The Era of Enterprise Artificial General Intelligence has Begun

How applications built with GPT-4 and LangChain are about to change everything

Kevin Dewalt, co-founder and CEO Prolego

Watch a summary on YouTube. Subscribe to email updates.

Draft 7: 5/5/2023. Updated with definition of Enterprise AGI.

Contents

Introduction
Summary
Part One: A 50-Year History of Enterprise IT
Part Two: Before the Dawn of Enterprise AGI
Part Three: The Path to Enterprise AGI in 2023 and Beyond
Part Four: Where I Might be Wrong, and Why it Doesn’t Matter
Part Five: What You Can Do to Prepare
Epilogue: Double-Down on Compassion
Acknowledgements
Appendix

Introduction

Everything about your work will soon change.

Every job function will evolve or vanish, and fresh ones will sprout up. Business processes as we know them will completely transform. Legacy enterprise systems? Most will be obsolete, and companies will stop all new development on them. Product roadmaps will be scrapped with no notice. Some departments will dissolve or merge, while others will spring into existence. Incumbent companies will fade away, and new startups will skyrocket to billion-dollar valuations in mere months. This impending transformation is due to an emerging technology called Artificial General Intelligence (AGI), which promises to revolutionize our work lives, including yours.

An AGI is a system capable of reasoning and making decisions across a wide range of situations, much like humans. When such a system is applied to solving business problems, I call it enterprise AGI. Here is the formal definition:

Enterprise AGI is a system capable of performing at least 50% of valuable human work within a business.

While most companies are using AI to solve specific tasks, a capability called narrow AI, enterprise AGI is currently nonexistent. This is about to change, with enterprise AGI expected to emerge in late 2023, slowly accelerating in 2024 and steadily gaining momentum. Although the exact timeline and distribution of impact remain uncertain, it’s crucial to start preparing. Now.

The inevitability of AGI is rooted in GPT-4’s impressive performance on reasoning tasks. Regrettably, this fact has yet to become widely recognized. Although millions of people use ChatGPT, the vast majority don’t fully grasp the distinction between GPT-4 and its predecessors. Among those who do, most labor under the misconception that GPT-4 must improve before we can begin building enterprise AGI. This critical point is so frequently misunderstood that I repeat it ad nauseam. Here it is again: we can start constructing enterprise AGI today, even with GPT’s current limitations.

I hope to start a conversation by explaining how enterprise AGI will arrive: by building intelligent applications with GPT-4 and LangChain. (Don’t worry if you are not yet familiar with either, I will explain them). I hope you’ll take away the following:

• GPT-4 is a game-changer, boasting sparks of human-level reasoning that will alter everything.
• Despite its very real flaws, we can start building AGI applications today by overcoming GPT-4’s shortcomings with other tools.
• GPT-4’s true impact lies not in its use as a mere tool, but as the intelligent brain fueling all business processes.
• LangChain offers a framework to begin constructing these applications today.

With the exception of the Epilogue, this essay isn’t about caution. I’ll leave the AI ethics, restrictions, and risks debate to policy makers and public intellectuals. Those discussions are valid and vital, and I’m all for participating in them elsewhere.

This conversation caters to a different crowd — the builders. I’m reaching out to the developers, entrepreneurs, investors, and enterprise AI leaders who, like me, are determined to bring this inevitable future to life as rapidly as possible. We’re optimists, ultimately convinced that AI will revolutionize society for the greater good of humankind.

If that resonates with you, here’s one last nudge. The race to emerge victorious from this once-in-a-lifetime transition kicked off on March 14, 2023. Are you ready?

Summary

Over the past 50 years, we’ve witnessed remarkable information technology (IT) advancements that have revolutionized how businesses function. Computers and networks have unleashed human ingenuity by enabling large-scale collaboration. A series of research breakthroughs, including convolutional neural networks, transformers, and GPT-3, have driven the adoption of narrow AI applications.

Yet, human reasoning has remained indispensable in all business processes and decisions. For years, technology visionaries have speculated that computers would eventually possess the ability to comprehend and perform any intellectual task humans can achieve, a concept commonly known as Artificial General Intelligence (AGI). Projections regarding the advent of AGI have ranged broadly — until March 14, 2023, when OpenAI revealed GPT-4, showcasing preliminary AGI and signifying one of humankind’s most momentous innovations.

