HR AI Article 1: The Crown Jewels

Just over a year ago, Greg Brockman of OpenAI introduced ChatGPT-4 via live stream, heralding the start of a groundbreaking era in technology. It became the fastest growing consumer application on the internet…. ever. Within just two months of its launch, it garnered an astounding 100 million monthly active users. To put it in perspective:

• Instagram took two and a half years to reach 100 million users.
• TikTok achieved the same milestone in nine months.
• ChatGPT accomplished it in a mere two months.

In this article we’re going to focus on the application of Gen AI, not from the consumer perspective but from the perspective of the corporate enterprise. For this article, let's assume this new wave of disruptive technology will follow a similar path to that of previous waves that transformed the corporate world, such as client-server, internet, mobile, and cloud computing. Like these past waves, it's expected to evolve through various stages, ultimately reaching the "Plateau of Productivity." At this stage, the technology matures, and its true business value becomes fully realized.

Traditionally, companies have approached disruptive tech cycles in two ways:

1) Some adopt the “fail early and fail fast” philosophy, but the high stakes and potential safety risks of this particular cycle make it a less viable option for many.

2) Others strive to identify and focus on the primary attribute that sets the new technology apart, thus reducing the risk of getting trapped in the cycle’s highs and lows. Avoiding the trap of exploring every theoretical application of the technology—many of which may overlap with existing solutions.

This article is for those following the second approach.

What then, is Gen AI's uniquely valuable primary attribute? Given the dominant use cases surfacing across the corporate world in the past year, it's easy to assume it centers around Gen AI's proficiency in summarizing vast amounts of text. Many companies are integrating Gen AI with their knowledge repositories—extensive libraries of text where Gen AI can swiftly locate the most pertinent information, summarize it, and relay it back. These use-cases look like enhanced Helpdesk Chatbots that can fetch and summerize policy documentation, or “easier to build” reports that summarize dense textual materials like large-scale survey results.

Sidebar Recommendation: Refrain from integrating personally identifiable information with Gen AI, unless you possess comprehensive expertise in the technology's inner workings. Even with such knowledge, proceed only with robust support and collaboration from your corporate legal/compliance and IT data security teams.

This search and summarization trend isn't unexpected; historically, when groundbreaking technology emerges, we often gauge its worth by comparing it to familiar tools. For instance, early internet search engines were likened to "the world's largest Yellow Pages," and the power of engines were measured in "horsepower." We're seeing a similar pattern with Gen AI, where its prompt-based interface leads many to liken it to a search engine. However, Gen AI's capabilities extend far beyond text search and summarization.

The real distinction of Gen AI lies not merely in text summarization but in its unparalleled capacity for continuous learning and adaptation through interaction and feedback, thanks to its neural network foundation. My view is that by concentrating on something we’ll call "Reward Signal Capture (RSC)," which enables Gen AI’s ability to learn, businesses can sidestep the time lost at the cycle's peak and trough, unlocking immense business potential. RSC is crucial; the enterprises that master RSC will undoubtedly lead the pack.

If data is the gold of the digital age, then reward signals are its Crown Jewels.

We'll delve deeply into what exactly reward signals are, but for now, think of them as the feedback that sharpens AI's abilities, like a navigator refining a ship's course. These signals are vital for unlocking AI's full potential, enabling it to adapt and accrue significantly in value.

In the past year, the corporate world has seen the emergence of a handful of dominating "implementation patterns" for Gen AI. However, not all are equally adept at capturing the reward signals generated within a company.

How well does your approach to Gen AI leverage Reward Signal Capture (RSC) to unlock this technology's unparalleled business value?

Neglecting this could inadvertently hand over a competitive advantage to your rivals. Through the lens of RSC, we'll delve into current Gen AI implementation patterns. While our examples mainly target the HR domain (my home turf), the principles easily apply across various business domains.

This article is intended for non-IT business professionals seeking to maximize Gen AI's value and IT professionals seeking to bridge the gap between technological and business challenges.

Before we dive in, I want to express my gratitude to my incredible colleagues, friends, and mentors who have shared their experiences and insights on this extraordinary technology.

