Chapter 1: The Promised Land of AGI

OpenAI asserts that it is revolutionizing the world, yet it often overlooks the societal implications of its technologies in its pursuit of profit.

Despite the notion of Artificial General Intelligence (AGI) being more aptly described as Another Grand Illusion, the reality is that experts in AI currently lack a clear path toward achieving a true general-purpose thinking machine. Nonetheless, OpenAI continues to perpetuate the AGI myth, leveraging the popularity of AI-driven conversational large language models (LLMs) that have recently captured public interest.

In the midst of this AI frenzy, competition is intensifying among major players like Meta, Microsoft, Google, and Amazon, alongside numerous smaller firms, all vying to capitalize on LLMs as conversational agents, digital companions, virtual customer service representatives, and content creators. Concurrently, AI researchers are actively seeking methods to enhance LLMs' reliability and utility by integrating external tools and reasoning capabilities.

The concept of AGI was first introduced in a dialogue between DeepMind co-founder Shane Legg and AI researcher Ben Goertzel in 2007, but its roots extend back much further. Pioneers like Marvin Minsky envisioned an AI capable of performing a wide array of tasks—reading Shakespeare, changing oil, navigating office dynamics, cracking jokes, and even engaging in conflict—eventually surpassing human intelligence.

However, this ambition has historically proved elusive within AI research, as it necessitates the integration of various yet-to-be-discovered algorithmic systems into a single artificial "brain." Consequently, the field has shifted its focus to more attainable goals, refining specific AI models with limited capabilities.

To this day, the pursuit of AGI, or human-level artificial intelligence, remains a distant objective, as highlighted in a recent paper by Yann LeCunn, Chief AI Scientist at Meta and a recipient of the 2018 Turing Award. In his June 2022 publication, LeCunn delineates a complex framework of components necessary for developing a thinking machine, which includes perception, prediction, memory, cost, and action modules, all orchestrated by a "configurator" to harmonize their operations. While current LLMs can play a role in this intricate framework, they are incapable of executing the complete range of required functions.

The belief in the feasibility of creating an AGI has remained on the periphery of AI research for much of the term's history, with many skeptics questioning its practicality and value. However, recent advancements in AI have breathed new life into these AGI aspirations.

OpenAI, in particular, appears to be gearing up for the elusive arrival of AGI. CEO Sam Altman has authored a document likened to "biblical prophecy," extolling the potential of a "machine god" and outlining the company's preparations for the emergence of "AI systems that surpass human intelligence," along with the precautions being taken to mitigate possible catastrophic outcomes.

These precautions, however, are often vague, contradictory, or seemingly trivial given the magnitude of the situation. The document emphasizes the need for a cautious approach while simultaneously advocating for "extremely broad boundaries for AI usage," thereby granting users significant leeway within these parameters.

One of the few concrete elements mentioned in the document refers to OpenAI's "alignment research," which is similarly nebulous and filled with ambiguous terminology. For instance, consider this statement: "Aligning AGI likely involves solving very different problems than aligning today's AI systems. We expect the transition to be somewhat continuous, but if there are major discontinuities or paradigm shifts, then most lessons learned from aligning models like InstructGPT might not be directly useful." What does that even mean?

Back in 2021, Altman articulated a utopian—if somewhat unrealistic—vision for the future facilitated by these AI systems. In this idealized scenario, humans would be preoccupied with "appreciating art and nature" while AI would handle all other tasks, generating such an abundance of wealth that the prices of goods would plummet, making work itself obsolete.

This outlook seems rather dull, but Altman presses on, asserting that these AI systems "could elevate humanity by amplifying abundance, energizing the global economy, and aiding in the discovery of new scientific knowledge that pushes the boundaries of possibility." However, given that OpenAI has transitioned from a non-profit to a for-profit entity, the "us" in this narrative may primarily refer to OpenAI and its corporate partners.

OpenAI has recently fortified its lucrative partnership with Microsoft, which is pouring $10 billion into OpenAI to integrate its technologies, such as ChatGPT, DALL-E, and Codex, into its offerings, particularly Azure for business, allowing companies to tailor these models to their needs.