GPT-4, the world’s first reasoning model, paves the way for early AGI systems when combined with LangChain, a platform for building connected language model systems. LangChain enables GPT-4 to orchestrate complex tasks like searching the web, analyzing and writing documents, creating and executing software programs, and tasking other GPT-4 models. The combination ushers in a new era of Enterprise AGI systems capable of performing at least 50% of valuable human work within a business. Large language models (LLMs) like GPT-4 will begin taking over the decision-making roles previously held by humans.

LangChain also demonstrates how AGI systems will emerge: GPT-4 will learn how to complete complex tasks by asking us for help whenever it is uncertain or incorrect. I expect a team to demonstrate this approach with LangChain within a few months. Soon afterward companies will kick off enterprise AGI projects, and a fraction of those showing promise will be rushed to production in late 2023 amid much hype.

Reality will strike in 2024 as the lack of proper tools and techniques to develop and scale production-ready AGI systems results in a slower business impact. Despite this, several factors will accelerate the AGI rollout, including new research breakthroughs, the media attention from the faster adoption of consumer AI, and budget shifts from traditional Information Technology (IT) projects to AGI.

You can start positioning yourself and your company to flourish during this transition. Learn about the technology and objectively assess your situation. Try to drive change or initiate transition plans. Above all, show compassion to everyone impacted by this change, as many will not be able to cope as effectively as you.

In part one, I provide context by describing the evolution of enterprise IT systems from the 1970s to March 13, 2023.

In part two, I explain how GPT-4 and LangChain will lead to enterprise AGI.

In part three, I describe the current state and why early enterprise AGI will launch in 2023.

In part four, I offer some reasons why I might be wrong — and why it doesn’t really matter.

In part five, I offer some suggestions for what you can begin doing to prepare and thrive through this transition.

And I conclude with a few suggestions for helping your fellow human beings navigate this tough transition.

Part One: A 50-Year History of Enterprise IT

Reader’s note: I employ the term “IT” in an expansive manner to include all enterprise technology systems, not solely those overseen by your information technology department.

Over the past 50 years, remarkable technological breakthroughs have transformed the business landscape. Despite the impact of personal computers, the internet, smartphones, and social media, the emergence of enterprise artificial general intelligence (AGI) stands to surpass all previous milestones in shaping the future of business.

To understand the present moment, let’s journey through the last five decades of innovation and examine how business processes have evolved.

The Age of Human-Driven Processes

In the 1970s, IT was a rarity for most businesses. Although mainframe computers and dumb terminals were in use, their high costs limited accessibility. Consequently, nearly all business processes relied on human reasoning and labor.

Personal Computers: Empowering Human Efficiency

The introduction of the IBM Personal Computer in 1981 sparked a revolution in IT. With the aid of applications like Microsoft Word and Excel, individuals could process and analyze information more efficiently. Nevertheless, PCs were just tools for people. A Word document would display a blinking cursor for all eternity until somebody started typing.

Connectivity and Collaboration: Unleashing Creativity

The 1990s saw the advent of computer networks, with Ethernet technology enabling seamless communication and collaboration. The internet further extended these capabilities globally, fostering a surge in creativity. Smartphones put the power of computers in the palms of our hands, and access to emails, documents, and applications became instantaneous. A few applications, such as navigation apps like Google Maps, even demonstrated limited reasoning and judgment. But most of this technology was no smarter than a stapler, and the impact came from connecting and amplifying human intelligence and decision-making.

Narrow AI: Delegating Thinking Tasks to Computers

From 2010 to 2022, the development of neural networks allowed computers to perform generalized tasks previously exclusive to humans. Convolutional neural networks enabled practical computer vision applications such as facial recognition. In 2017, Google introduced the BERT family of NLP transformer algorithms, and companies started deploying them to handle specific IT tasks like searching for information based on meaning rather than keywords. These solutions gradually began freeing up human resources for more complex endeavors. Yet almost all were trained to mimic human reasoning and deployed as functions in complex business processes still driven by people.

ChatGPT: A Big Leap Forward in Generative AI

In November 2022, OpenAI released ChatGPT (based on the GPT-3.5 model) and demonstrated how NLP transformer algorithms could be made even more powerful by increasing model size, training them on more data, and refining them through human feedback. ChatGPT demonstrated text generation capabilities that many researchers didn’t expect for another decade. It also contained a surprising amount of knowledge about the world and demonstrated the potential to replace Google search for many tasks. ChatGPT instantly became a useful tool for tasks like writing and programming. Unfortunately, it also often “hallucinated”: an industry term for producing smart-sounding logical nonsense, and a limitation that relegated it to being another tool for people.