Understanding Reward Signals: The Path to Learning

Reward signals serve as crucial feedback, guiding AI systems like LLMs to refine their learning process, much in the same way that humans benefit from both praise and constructive feedback. Think of an LLM as a student attempting to solve a problem. In the case of the LLM it generates what it believes are the most likely answers then, using mathematics, evaluates the probability for each answer and ranks them. While the LLM is being trained, this evaluation involves calculating what's known as a "loss function" value, which measures how far off the answer it selected is from the actual correct answer presented in its training data.

If the LLM's predicted answer is quite close to the correct answer, resulting in a low loss function value, it receives a positive signal, akin to a "well done." This indicates that its guess is accurate and it's on the right track. Conversely, if the predicted answer is far from the correct answer, indicated by a high loss function value, the LLM receives a signal that it needs to change its approach for next time. This feedback mechanism, rewarding low loss values (accurate answers) and correcting high loss values (inaccurate answers), encourages the LLM to continuously improve its accuracy over time by adjusting the probabilities and rankings it used.

These signals arise not only from the accuracy of an AI's predictions compared to a training dataset but also from user interactions, through a process called “Reinforcement Learning From Human Feedback” (RLHF). Positive responses (like a “thumbs up”) reinforce desirable behaviors, while criticisms (a “thumbs down”) or direct edits and corrections signal the AI to refine its approach. For instance, feedback pointing out missing essential skills in an AI-generated job description prompts the AI to emphasize those skills in future outputs.

Due to technical and cost constraints, these cycles of feedback and adjustment are currently processed in training batches, rather than in real-time.

Regardless of how often they are applied, Reward Signals remain crucial for evolving AI's capabilities, effectively transforming it from a static repository of data into a dynamic value-generator that grows increasingly intelligent over time.

Reward Signals – A deeper dive with practical examples:

The following illustration shows how reward signals are generated and used. An input query like "What is the capital of the USA?" is encoded (i.e. translated into a mathematical representation called a “high dimensional vector”) by an LLM, which then predicts an answer. If the prediction—such as "Washington D.C."—matches the training data, the LLM receives a positive reward signal, affirming that the prediction should be reinforced

Reward signals also arise from RLHF through user interactions. For example, if an AI's correct prediction is validated by a user's "thumbs up," (i.e. clicking on the screen’s thumbs up icon, or other forms of deep engagement) it receives positive reinforcement. Conversely, a "thumbs down" or direct user corrections to the text of incorrect predictions provide a negative reward signal, prompting the AI to recalibrate.

Consider an instance where a user queries the essential skills for a Semiconductor Facility Manager, and the AI’s response, though correct, is incomplete. If the user suggests additional vital skills, this input becomes a reward signal, indicating to the AI the need to include these skills in future outputs. Such user feedback loops permit the AI to constantly learn and refine beyond the initial training data, enhancing the precision and value of future interactions.

Reinforcement Learning via Human Feeback (RLHF) Example:

A user inquires: "What is the most important skill for a Semiconductor Facility Manager to have?"

Gen AI’s response: "Ensure adherence to safety protocols, manage production schedules, and verify that machinery is operating efficiently."

User’s feedback: (clicks the thumbs down and types the following…) “That’s a solid start, but also consider that a firm grasp of SEMI standards, along with proficiency in ISO 9001 and Six Sigma, is critical for a Semiconductor Facility Manager.”

In this example the feedback can be harvested as reward signal and expressed back to the LLM / neural network in what is known as an input/output pair (a type of training data format) then used to improve future predictions.

Examining the dominant Gen AI implementation patterns through a RSC lens.

[A comparative chart of the patterns is provided at the end of the article.]

Unlike most technologies, which depreciate over time, Gen AI solutions can significantly appreciate in value as they continuously learn from reward signals. But who owns this increasing value, and who stands to benefit? We will explore this by examining common Gen AI implementation patterns in the corporate world from this angle.

