The AI Race: If you're not first, you're last?

[We’ll resume the ethical considerations next post]

You see it in headlines from top news websites: the US, or startups, or established players, or your mom’s third cousin on her father’s side twice removed, is in an AI race. This is nonsensical.?

Saying there is a "race" implies that someone must "win" AI. This makes about as much sense as telling leaders we need AI to "finish strong!" or that AI must "push through" to bask in glory. If this is a race, remember to chant this to yourself as you grit it out: pre-training runs are temporary, but inference is forever!

But there is no “finish line”. We have no way of knowing who is “winning” or when someone has “won”. It’s like saying there is a breath-inhaling race between the US and China. Everybody, breathe harder!

But more troubling than endlessly parroting a nonsense phrase that does almost nothing to clarify what’s going on, why it’s important, and how it’s unfolding is that adopting the phrase has strong implications for how readers and policymakers should react.?

Nobody wants to lose a high-stakes race. And as I’ve written elsewhere, AI company leaders have compared GenAI to the advances of fire, steam engines, electricity, and nuclear power.[1] Not only that, but even leaders of nonprofits also focused on AI have agreed the hype is 100% justified. There aren’t many stakes much higher than winning who will get to control a technology that could be as fundamental to civilization as fire was to humanity.?

There are at least three major assumptions that underlie the AI race (innovation = progress, GenAI will be meaningfully useful in the near term,?and “winning” is important for values) and those assumptions lead to at least three major reactions (need to spend a lot, regulations should be limited, we should implement quickly and broadly).?

Assumption 1: AI = Innovation = Progress = Good; Ergo, AI = Good

Inherent in the belief that we should “race” toward some AI finish line is that AI is innovation and innovation is progress, and progress is always good. There are, of course, many bad uses of AI, but let’s set that aside to focus on the even broader claim that innovation is itself innately good. Yes, innovation can be good, but that’s not an intrinsic feature. Rather, innovation is generally value neutral. For example, there has been lots of innovation in virtual reality over the last 30+ years. How often do you chill in the metaverse? See lots of people doing anything meaningful there with any regularity? Didn’t think so…

Can AI be a force for good? Absolutely. Can it also lead to harm? Absolutely. Is it clear which use case will prevail if we race toward it versus approaching it thoughtfully? Not at all.?

Assumption 2: Near-Term Meaningful Usefulness

Of course AI has enormous potential to be wondrous. But that’s different from actually being wondrous and useful in a large-scale, meaningful way.?

Most of the most interesting uses shared online, even by so-called AI hype men (er…influencers…) like Ethan Mollick, consist of inconsequential outputs.[2] It can role role-play in a video game (with errors)? Cool. It can “speak” in different voices on command? Truly cool. We can use it to edit videos to change a few lines? Truly interesting (and perhaps you’ve already noticed the potential for misuse). You can turn images into video but it’s “far from perfect or controllable enough for serious use”? Still truly cool. It can write poetry that meets exacting specifications ? Truly cool. Honestly, these are all fascinating and can feel a bit like magic to the non-jaded. They are impressive engineering feats, fun to experience, and they could ultimately lead to some profound use case (though whether it will be profoundly beneficial or profoundly harmful is unclear at the moment).

But some of Mollick’s posts are just not accurate. LLMs aren’t good at grad school just because they can do well on the GRE . (If you think grad school = the GRE, you’re in for a rude awakening. But also, we don’t know what data Llama (the model referenced by Mollick) used, so it’s possible it was just data leakage driving the improvement.) We now have widespread access to driverless cars ? Would be cool, but it’s not true unless you live in a specific geo-fenced region (which has been the case for many years, btw). Need help with the LSAT? Consult Claude 3.5 (there is no evidence this is a good idea just because the model did well on that exam). Self-reporting surveys say LLMs decrease working times in 37% of tasks? Very likely not an accurate reflection of reality for most tasks at most jobs (if it was, you might expect bosses everywhere to force most employees to use it for most tasks). It can accurately perform spreadsheet tasks 80% of the time and that “may be the start of something quite significant”? Maybe. But also, maybe not.[3]?

                                                       The 80% Solution
The spreadsheet example above is like a lot of AI hype. ChatGPT, released in November 2022, could output human-like prose and people readily extrapolated that to how AI might soon write award-winning literature. But now, 20+ months, many billions of dollars, and an endless stream of podcasts, news anchors, news articles, tweets, and so on claiming rapid advances in AI later, we still don’t have AI that can output a captivating novella, let alone a novel-length masterpiece.?

It instead seems LLMs may be approaching the upper limit of its capabilities.

The top MMLU benchmark score from March 2023? 86.4. Today? 88.7.?
HellaSwag benchmark in March 2023? 95.3. Today? 95.4.?
ARC-Challenge in March 2023? 96.4. Today? 96.5.
GSMK in March 2023? 92. Today? 96.