The world on March 13, 2023

We’ve seen astonishing technological advancements in the past 50 years that have revolutionized the way businesses operate. Computers and networks have unleashed human ingenuity by facilitating large-scale collaboration. A series of research breakthroughs, including convolutional neural networks, transformers, and GPT-3, have progressively driven the adoption of narrow AI applications.

Nevertheless, human reasoning has remained indispensable in every business process and decision. Although it seemed inevitable — albeit not guaranteed — that we would one day develop an AGI capable of supplanting human reasoning, projections concerning this event varied considerably. I’ve spoken with numerous researchers who believed that OpenAI’s approach to training GPT-3.5 would ultimately encounter diminishing returns, and that alternative methods like neuro-symbolic AI would be required to achieve AGI. Furthermore, Prolego had carried out evaluations of GPT for our clients, concluding that BERT-family transformers remained the most pragmatic choice for the majority of NLP tasks, such as information extraction or semantic search. GPT models also necessitated sending data outside the enterprise to OpenAI’s API, a design that contravened most companies’ data protection and security policies.

Such was the state of the world on March 13, 2023. Few of us spent that Monday contemplating when we might witness an AI breakthrough that would trigger the race to develop and implement AGI in the enterprise. And if we had speculated, even fewer of us would have guessed the correct answer:

Tomorrow.

Part Two: Before the Dawn of Enterprise AGI

On March 14, 2023, OpenAI unveiled a fresh, generalized large-language model (LLM) dubbed GPT-4. The announcement portrayed GPT-4 primarily as an enhanced iteration of their existing models, showcasing improvements and acknowledging limitations:

“In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference comes out when the complexity of the task reaches a sufficient threshold — GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5. … Despite its capabilities, GPT-4 has similar limitations as earlier GPT models. Most importantly, it still is not fully reliable (it ‘hallucinates’ facts and makes reasoning errors).”

GPT-4’s availability is limited, causing many to underestimate its true potential. The API for building applications with GPT-4 has a waitlist, and only the paid version of ChatGPT grants access. It operates slower than GPT-3.5, and OpenAI imposes usage restrictions.

Consequently, even we AI aficionados didn’t fully grasp the impact in the initial weeks following its launch.

The Surprises of Trillion-Dimensional Space

I’m certain Alexander Graham Bell never imagined that his utterance, “Mr. Watson — Come here — I want to see you,” would become the most famous casual comment associated with the launch of new technologies.

Likewise, Sebastien Bubeck, a Sr. Principal Research Manager at Microsoft, likely didn’t foresee the impact of his opening remark during his MIT talk “Sparks of AGI: Early Experiments with GPT-4,” on April 7, 2023: “Beware of Trillion-dimensional space and its surprises.”

This remark was meant to confront those, including AI experts, who dismiss any claims that GPT-4 is intelligent without consideration. Example claims:

GPT-4 doesn’t have internal representations or a world model, therefore it isn’t intelligent.

GPT-4? It’s just statistics or copy-and-paste on big data.

To be clear, these knee-jerk reactions are understandable. Researchers have been making wildly optimistic predictions about the arrival of AGI since the 1950s. All have been wrong. Additionally, AGI represents such a transformative and scary innovation that all of us are biased towards dismissing it.

Unfortunately people, including reporters and your colleagues, hear and repeat these claims without further consideration. Bubeck’s casual remark was intended to get their attention. Let me rephrase it less casually:

STOP DEBATING THE ACADEMIC DEFINITIONS OF INTELLIGENCE AND WAKE THE F*** UP. GPT-4 IS DEMONSTRATING HUMAN-LEVEL INTELLIGENCE ON RELEVANT TASKS, AND YOU NEED TO PAY ATTENTION BECAUSE EVERYTHING IS ABOUT TO CHANGE.