For illustrative purposes, we’ll apply a consistent business scenario across each pattern to simplify the comparison. Our hypothetical company, XYZ Corp, aims to utilize Gen AI to assist their hiring managers in creating job descriptions that evolve with the company's needs.

Please note that the implementation patterns described here are not exhaustive; new patterns are constantly emerging. The patterns depicted here are illustrative, and various components of these patterns can be tailored to overcome the pattern’s limitations.

Pattern 1: Leveraging a General-Purpose Large Language Model (LLM)

In this scenario, XYZ Corp, a longtime client of FutureTech, has adopted the latter's new software feature that leverages Gen AI to write job descriptions. Behind the scenes, this feature is enabled by a software connection built between FutureTech’s product and one of the well-known General Purpose LLMs we'll call this one, Gigantamax 2.0.

Here's the process as experienced by an XYZ Corp hiring manager:

Step 1: Launch

XYZ Corp informs all hiring managers about a new AI-driven approach to generating job descriptions.

Step 2: Write a Prompt

A hiring manager at XYZ Corp composes a prompt for a job description tailored for a Semiconductor Facility Manager. Since XYZ’s hiring managers are new to “prompt writing” the prompt itself is rather “sloppy” for example:

Mgr’s Sloppy Prompt: “Urgently needed: SMF expert with hands-on DUV tech experience, a pro in green manufacturing practices, and lives by agile innovation. Must be fluent in EHS protocols, proficient in 5S, and up-to-date with REACH regulations.”

Defintion: Sloppy Prompt [slop-ee prom-pt] noun : A loosely structured or vaguely specified input given to a Generative Artificial Intelligence (Gen AI) system, which may lack clear direction, specific context, or precise terminology needed to generate highly accurate and relevant outputs. Such prompts can result from users' unfamiliarity with the nuances of prompt crafting, leading to outputs that may not fully capture the intended requirements, nuanced context, or specificities of a task or query. These prompts may also include jargon, abbreviations, or references that are not universally understood or are too specific without adequate explanation, further complicating the AI's task of generating useful responses.

Step 3: Generate a Predicted Response

FutureTech's software, utilizing its software “bridge” to the Gigantamax 2.0 LLM, creates a job description. In this scenario Gigantamax 2.0 will not be able to express the specific aspects of XYZ Corp, for example the importance of its distinctive corporate culture or the most recent requirements of the role, nor will it be able to decode any of XYZ’s company-specific acronyms. Not only because the prompt is sloppy but equally important the LLM has not been trained on XYZ Corp’s data. Additionally it will not be aware of regulatory changes made in the global business environment that were enacted 6 months ago because Gigantamax 2.0's training cutoff date was 12 months ago.

Accordingly, Gigantamax 2.0’s predicted response is as follows: “Seeking an experienced Semiconductor Facility Manager familiar with Environmental Health and Safety (EHS) protocols and 5S lean manufacturing principles. The ideal candidate will drive innovation and agility within the team.”

While this response captures a broad outline of the role, touching on the EHS protocols and lean manufacturing principles mentioned in the Hiring Manager’s sloppy prompt it missed:

·?????? Hands-on experience in DUV lithography.

·?????? Commitment to the company's important 'Think Green' initiative.

·?????? Fluency in the latest REACH regulations.

·?????? Need to highlight company culture around rapid tech adoption.

Step 4: Reward Signal Generation

If the job description doesn't meet expectations, the hiring manager can provide feedback through edits or a "thumbs down."

In this case the Hiring Manager clicks the thumbs down icon on FutureTech’s software interface and types the following back into the prompt: “This is too generic. We need specific experience in DUV lithography and a strong alignment with our new 'Think Green' initiative. Also, emphasize our urgency and the need for a culture that fosters rapid tech adoption. Plus the regs!!”

Step 5: Reward Signal Capture

FutureTech's system captures the manager’s feedback, containing XYZ Corp’s specific insights, and stores them to be eventually used to generate reward signals.

Step 6: Value Generation

These feedback signals are pooled with data from other clients of FutureTech, leading to improvements in FutureTech’s instance of the Gigantamax 2.0 LLM that are implemented system-wide twice a year.