We don’t think benchmarks are super useful as currently implemented (we don’t know the underlying data, the testing methods, etc. for most major models). Also, it’s not clear if great performance on a benchmark translates into real-world usefulness), but every major GenAI company broadcasts them with each release, so they’re fair game to point out.

How useful is an 80% solution? Not very, unless it's for low-risk tasks with high fault tolerance. But those uses are not the most important markets to help humanity (e.g., medicine). 

This isn’t to say that further tremendous advances aren’t possible. But it is to say that past performance does not guarantee future results. Often, it doesn’t. It’s at least as likely that we’ve already neared the upper reaches of current GenAI architecture as it is that there will be some incredible breakthrough. And even if there is such a breakthrough, it’s not clear it’d be cost-efficient or safe, and therefore it may not be useful for productive tasks.?

As always, capabilities should be considered separately from usefulness.        

Assumption 3: The AI Race Winner Will Control AI Values

Some may think the AI race is important because the “winner” will get to determine the “values” of the AI. Meaning, they get to decide how the AI responds to different inputs, and in the US we want US AI companies to determine the values output by AI at our workplaces and schools and government and hospitals, not French or Chinese companies.?

Of course, few media outlets pause to ponder why we should want any private entity to determine an AI’s values rather than, say, the people implementing or using the AI. And if the answer is that the AI companies will allow the implementers or users to pick their values, then it doesn’t really matter who “wins” the “race” for values purposes.

Moreover, if anyone thinks the US government, or hospitals, or military or Fortune 500 businesses would choose a foreign AI system over a US system that is, in all likelihood, only marginally worse if at all, they aren’t being serious. The security, privacy, and economic stakes are too high to trust another nation with such precious data personal, health, security, and commercial data.?

But suppose you do believe you are either going to win or lose the most important technological advancement in the last 70 years, as the media is happy suggest? What should you do? Well, obviously you throw cargo ship loads of cash at it. You also take long hesitations before even considering possibly passing new ones and you champion applying AI to everything so you can eek out every advantage.??

Reaction 1: Make It Rain

It’s been said that mo’ money mo’ problems. And this must be the reason OpenAI is in a hurry to incinerate up to $5 billion this year : no money, no problems??

Notably, OpenAI will be able to raise more money (unlike, say, most people living in poverty) because its investors must win the AI race. This also means all major cloud providers must spend heavily on the data centers to satisfy the needs of AI. Why? Well, because it’d be bad if Anthropic or Google or Meta wins. Or something like that.?

The real reason, we’d venture to guess, is that the company with the greatest adoption rate can turn that into more money via ads or subscriptions. And GenAI is not like fashion where you usually buy a variety of brands. Most people only need one GenAI solution, so locking in customers is important.?

That is, if you thought the race was about virtuously benefiting humanity, you’re probably mistaken. But man, some individuals will become even extra richer and if that’s not an American value, what is? In other words, we must encourage a US company to win the AI race so mostly rich Americans get richer, not those French or Chinese entities.?

Reaction 2: Encouraged to Consider Possibly Maybe Taking Initial Steps to Explore

The “AI race” is probably partly why the AI Roadmap from the US Senate AI Working Group hedges so intensely: all regulations are apparently presumed inherently bad for innovation (and, again, innovation is presumed to always be good).?

The Senate AI Working Group didn’t propose any plans with its report. Rather, it stated that it “encourages the relevant committee” to “consider” such minor itsy bitsy non-urgent concerns such as “legislation that both supports further deployment of AI in health care and implements appropriate guardrails and safety measures to protect patients…” and to “consider legislation that would provide transparency for providers and the public about the use of AI in medical products and clinical support services, including the data used to train the AI models.” (emphasis added)

Are you concerned about privacy or intellectual property? The US Senate is, too. Slightly. Again, the “relevant committees” are “encouraged” to “consider federal policy issues related to the data sets used by AI developers to train their models, including data sets that might contain sensitive personal data or are protected by copyright, and evaluate whether there is a need for transparency requirements.” (emphasis added)

If deepfakes are on your brain, no worries. The “relevant committees” are “encouraged” to “consider whether there is a need for legislation that protects against the unauthorized use of one’s name, image, likeness, and voice, consistent with First Amendment principles, as it relates to AI.“ Note, this is just to consider if there’s a need, not to consider what should or could be done about known issues of, say, images of teen girls being used on non-consensual sexual deepfakes who are discovering that few states have laws to protect them.?

Reaction 3: Implementation Nation

What else might you do if you think there is a race and failing to win will be devastating for you and your compatriots? Well, you try to make sure you keep abreast of it so you have an edge. What’s one way to do that? Shove it into absolutely everything you can, as quickly as you can, to the greatest extent you can. Do you brush your teeth? Do it with AI ! Want to walk faster? Use AI ! Want to encourage your kid to dream big and write a letter to their favorite track star? Have AI do it instead !