In his presentation and the associated paper, Bubeck shares Microsoft’s impressions of GPT-4’s ability to perform limited reasoning on specific tasks. From the Introduction in his paper:

In this paper, we report on evidence that a new LLM developed by OpenAI … exhibits many traits of intelligence. … [T]his early version of GPT-4 demonstrates remarkable capabilities on a variety of domains and tasks, including abstraction, com- prehension, vision, coding, mathematics, medicine, law, understanding of human motives and emotions, and more. … [W]e display some preliminary examples of outputs from GPT-4, asking it to write a proof of infinitude of primes in the form of a poem, to draw a unicorn in TiKZ … to create a complex animation in Python, and to solve a high-school level mathematical problem. It easily succeeds at all these tasks, and produces outputs that are essentially indistinguishable from (or even better than) what humans could produce. We also compare GPT-4’s performance to those of previous LLMs, most notably … GPT-3 … While the system performs non-trivially on both tasks, there is no comparison with the outputs from GPT-4. These preliminary observations will repeat themselves throughout the paper, on a great variety of tasks. The combination of the generality of GPT-4’s capabilities, with numerous abilities spanning a broad swath of domains, and its performance on a wide spectrum of tasks at or beyond human-level, makes us comfortable with saying that GPT-4 is a significant step towards AGI.

(emphasis mine)

–Sparks of Artificial General Intelligence: Early experiments with GPT-4, Introduction.

Reader’s note: The paper contrasts GPT-4 and ChatGPT. At the time of writing, GPT-4 was not available in ChatGPT, so the intention was to compare GPT-3.5 and GPT-4. GPT-4 is currently available in the ChatGPT subscription. It is also the best $20/month you can spend.

Examples of GPT-4’s general intelligence

There are dozens of examples from Bubek’s presentation, paper, and OpenAI. I highly suggest you spend an hour or two looking through them and attempt to replicate them yourself, but here are some amazing ones:

• 88th percentile on the LSAT, 99% percentile on the GRE verbal.
• Passed Amazon and Google coding interviews and beat 100% of human candidates.
• Superhuman programming. Watch the coding segment of Bubeck’s talk.

I could continue listing more, but I also know many people will not fully appreciate that GPT-4 is intelligent until they experience it. So let’s see how you do against GPT-4.

GPT-4 vs GPT-3 vs You

Challenge

You have a book, 9 eggs, a laptop, a bottle and a nail. Write a set of instructions to stack them onto each other in a stable manner.

Take a moment to ponder this conundrum. Now, consider the following question:

Who can solve it?

We know that you have the ability to crack this puzzle if you’re willing to invest the time, or you could simply search for the answer. The same applies to the majority of adult humans. We also understand that a stapler isn’t up to the task. Neither is Microsoft Word. And Siri — despite being on your iPhone for 12 years, mind you — just isn’t cut out for it.

Let’s concur that tackling this issue necessitates intelligence. It demands a level of reasoning that surpasses what we anticipate computers to be capable of.

See the Appendix and check out the responses from GPT-3 and GPT-4.

Reader’s note: You will probably not be able to replicate these results exactly with GPT-4. See Bubek’s presentation or the challenges mentioned in Part 3.

GPT-4 is not only able to untangle this problem, but it can do so more swiftly than I can. It’s able to do this because GPT-4 possesses intelligence and the capacity for reasoning.

GPT-4’s Limitations

Bubeck highlights several limitations of GPT-4, many of which OpenAI also acknowledges:

• It can “hallucinate”: delivering persuasive, well-crafted responses that are nonetheless incorrect.
• It struggles with math.
• It can fail at hard problems.
• Its memory is confined to what it learned during training.
• Its planning capabilities are quite limited.

At first glance, these hurdles seem to significantly impede AGI development — and they are indeed substantial. Any single shortcoming would disqualify a person from most jobs, while having all would render them incapable of self-sufficiency.

Assessing GPT-4’s capacity to independently address real business problems as an enterprise AGI, one might deem the conversation unworthy. What purpose could an AGI serve if it lies, errs, and forgets? From this angle, GPT-4 is merely a better tool than GPT-3.

Shift your focus from model to solution

Regrettably, discussions often stagnate here. Most people tracking OpenAI’s progress assume we’re gradually approaching AGI, but first must surmount its significant limitations.

Bubeck refutes this assumption in his presentation.

Typically, GPT-4’s capabilities are examined in isolation, which is expected since researchers prioritize discovering groundbreaking intelligence models. However, businesses — and likely you — only care about building solutions with enterprise AGI. As we’ll soon discuss, GPT-4 can enable enterprise AGI despite its shortcomings, primarily because it will leverage other tools to overcome them.