Pattern 1 Summary:

As XYZ Corp continues to use FutureTech's software, their feedback joins a larger pool of reward signals that benefits all of FutureTech's clients, including XYZ Corp's competitors. With Pattern 1, XYZ Corp both contributes to and benefits from the improved capabilities of FutureTech's instance of Gigantamax 2.0 LLM. However, this arrangement hinges on their ongoing partnership with FutureTech and FutureTech's relationship with Gigantamax.

If XYZ Corp terminates the relationship, they would lose access to the reward signals they generated, as these are retained by FutureTech and likely non-transferable back to XYZ Corp in a format useful for training another AI system. Additionally, if Gigantamax 2.0 is succeeded by Gigantamax 3.0, the improved 2.0 instance—enhanced with XYZ Corp's Reward Signals—might become obsolete, rendering those reward signals inaccessible due to version incompatibilities.

Despite these potential pitfalls, this pattern of sharing and benefiting from collective reward signals makes sense in a variety of scenarios. For instance, when the task at hand does not differentiate XYZ Corp from its competitors—such as basic customer service inquiries or standard operational processes—sharing reward signals does not compromise competitive advantage. Additionally, if XYZ Corp is not large enough to generate a substantial amount of reward signals on its own, participating in a shared pool allows them to benefit from the collective intelligence and enhancements derived from larger datasets than they could provide alone. Another scenario where this could be advantageous is when a company is entering new markets or sectors and lacks the specific domain expertise. Contributing to and accessing a shared pool of reward signals can accelerate learning and adaptation in unfamiliar territories.

Pattern 2: Basic Retrieval Augmented Generation (B-RAG)

In Pattern 2, XYZ Corp deploys a system developed by their IT team and a tech partner, integrating a vector database into their application. This database stores XYZ Corp-specific information (such as existing job descriptions, training manuals, annual reports, legal briefs regarding regulatory changes, et. al.) for use by a General-Purpose LLM in responding to hiring managers.

Step 1: Launch

XYZ Corp notifies hiring managers about a novel AI-driven process for creating job descriptions.

Step 2: Write a Prompt

A hiring manager submits a request for a Semiconductor Facility Manager job description using a sloppy prompt, mirroring the prompt in Pattern 1.

Mgr’s Sloppy Prompt: “Urgently needed: SMF expert with hands-on DUV tech experience, a pro in green manufacturing practices, and lives by agile innovation. Must be fluent in EHS protocols, proficient in 5S, and up-to-date with REACH regulations.”

Step 3: Generate a Predicted Response

The B-RAG tool analyzes the prompt against the vector database, pulling similar text from documents loaded into the vector database like XYZ’s 2023 Sustainability Report, 2023 Annual Report, a recent Regulatory Brief and some rather dated Semiconductor Facility Manager training manuals from the early 2000’s. These snippets are combined with the Hiring Manager’s sloppy prompt and sent to the LLM, which crafts a job description.

Example - finding similar text (vector similarity):

The following illustration depicts B-RAG in action.

1) Retrieval – arguably the most important step in B-RAG – from the vector database

2) Augment, when the system prompt (denoted by the three solid lines, created by the LLM provider to govern all output), the Query (i.e. “mgr's sloppy prompt”) and the context (i.e. text snippets from the vector database) are combined

3) Generate – when the LLM summarizes all of the text

The following lists the contents of the system prompt, the query (sloppy prompt) and the context (extracted snippets from the vector database) prompt send to the LLM for summarization (note: the prompt contents are not literal text, the contents are numeric representations of the text called vectors / embeddings)

Step 4: Reward Signal Generation

Upon evaluation, the hiring manager recognizes improvements due to the vector database integration but also notices gaps stemming from the absence of specific leadership skills, which prompts the manager to provide edits as feedback.

Accordingly, the manager clicks the thumbs down icon and types the following back into the prompt: “Improvement noted, but it missed leadership skills? If XYZ Corp leads with agile solutions, our managers would obviously need related leadership skills like a focus on continuous improvement – why did you exclude these skills?”

