by In late 2022, OpenAI released a Transformer-based Large Language Model (LLM) called “ChatGPT.” Against the expectations of the OpenAI team, ChatGPT became the fastest growing web-based app in history, rising to 100 million active users in two months (beaten only by Meta’s Threads). The first public impressions of ChatGPT had a relaxed quality and fake badges. In February 2023, Henry Kissinger, Eric Schmidt, and Daniel Huttenlocher wrote that artificial genetic intelligence (AI) is comparable to the intellectual revolution initiated by the printing press, this time condensing and ‘distilling’ the storehouse of human knowledge. In March 2023, Eliezer Yudkowsky, predicting annihilation-level risks, urged the governments and militaries of the world to shut down the AI project and “be prepared to destroy a rogue data center by airstrike.”
These first impressions represent two ends of the spectrum, but the reasoning that lies in the space between them is common in technology policy analysis: personal impressions of AI flow from the generation of the background assumptions from which policy analyzes are made. When assumptions of fundamental importance are unquestioned, it is all too easy to fall into the trap of extrapolating from current technological conditions to future technological marvels. Technology policy analysts of all stripes do great work, but it’s time to recognize the gaps in our reasoning and reach for our higher purpose individually and collectively.
An example shows the general tendency. The Center for New American Security Paul Scharre, in his book “Four Battlegrounds” — which is overall a fund of insights — hedges on the future of AI, although he continues towards the idea that “Databases can be built more and more diverse. resulting in more robust models. Multimodal datasets may help create models that can associate concepts represented in multiple formats, such as text, images, video and audio.” This prospect follows the idea that augmenting AI systems will lead to new capabilities (increasing their internal capabilities and training datasets) while positively referencing Richard Sutton’s famous argument in “The Bitter Lesson” about the benefits related to such techniques.
Shortly thereafter, Microsoft researchers helped set the tone for a flurry of overly optimistic claims about the future of LLMs with their provocatively titled paper, “Sparks of Artificial General Intelligence” on GPT-4. It’s not hard to see how a personal understanding of GPT-4 could lead to “We’re on to something big here.” However, this is no reason to allow the assumptions attached in this view to be collected in your analyses.
Extensive research highlights the limitations of LLMs and other Transformer-based systems. Hallucinations (authoritative but factually incorrect statements) continue to plague LLMs, and some researchers are suggesting that these are simply inherent aspects of this technology. According to one recent study, voters who use chatbots to get basic information about the 2024 elections can be misinformed about illusory polling places and other false or outdated information. Other research shows that LLM’s ability to form abstractions and generalize them is beyond that of humans; the reasoning capabilities of multimodal systems is a similar story. While OpenAI’s latest development—the “Sora” text-to-video generator—is impressive in its realism, it conjures things and people out of thin air and doesn’t conform to real-world physics.
So much for the idea that new methods like image and video will lead to the reliable, robust and interpretable AI systems we need.
None of this suggests that it is only hype in the world of technology. Carnegie’s Matt O’Shaughnessy rightly notes that talk of “awareness” is likely to have a negative impact on policymaking. as fundamental limitations of machine learning. Additionally, the Biden administration’s sweeping October 2023 executive order on AI, while largely invoking the Defense Production Act to authorize monitoring of certain computer-powered AI systems, was more varied in tone than expected. expecting it.
But, the problem we identify here is not a hype problem per se. There is hype result getting stuck in analytical frames that are all too easy to ignore in favor of quick publications and individual or organizational self-promotion. Lest we believe this is just a strange LLM-specific bias, the dismay of AI-enabled and autonomous drones on the battlefield in Ukraine should raise an eyebrow at the alleged speed of the underlying attacks happened in 2023. In addition, it is easier to find nuance. in the field of quantum information science, but at the same time, little individual or collective thought seems to arise when its crown jewel of quantum computing begins to look to downgrade its future.
However, today’s generational AI is starting to look like a parody of Mao’s Continuous Revolution – transforming this technology into human-like “general” intelligence or some other marvel of technological imagination is always a model upgrade, and it cannot be allowed. to comply with challenges from regulatory bodies or popular movements.
The important thing is that policy analysts make choices when assessing technology. Choosing certain assumptions over others presents the analyst with a set of possible policy options at the expense of others. Individuals will inevitably have first impressions of new technologies and this can be a source of diversity of opinion. The problem of policy analysis arises when practitioners fail to pour their first (or second, or third, etc.) opinions into a shared crucible that exposes unstable ideas to high-temperature intellectual criticism, leading them to policy challenges express specific and. solutions without unduly neglecting other wholesale possibilities.
Policy analysis is generally a combination of ingredients from industry, domestic politics and international affairs. It is merely recognizing that there is a policy challenge de novo but from an intuitive connection between the needs and values of a society and the anticipated or actual impacts of developments within its borders or abroad. That intelligence—we all have it—should be the focus of our honest, shared scrutiny.
Vincent J. Carchidi is a Non-Resident Scholar at the Middle East Institute’s Strategic Technologies and Cyber Security Program. He is also a member of the Foreign Policy Cohort for America’s NextGen 2024 Initiative.
For the latest news, weather, sports, and video streaming, go to The Hill.