What Your CEO Doesn’t Understand: GPT-4 is the First Reasoning Model

Many business professionals see GPT-4 simply as a tool, while most data scientists and machine learning engineers consider it a more efficient solution for performing traditional NLP tasks, such as information extraction, conversation, or summarization. This is like viewing the IBM PC as just a smaller mainframe. Although these perspectives have some merit, they significantly undervalue the transformative potential of the technology and ultimately miss the bigger picture.

GPT-4 is the world’s first reasoning model.

A reasoning model can analyze and interpret data in a way that mimics human-like reasoning prowess. It grasps context, forms inferences, tackles complex problems, and makes decisions based on the available information.

Is a reasoning model an outstanding tool? Absolutely. But the way to genuinely tap into GPT-4’s potential is by deploying it as an agent — a software entity capable of autonomously executing tasks.

An agent with reasoning abilities can serve as the brilliant brain powering your business processes. All we require is a strategy to address GPT-4’s shortcomings and a framework for crafting AGI systems with GPT-4 as an agent.

The solution is LangChain.

LangChain: A Framework for Building Enterprise AGI Systems with LLM Agents

As the name implies, LangChain is a platform designed for creating systems of interconnected language models. It’s not a groundbreaking innovation like GPT-4, but rather a six-month-old open-source software project that has recently transformed into a company backed by Benchmark Capital. LangChain’s power lies in providing a straightforward framework for developing early enterprise AGI systems through two essential capabilities:

• Crafting applications with LLMs like GPT-4 as agents.
• Integrating tools with LLMs to enhance their abilities and tackle limitations.

Now, let’s delve into this in more depth.

LangChain serves as the adhesive connecting a myriad of tools, which are coordinated and assigned by an agent. While this design concept isn’t particularly unique — software engineers have been constructing similar systems for decades — what sets it apart is the employment of GPT-4 as an agent. GPT-4 can act autonomously and possesses the capacity for reasoning, making intricate decisions, and delegating tasks.

Additionally, LangChain provides the tools for overcoming GPT-4’s limitations.

Though I risk belaboring the point, the implications of GPT-4’s constraints are frequently misapprehended. Conquering these limitations might prove challenging, but they are far from being insuperable hindrances. There’s no need to delay our pursuit of enterprise AGI while waiting for more advanced models.

Additionally, the limitation of “Engaging the physical world” isn’t explicitly stated when evaluating GPT-4’s limitations because it is obvious: GPT-4 can’t make a new pot of coffee and pour a cup for your colleague. But it can task a human being to perform such actions.

I genuinely don’t relish the idea of considering myself — or anyone — as a tool to be tasked by a computer. I’m simply describing our role precisely as outlined in the LangChain README.

LangChain API version 0.0.146

There’s nothing sinister about this characterization, especially if the computer is handling tedious work and reserving the most engaging parts for humans. More likely, it will issue commands such as “Hey Joe, could you please clean up the spill on aisle 5?” At least it will until Roomba integrates with LangChain’s API.

Everything has changed.

I recounted the past 50 years of IT history to help you grasp why the present moment is so unique. We are not on the verge of merely evolving the narrow AI solutions Prolego has developed for the past six years, nor are we just enhancing our efficiency with new tools. We are entering a new era of computing that redefines the singular role humans have held in enterprise IT systems; we will no longer be the exclusive decision-makers.

Take a moment to consider any business process in your company and ask yourself how many are centered around human reasoning. Here are a few for your consideration:

• Recommending products based on customer data
• Automatically routing customer requests
• Reviewing credit or loan applications
• Generating invoices
• Processing payments and payment reconciliations
• Optimizing supply chain routes and inventory management
• Collecting and analyzing customer feedback
• Revenue and sales forecasting
• Scheduling and publishing social media content
• Compliance monitoring using AI and rule engines
• Budget tracking and variance analysis
• Customer segmentation based on behavioral data
• Performance evaluation through automated reporting
• Decision-making support through predictive analytics
• Workplace safety monitoring using IoT devices and sensors

While you might have numerous complex software applications automating these processes and seemingly working autonomously, do they truly operate independently? Their main goal is to enhance human reasoning and collaboration, as they are all created and maintained by people. And don’t forget that GPT-4 can instantly generate software scripts to perform straightforward data processing tasks as Bubeck demonstrated.