Step 5: Reward Signal Capture

The current B-RAG setup does not incorporate a method for RSC and learning from the manager’s feedback, meaning this valuable insight remains unused for future enhancements.

Step 6: Value Generation

To refine future job descriptions, XYZ Corp faces the task of regularly updating the vector database with new and pertinent information, (e.g. regulatory briefs, updated training manuals, etc.) a process that proves to be hit-or-miss and may require considerable manual effort. The LLM’s learning and improvement are entirely reliant on the initial prompt's quality and the vector database content – the LLM never “learns” in this pattern, these steps are repeated each time.

Pattern 2 Summary:

XYZ Corp capitalizes on the synthesis of their unique data with the prowess of a General-Purpose LLM, producing job descriptions that not only resonate with the company’s specific lexicon but also reflect its unique culture. However, a significant drawback of this approach is its static nature—the LLM does not actually learn or retain knowledge from one query to the next. Each query initiates an identical search through the vector database, with the LLM summarizing the results. This renders the LLM a complex tool for search and summarization, rather than an entity capable of learning and improving over time.

The system’s evolution is confined to labor-intensive updates to the vector database, which aim to enhance the likelihood of retrieving beneficial data. The improvement of the AI’s performance hinges on this continuous manual data enrichment.

If you’re thinking that the entire purpose of this pattern is to compensate for poorly written sloppy prompts, and the LLM not being trained on XYZ's data … you’re right.

Pattern 3: Advanced Retrieval Augmented Generation (A-RAG)

Pattern 3 elevates the B-RAG model by introducing a cross-encoder into XYZ Corp's infrastructure, enabling the system to learn from Reward Signal Capture. To understand this pattern, you need to understand that “similarity” does not equate to “relevance.” Although vector databases select text based on vector similarity, they may not always capture contextually pertinent segments. For instance, the previous pattern's Hiring Manager feedback questioned the absence of “a focus on continuous improvement” in the response, underscoring the manager's expectation that “agile solutions” would inherently require a focus on “continuous improvement.” Despite a low vector similarity, the hiring manager sees a significant relevance between “continuous improvement” and “agile solutions.” With the aid of RLHF, the cross-encoder is designed to link these concepts.

Here’s what it would look like from the perspective of a hiring manager at XYZ Corp.

Step 1: Launch

XYZ Corp announces to its hiring managers the rollout of a new AI-powered method for generating job descriptions.

Step 2: Write a Prompt

Just like in Pattern 1 and 2, an XYZ Corp hiring manager requests a job description for a Semiconductor Facility Manager through a sloppily written prompt.

Mgr’s Sloppy Prompt: “Urgently needed: SMF expert with hands-on DUV tech experience, a pro in green manufacturing practices, and lives by agile innovation. Must be fluent in EHS protocols, proficient in 5S, and up-to-date with REACH regulations.”

Step 3: Generate a Predicted Response

The system initially extracts content based on similarity. Now, leveraging previously generated reward signals, the cross-encoder re-ranks the vector database’s retrieved text segments and places references containing “continuous improvement” higher on the list. Accordingly, it produces the following:

Pattern 3 Gen AI’s Predicted Response: “Immediate need for a Semiconductor Facility Leader versed in DUV lithography. Agile innovation is our mantra, so candidates must champion EHS, embody 'Think Green,' and lead with 'continuous improvement.' Critical to navigate new regulatory landscapes, especially REACH and push the envelope of innovation and tech adaptation.

Step 4: Reward Signal Generation

Mgr. clicks the thumbs up icon and types the following back into the prompt: “This captures the essence of what we're seeking. The emphasis on modern and visionary leadership within ‘green manufacturing’ represents our current initiatives. The detailed focus on ‘agile innovation’ conveys our proactive approach to staying at the forefront of industry innovation."

Step 5: Reward Signal Capture

This positive feedback strengthens the system's future responses, ingraining a pattern of aligning “continuous improvement” closely with “Agile innovation”.