But why stop there? Why teach kids about how AI is made, best use cases, its limitations, the many ethical and legal considerations, and so on when you can just have them use it to do their work for them? Sure, the kids will only use the AI to brainstorm and create outlines once they realize it can also compose essays and answer quizzes. Right. *wink wink*?

                                                  GenAI and Education

Humans already over-rely on technology to handle physical tasks. This is why so many of us huff and puff when the escalator isn’t working and we’re forced to take the stairs. The muscles, cardiopulmonary, and cardiovascular system have atrophied. Is there any reason to believe that our cognition won’t similarly atrophy if we replace critical thinking with AI outputs?

This probably sounds like the alarmists who panicked over calculators and Google Search in years past, but those comparisons miss the point. Calculators didn’t write equations for you or give you all the steps. A calculator is useless in trigonometry unless you know how to do trigonometry. Google Search required people to open webpages and parse out the content. Without critical thinking, Google was useless for useful results.?

GenAI, on the other hand, does the heavy lifting for the user. You don’t have to know anything about Huckleberry Finn to have GenAI output an entire essay on it. Moreover, applying deep thinking about a topic to write a research essay is a different (and much more important) skillset than merely editing an AI output.?

You can expect a full article on this topic in the future…        

What Does It All Mean?

The important takeaway from all the above is that GenAI can do lots of superficially cool things that demonstrate a lot of truly impressive improvements from AI research, but GenAI is limited when it comes to doing anything useful at a large scale with high accuracy at reasonable costs. There is no high-quality research showing that GenAI reliably improves student outcomes, company productivity, life satisfaction, health outcomes, or any other meaningful metric. This, despite mountains of cash thrown into the AI furnace and breathless trumpeting from the media and hype men.?

What we need is to take a few deep breaths and spend more time researching and testing to figure out what works and what doesn’t, rather than creating something and tossing it over the fence for the public to figure out at great expense both financially (e.g., from implementing AI systems that are unproven–including monetary and opportunity costs) and socially (from misinformation, disinformation, undermining the essential fabric of trust in society, etc.).?

It’s possible to be supportive and optimistic about AI without treating it as if it will solve many of humanity’s thorniest problems in the near term. It won’t, unless we’re counting the superficial, like relationship chatbots. Curing loneliness with chatbots is like curing unhappiness with heroin.?

Technology alone will never be able to solve diplomacy (“ChatGPT, negotiate a ceasefire in the Middle East on behalf of all countries.”); zoning regulations; access to high-quality healthcare, education, and nutrition; racism, sexism, and all other isms based on features that tell us nothing about the content of the character of the person or group being judged; allocation of taxpayer funds; prioritizing economic, defense, transportation, communication, and energy policies; or any number of the most pressing problems. AI is a tool. It’s not the solution, because for all the most important problems there is no single solution, and no solution can be achieved by technology alone. Cultures matter. Social relationships matter. Manufacturing, distribution, administration, and oversight can't be 100% tech driven and each step is essential.

We’re partial to Ai2 , but that doesn’t undermine some recent limitations Ai2 researchers have published:

• [PDF] MacGyver: Are Large Language Models Creative Problem Solvers? | Semantic Scholar (limitations in creative problem solving)
• https://www.semanticscholar.org/paper/b2ca176ed2607840574fcaa4ad4fb7203cb29c0d ? (bias in training data affective model performance)?
• [PDF] Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs | Semantic Scholar (implicit bias in models)
• [PDF] Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory | Semantic Scholar (revealing private information)
• [PDF] The Generative AI Paradox: "What It Can Create, It May Not Understand" | Semantic Scholar (models lacking understanding)

Conclusion

There is no AI race. It’s not the Manhattan Project. It's a team event with nonprofits, universities, governments, and for-profits from around the world interacting in complex ways to help solve important (and not-so-important) problems.?

So, can AI assist in all these problems? YES! And it will. But if the “AI race” remains about advancing capabilities and less about advancing usefulness, we’ll miss out on its potential. We don’t need AI that write a poem where every line ends with the letter x. We want AI that can make life more enjoyable/satisfying/meaningful for everyone. We’re not there yet, and the “AI race” narrative doesn’t help.

[1] The focus is almost always on GenAI, like GPT-4, Gemini, Anthropic, etc. This post uses AI and GenAI interchangeably for the purposes of this topic, but they're distinct in reality.

[2] we pick on Mollick a bit here mostly because his posts are a simple way to find a lot of diverse hype at a single source. For the purposes of this article he’s interchangeable with any number of other influencers. We don’t know him and we have nothing against him personally. While he definitely has a personal interest in making AI seem uber important because he has a book on the topic, we don’t pretend to know precisely why he has made any particular post or statement.

[3] Stopping our dive here. That’s only from his last two weeks on LinkedIn. He’s a prolific poster.