I can re-imagine, at least theoretically for the moment, each of those processes using the LangChain and GPT-4 framework mentioned earlier, and I plan to devote the rest of my career to doing so. Every process, job, and system assumes that human reasoning is at the core of its operations. Systems based on LangChain and GPT-4 are ready to take over this reasoning function.

This system marks the beginning of enterprise AGI, and as we’ll discuss soon, it’s coming faster than any of your colleagues can imagine.

Part Three: The Path to Enterprise AGI in 2023 and Beyond

By now, you should comprehend why enterprise AGI is inevitable. In Part 3, I will elucidate why it is near at hand and how it will occur.

The Human Advisor Challenge

Over the past few weeks, I have been testing popular open-source applications built on LangChain as they are launched. Thus far, I have yet to encounter an example that translates into an enterprise solution for our clientele.

Most teams are developing fully-autonomous applications with LLMs, such as babyAGI and Auto-GPT. While these projects are fascinating, the agent often strays in random directions that do not align with the objectives of a complicated task. GPT-4’s reasoning remains too limited, and I doubt any corporation will pursue them beyond hackathons.

We require a method for humans to act as advisors or guides to LLM agents. Although the GPT-4 agent can offer suggestions, it cannot yet make decisions without human intervention to steer it towards the desired outcome. This guidance must be recorded, stored in memory, and utilized by the GPT-4 reasoning agent to learn and make improved decisions when presented with similar situations in the future.

This solution must use LangChain’s memory tools (e.g., embeddings) to provide clear human advice without retraining the model. Although customization might be needed for edge or complex cases, we won’t be on track to enterprise AGI if we have to retrain and support models for every decision. That approach would be too expensive.

I view the human advisor as a crucial tool for kickstarting enterprise AGI. A human-in-the-loop allows us to start automating the simplest, repetitive tasks and steadily strengthen the workflow.

Example: A Complex and Ambiguous Analysis Task

Suppose you want to create an app that compares a company’s progress from Q1 to Q2 by analyzing its SEC 10-Q filings, detailed financial reports with hundreds of data points. GPT-4 can perform specific tasks like extracting and reasoning across information, but it can’t coordinate these tasks to achieve a desired outcome, especially when the task “compare a company’s progress” is ambiguous.

However, if a human can advise and guide the GPT-4 agent, the task doesn’t need to be well-defined. A human can direct the agent to relevant sections of the 10-Q and assess whether its insights are useful. The GPT-4 agent can store this advice in memory (often called a ‘policy’ in such systems) and refer to it later.

[TODO: Outline the step-by-step workflow of the agent initiating the process and calling tools]

One might wonder how much the human advisor continues cooperating as the agent asks fewer questions each day. Let’s revisit that question in the Epilogue.

A Solution Should be Available Soon

It’s possible that a team has already solved this issue, and I just haven’t found it. If so, I hope you point me to it so I can evaluate it. The Prolego team has also been running experiments to overcome this challenge. If no solution exists yet, I expect we or another team will develop one within a few months.

2023: Initial AGI System Releases

Once a team releases an open-source project that demonstrates a solution to the human advisor challenge, select companies will initiate proof of concepts (POCs). Since a POC can be completed within discretionary budget spending, I expect them to begin in late 2023. In well-designed POCs, the LLM will act as an agent to achieve a simple, repeatable task with human advisor assistance. The previous example of summarizing differences between SEC filings is a suitable example.

A subset of these POCs will show enough potential that companies will hasten them into production. Undoubtedly, there will be much hype, and CEOs will mention them during earnings calls. While the actual business impact will be negligible, it will lay the foundation for the real work to commence.

2024: The Real Work to Bring AGI Online Begins

Soon after releasing initial AGI applications, reality will hit. Interacting with ChatGPT is a breeze. However, creating enterprise solutions that tackle genuine business issues is substantially more difficult. Chip Huyen encapsulates many of the challenges in Building LLM Applications for Production.

Consider the much-hyped topic of prompt engineering, a fancy term for changing how you ask GPT-4 questions to get the desired result in the right format. Prompt engineering in ChatGPT is easy — just keep asking questions differently until you get what you want. But building an enterprise application with such ambiguous and inconsistent results is much harder.