Step 6: Value Generation

With accumulated feedback, the cross-encoder is fine-tuned to enhance its relevance recognition. Consequently, when a different manager requests a job description, the system effectively balances technical and leadership skills.

Pattern 3 Summary:

A-RAG considerably improves upon its predecessors by merging the LLM's summarization functions with the contextually rich data from XYZ Corp's vector database. It goes beyond mere similarity searches, now capable of recognizing and prioritizing relevancy. This enhancement results in a system that not only keeps pace with XYZ Corp's needs but also provides responses with heightened accuracy and relevance, positioning A-RAG as a growing asset to the company.

Pattern 4: Single-Purpose Small Language Model (SLM)

This pattern represents a shift from using vector databases and cross-encoders to focusing on "Single-Purpose Small Language Models (SLMs)" that incorporate Reward Signal Capture.

SLMs are significantly more streamlined than traditional LLMs, having less than 100 billion parameters (compared to LLMs that can approach 1 trillion parameters), and are custom-built for specific tasks. For instance, while a trillion parameter LLM is capable of something as amazing as mimicking the speaking style of well-known theoretical physicists and teach quantum mechanics at various educational levels, it would lack a nuanced understanding of your company's unique culture and the specific skillset your workforce needs. In contrast, an SLM would excel in producing outputs that reflect your company's culture and workforce skill requirements in detail but would be unable to answer questions related to quantum mechanics.

From the perspective of a hiring manager at XYZ Corp, the process would unfold as follows:

Step 1: Launch

XYZ Corp informs its hiring managers about the introduction of a new AI-driven method for generating job descriptions.

Step 2: Write a Prompt

A hiring manager at XYZ Corp drafts a sloppy prompt for a Semiconductor Facility Manager’s job description, following the approach used in prior patterns.

Mgr’s Sloppy Prompt: “Urgently needed: SMF expert with hands-on DUV tech experience, a pro in green manufacturing practices, and lives by agile innovation. Must be fluent in EHS protocols, proficient in 5S, and up-to-date with REACH regulations.”

Step 3: Generate a Predicted Response

The SLM crafts a job description that closely aligns with XYZ Corp’s specific corporate culture and skill needs.

Pattern 4 Gen AI predicted response: “XYZ Corp urgently seeks a Semiconductor Facility Manager with specialized experience in DUV lithography. A visionary in EHS and our 'Think Green' initiative, the ideal candidate leads with innovation, mastering regulatory challenges introduced by REACH, and exemplifying our ethos of agility. Exceptional in driving 'continuous improvement,' they are the linchpin in our quest to redefine industry standards through leadership that marries technical prowess with sustainable practices.”

Step 4: Reward Signal Generation

Upon reviewing the output, the hiring manager acknowledges its tailored accuracy.

Mgr. clicks the thumbs up icon and types the following back into the prompt: “This perfectly captures what we're looking for. It addresses all of our requirements, including the urgency, specific skills, and our cultural ethos. Well done!”

Step 5: Reward Signal Capture

This feedback is captured by the SLM as a reward signal, informing future enhancements.

Step 6: Value Generation

Based on the accumulated reward signals, system administrators periodically refine the SLM. If the feedback indicates the necessity, they may conduct a targeted fine-tuning session to address the gaps pointed out by the hiring managers.

Pattern 4 Summary:

XYZ Corp's SLM platform, continually refined with specific reward signals, does not just sustain but enhances its performance, thus becoming a valuable and growing asset. These proprietary reward signals remain with XYZ Corp, contrasting with the shared improvement model of general-purpose LLMs. Leveraged by the emergence of specialized Language Processing Units (LPUs), XYZ Corp’s SLM strategy avoids the expense of extensive fine-tuning, instead offering an economical and targeted Gen AI tool that gains value over time. This approach signifies a strategic pivot to specialized, efficient SLMs that significantly cut down on compute costs and legal risk by focusing exclusively on relevant data, as opposed to the more expansive content LLMs are trained on.

Recent research that highlights the effectiveness of using smaller models with carefully curated data for specific tasks.

