfuls of pious hypocrisy into the mix as well.
– – –
Note: This post is also available on my AI-focused site at https://dorukai.dorukakan.com/.