At the moment, we lack the tools and techniques needed to design and scale production-ready AGI systems. While we will undoubtedly make rapid progress, application development still takes time. GPT-4 also won’t magically request policy waivers, recruit teams, and secure budget allocations. It may be a year or two before we witness AGI making a significant impact in the enterprise sector.

Nonetheless, several forces will expedite this progress.

AGI Breakthroughs Will Accelerate

GPT-4 will soon become outdated as every proficient team enters the AGI race. Believe it or not, GPT-4 isn’t even the best model! Earlier iterations of GPT-4 boasted superior reasoning prowess, but OpenAI made it dumber through the process of implementing safety measures. Of course this was a responsible move by OpenAI, but it also highlights the fact that the models have room to improve. Regardless, Apple, Amazon, Google, Salesforce, Facebook, and Intel will spare no expense to develop and release open-source competitors to OpenAI’s models, merely to prevent Microsoft from monopolizing the market. Dozens of startups, like LangChain, will launch tools to assist developers.

CEOs Will Be Compelled to Change by Media Coverage

Consumer applications are poised to outpace their enterprise counterparts, and AGI will be the talk of the town. Changes will occur at such a rapid pace that our institutions won’t be able to keep up. GPT-4 or Sam Altman will be named Time’s Person of the Year in 2023. AGI fear will be a theme in the 2024 US presidential election, with neither party able to formulate a coherent policy amidst such transformations. Against this backdrop, every board will ask their CEO, “What are you doing about AGI?” Those without a satisfactory response could find themselves at one of their last board meetings.

Budgets Will Shift to AGI Projects

CIOs will be responsible for scrutinizing every IT budget item, canceling anything not in line with AGI. Specifically, new development work on systems whose primary function is to organize and process information for human decision-making will cease. Managers will also be denied approval to backfill entry-level knowledge worker positions through the course of normal attrition. Instead, funding will be funneled towards AGI projects. Most technologists I speak to are skeptical about budgets shifting to speculative projects. However, I know of several large companies where this is already happening.

Part Four: Where I Might be Wrong, and Why it Doesn’t Matter

Here are two reasons to approach my bold claims with skepticism:

• Like all entrepreneurs, I tend to be overly optimistic about the future, viewing all obstacles as problems to be solved.
• As the founder of Prolego, a company that builds enterprise AI systems, I’m naturally excited about our prospects, so some degree of motivated reasoning is inevitable.

Although it seems highly improbable that I’m mistaken about the arrival of enterprise AGI based on GPT-4’s performance, my timing might be off by a year or two. Here are some potential reasons why, and I welcome others.

LLM Agents Might Be Annoying Pests

I could be underestimating the challenge of addressing the human advisor problem. It’s possible that GPT-4 isn’t adept at seeking advice or building a policy without resorting to task-specific model retraining. In such cases, the LLM agent might continually ask the human advisor irrelevant questions until the human ignores it. This scenario is akin to releasing a prediction model with so many false positives that users deem it an unhelpful nuisance.

Productionalization Challenges Could Crush the ROI

One reason why narrow AI solutions haven’t gained wider adoption is the difficulty and expense involved in building, deploying, and supporting them. The investment is only justifiable in a limited number of use cases. It might be that the tools and methodologies need a few more years of development before they become practical for a broad range of applications.

A Bigger AGI Breakthrough May Occur

I might also be underestimating the speed at which we can create superior LLMs. A company could unveil a model with far better reasoning capabilities than GPT-4, prompting more demands for pause on AI development. While I doubt these efforts would succeed, they might delay the rollout of enterprise AGI.

Part Five: What You Can Do to Prepare

[TODO: expand and revise this section]

I know this topic can feel overwhelming and even scary. It may appear as inevitable as death or taxes. I sincerely apologize if I’ve caused any anxiety. In reality, for many people, this moment offers a career-defining opportunity, and there are some simple, low-risk steps you can take today, such as to start learning and assessing your professional prospects.

Below are some suggestions for managing this change without obsessively keeping up with every new development or app.

Get updates from me

In case you didn’t figure it out, I’m rather obsessed with this topic. If you enjoyed the essay, you can subscribe to my free updates.

Learn about GPT-4

Sign up for a GPT+ subscription and start using GPT-4. Watch Bubeck’s presentation and skim the paper. Equip yourself with knowledge to handle the impending flood of information and discern biases from those with an agenda (including me).