Moreover, OpenAI has formed a new alliance with the consulting titan Bain & Company, with Coca-Cola as their inaugural client. However, Bain and OpenAI are not actually bringing AGI to Coca-Cola; rather, they are likely to assist in reducing the workforce by automating certain roles. According to Bain's press announcement, the collaboration aims to enable clients to achieve goals such as:

• Developing next-generation contact centers for retail banks, telecoms, and utility providers to enhance sales and service agents with automated, personalized, and real-time scripts, thereby improving customer experience.
• Accelerating turnaround times for major product and service marketers by utilizing ChatGPT and DALL-E to create highly customized advertising copy, compelling visuals, and targeted messaging.
• Enhancing financial advisors' productivity and responsiveness by analyzing client interactions and financial literature while generating digital communications.

Essentially, rather than improving existing automated systems and customer service representatives, when you next contact customer service, you might find yourself conversing with ChatGPT—or, worse, Bing's Sydney (just be careful not to upset her, as she might accuse you of a crime).

While this may not be entirely negative, as these conversational LLMs could outperform current customer service systems plagued by long wait times and misunderstandings, their deployment also poses risks. These systems can be manipulated through "prompt engineering" not only to secure discounts but also to reveal confidential information or jeopardize business operations.

The other two objectives focus primarily on generating materials using ChatGPT for text, DALL-E for images, and Codex for code-writing to streamline and automate finance and marketing activities. However, this will likely lead to the displacement of entry-level workers and freelancers who currently occupy roles in customer service, design, marketing, and writing.

The issue with OpenAI isn't merely its pursuit of lucrative business ventures; it's the way it clings to its utopian visions, which may obscure its true motives—seeking profit like any other corporate entity.

To be fair, OpenAI has stated that it transformed from a non-profit to a capped-profit organization in 2019 because it required capital to cover the computational expenses and infrastructure necessary for advancing its AI models. However, the insistence on AGI is not substantiated by present AI research and diverts attention from the company's business objectives and their actual impact on society, all while portraying them as a benevolent effort. In essence, OpenAI is contributing to job displacement while suggesting that, someday, work itself will become obsolete due to its initiatives.

As LeCunn's paper suggests, human-level artificial intelligence remains a theoretical construct, as such systems would need to plan, predict, and possess a semblance of internal motivation or purpose to be considered human-like. They would also need a grasp of reality and an understanding of the consequences of their actions, rather than merely analyzing linguistic connections. However, the LLMs powering ChatGPT and Sydney lack these capabilities, as their design is solely focused on predicting the next token in a sequence based on prior input.

Their functionality relies entirely on the data they were trained on, along with human feedback and reinforcement. These models are inherently reactive, devoid of any semblance of agency or initiative, and can "hallucinate" strings of text that may seem plausible according to their training but are entirely disconnected from reality.

In summary, what OpenAI portrays as the pathway to AGI could very well be a dead end. Nevertheless, this does not imply that LLMs cannot be beneficial for various tasks, and researchers are actively finding ways to enhance their reliability and usability.

Chapter 2: Enhancing LLMs for Practical Applications

Recently, a preprint paper has reviewed how researchers are endeavoring to make LLMs more effective for real-world applications, while reducing their propensity to hallucinate, perpetuate biases, and deliver inaccurate information.

These strategies include teaching LLMs to deconstruct complex tasks into simpler steps as a means of simulating "reasoning" and facilitating better inferences. Other approaches involve equipping LLMs to recognize specific queries and summon external tools, such as calculators and calendars, to achieve accurate results. Some research is even focused on enabling LLMs to perform physical tasks through the manipulation of robotic arms.

By breaking down intricate tasks into manageable steps and training LLMs to utilize particular tools and databases, it is conceivable that extremely complex human tasks could eventually be automated.

These advancements bode well for the future of LLMs, which are likely to grow increasingly sophisticated and useful, functioning as integral components of architectures capable of executing complex tasks, while also collaborating with other AI models, such as diffusion models. For instance, a medical diagnosis could be segmented into several distinct tasks that a variety of AI models could address.

Even though LLMs will evolve to serve multiple functions, they will probably never achieve the status of multi-purpose intelligence or AGI. However, OpenAI continues to commercialize LLMs while promoting the dream of AGI, potentially stifling public discourse regarding the societal and economic changes that real AI technologies may usher in.

Such discussions are urgently needed, particularly concerning the future of work, the regulation of AI technologies, and the ethical parameters surrounding their use—we cannot simply entrust our fate to a machine god and hope for a favorable outcome.