ARTICLE SUMMARY:

REWARD SIGNAL COMPARISION CHART:

Final Commentary:

• Pattern 1 General Purpose LLM: remains the most general and broadly applicable, suitable for companies looking for a quick and broad solution without deep customization needs.
• Pattern 2 B-RAG: allows for tailored responses using company-specific data but is hindered by its static nature and the effort required to keep the vector database relevant and updated.
• Pattern 3 A-RAG: provides a good balance between customization, learning capability, and scalability. It offers personalized solutions that evolve over time, though it requires moderate effort to integrate and maintain.
• Pattern 4 Single Purpose SLM: offers a compelling combination of customization, ownership of improvements, and efficiency. The initial cost and time to develop are significantly reduced, through the use of open-source models and the advent of LPUs. Potentially significant effort is needed to acquire and customize the training data to fit XYZ Corp's specific needs.

Disclosure: The views and opinions expressed in this post are solely my own and do not necessarily reflect those of my employer.

Thank you:

Wow, you've made it to the end. Thank you so much for joining me on this journey through the world of Generative AI. I deeply appreciate your willingness to dive into such complex topics—it means a great deal to me.

To express my appreciation, I'd like to share one additional insight with you regarding Gen AI's progression through the hype cycle to the "Plateau of Productivity." When will we know it has happened? When will we know Gen AI has reached the "Plateau of Productivity?"

History may offer us some clues. Do you recall the phrase “mobile first?” This term was often used by software sales teams to emphasize their company's commitment to mobile devices. Mobile was not just another feature in an existing product; it was the product. Early in the internet hype cycle, as established players added internet features to their products, a group of companies ventured into the space with products native to the internet. These companies became the pioneers of what would become the “cloud.” It is very difficult for established players to cannibalize their existing business to feed their ventures into the new world brought by disruptive technologies. Think about the big brick-and-mortar companies that once dominated retail—giants in their era, attempting to integrate e-commerce into their traditional business models. Then came the competition, born and raised in the digital world unburdened by the past, completely changing the game.

When will we know Gen AI has reached the "Plateau of Productivity?" We'll recognize it when companies start describing their products as “Gen AI-first.” Gone will be the days of employees navigating through software with drop-down menus and modules built on relational databases—mobile technology was just the beginning. Instead, these systems will be replaced by those that interact with users through Natural Language Processing (NLP). These advanced systems will have the capability to generate user interfaces on the fly, tailored to specific tasks, and will employ multi-modal technology to see and hear their users.

To understand how imminent this change is, try an experiment: draw something, take a picture, and submit it to any major general-purpose LLM. Ask it to describe your drawing and watch as it interprets your image. Here's one you can use if you'd like to try it immediately - it's from a whiteboarding training session I conducted for my team.

... keep in mind everything depicted on this whiteboard was explained in this ten plus page article - take a moment and try to interpret my scribble. Here's what the LLM responded with in less than 5 seconds:

"This image shows a whiteboard with various diagrams and notes that appear to relate to machine learning or artificial intelligence concepts. The board includes references to "vector", "pre-trained LM (Language Model)", "prediction", "embedding layer", "decoding layer", and "loss function". These are common terms and concepts used in the field of AI to describe components of neural networks, particularly those involved in natural language processing and understanding. There are also acronyms like "RAG" and "LLM", which might refer to specific types of models or methods, such as Retrieval-Augmented Generation for "RAG" and Large Language Model for "LLM". The nature of the sketch suggests it was used during a brainstorming session or a collaborative meeting to discuss AI model architecture or data flow."

We are rapidly progressing through the hype cycle. The real game-changers will be the companies that are "Gen AI-first." This is not about established companies merely adding AI to their existing operations. It's about new companies — or bold, existing ones — that embed Gen AI into the core of their identity, strategy, and operations from the outset. Keep an eye out for these "Gen AI-first" companies—they'll be the ones demonstrating that Gen AI has indeed become a key pillar of our digital and economic landscape.

Thank you again for your dedication and for accompanying me on this fascinating journey.

?

Remember,

"Audentes fortuna iuvat."