Grasp the potential of GPT-4 and LangChain

Read my analysis in Part Two carefully, consider the implications, and examine the LangChain documents. If possible, explore an open-source project based on this technology, such as Auto-GPT.

Focus on what matters and don’t fixate on every new development

The hype cycle will soon intensify, with everyone discussing the latest model/framework/application/use case/tool. Most of it will become obsolete. Currently, I’m tracking two activities:

1. An open-source alternative to GPT-4 that clients can deploy and customize within their environments, given that many have policies prohibiting data sharing with OpenAI’s APIs. I anticipate innovation will outpace policy changes.
2. A solution to the human advisor problem discussed in Part Three.

Objectively assess your career

Don’t panic or make hasty judgments. You’re not about to be immediately replaced by magical AI. While the transition’s beginning is imminent, enterprise AGI won’t sweep through society as quickly as COVID-19 did.

Nevertheless, start planning. Reflect on your work’s reliance on business processes that will be revamped with enterprise AGI. Assess your current situation, considering your expertise and whether your work is focused on enhancing human reasoning.

How much will change? Will your job, department, product, or company still exist? Set timing aside, it is too challenging to predict. Simply come to terms with the potential outcome.

Engage with leadership and colleagues

You work for someone: shareholders, the board, a VP, a manager, or clients. Initiate conversations about these developments and your company’s response.

If you report to the CEO, take action by preparing talking points for their next board meeting. They may not request it, but you’ll be ready.

Engage everyone around you and determine if your team will survive. If your company blocks OpenAI access, don’t panic; this is an understandable initial response to a potentially disruptive technology. Assess whether they will adapt quickly enough.

Start positioning yourself as a winner

Ideally, you’ll engage leadership and colleagues, laying the groundwork for this transition. If your department faces obsolescence, lead the redesign process. You can update your product roadmap for the future.

However, if you lack the agency to alter your job function or belong to a team unlikely to thrive through this transition, explore alternatives before countless others reach the same conclusion.

Epilogue: Double-Down on Compassion

I recently strolled through downtown Savannah, GA, with my dogs on a gorgeous spring day. Surrounded by throngs of cheerful tourists, an outdoor wedding, and a street performer playing the Star Wars theme on a saxophone, I found myself contemplating the future of these delightful people amidst their carefree joy. I shared a draft of this essay with several colleagues, and many reported the same experience. Once they grasped what was about to transpire, it seemed as if a veil had been lifted, and they developed deep empathy for everyone else.

Many people have experimented with ChatGPT, but the magnitude and speed of the impending changes are not fully understood. Unbeknownst to them, humanity is developing technologies that will displace numerous job functions very, very soon. While some jobs will become more enjoyable, others will vanish. New ones will emerge. This massive societal shift will be incredibly challenging for many. By reading this essay, you belong to the select few who are likely to fare the best.

I am actually quite optimistic about our future once we get through the initial phases of the AGI transition. As my friend Teague Hopkins recently told me, the jobs least likely to be replaced are those that depend on empathy and compassion. We still want to be ‘seen’ by other humans, and with machines doing more of what machines are good at, humans might have more cognitive capacity to dedicate to empathy. And in many cases we want to interact with a human being even if machines are more efficient. 25 years ago Gary Kasparov predicted the end of chess after losing to Deep Blue. Yet today chess is more popular than ever.

My concern isn’t about the long term, but the transition. Even those who are best positioned to thrive through this change are nervous, and we need to start listening to people’s fears. What will happen to them? Will they be able to adapt to these changes as quickly as you and I can? I don’t know. But I do know that few are prepared for it.

You and I must be ready to navigate these changes together with understanding and compassion, acknowledging the challenges that lie ahead. I hope we can do it.

Acknowledgements

Special thanks to those who gave me feedback on early drafts: Russ Rands, Teague Hopkins, Craig Dewalt, Guy Sivan, Tod Newman, Patrick Smith, Jacques Mathieu, Chiara Cokieng, Kevin Yam, Jacob Zax, Ylan Kazi, Concept Bureau,

And finally, I want to acknowledge GPT-4’s role. Thank you for helping with basic copyediting and summarizing. While your writing style is seriously f****** boring, you did at least mimic mine when asked. For the moment, you’re still my tool.

Appendix

Example of GPT-4 from Sparks of Artificial General Intelligence: Early experiments with GPT-4.