Generative AI Roadmap for Business

Hi, I'm market research analyst Mark Beccue. I've been providing market analysis of Artificial Intelligence since 2016, focusing on operationalizing AI in the enterprise. Welcome to my newsletter! Here you will find insights guiding businesses on how to navigate the Generative AI landscape.

Today, I'm sharing foundational advice -- a roadmap for making decisions about how a business can leverage Generative AI.

AI is having a moment. Sparked by OpenAI's introduction of ChatGPT in October, 2022, every business is now looking at what seems like magical outputs from Generative AI, dreaming of the possibilities for their business but at the same time feeling a great deal of FOMO.

Generative AI is a new experience for the vast majority of businesses. To make good decisions about Generative AI, you need to first understand a few things -- why Generative AI is different, what makes Generative AI a challenge and gauging the Generative AI trends. We will cover those issues and then talk about a roadmap.

Here's an outline of what we will cover:

• Legacy AI vs. Generative AI -- what has changed
• Generative AI challenges
• Generative AI trends
• Generative AI roadmap for business
• Conclusions

Legacy AI vs. Generative AI -- What Has Changed

There are important lessons learned about the similarities and differences between Legacy AI (AI-pre Gen AI) and Generative AI:

• Enterprises have been working to implement Legacy AI for nearly 10 years (since 2015), whereas they have been thinking about Generative for less than two years (November, 2022). Impact: No track record for Gen AI, unproven use cases.
• Legacy AI models were built almost entirely on custom data sets -- data owned or acquired by that particular enterprise. In contrast, a significant portion of most Generative AI models are built on a shared data set of public data. Impact: In theory, Gen AI models are faster to build and easier to scale than models built solely on custom data sets.
• Building and operating Legacy AI models typically require data science expertise, building and operating Generative AI models typically do not. Impact: Data science resources are extremely limited. Possibly the biggest factor driving Generative AI interest is the fact that this market barrier has been lifted.
• The market evolution for Legacy AI was, by technology standards, a reasonable pace. In contrast, the pace of market evolution for Generative AI is widely considered one of the fastest moving in history. Impact: Generative AI is a very unsettled market, technology, use cases, costs, risks and impacts are unknown or in flux, causing significant anxiety for enterprises.

Generative AI Challenges

As stated, Generative AI's unprecedented market evolution means there are many unknowns:

• AI models are morphing and multiplying nearly weekly.
• There is a lack of proven use cases or enterprise-grade end to end solutions.
• The legal and regulatory path is opaque.
• Compute for AI workloads, which currently depend a great deal on a limited supply of GPUs, is scarce and cost-prohibitive.
• AI models face challenges for accuracy, hallucination, bias. Without augmentation, they suffer from a lack of real-time knowledge.
• Companies are completely unprepared to understand and navigate AI risk management.
• The market is poorly prepared for new, increased demands in data management and governance.

Generative AI Trends

• Models. Models are evolving extremely rapidly. Consider that in November 2022, the primary LLM, ChatGPT, trained on 175B parameters. Fast-forward 13 months to December 2023 and Microsoft Phi-2 is trained on 2.7B parameters. There is significant LLM competition spurred by open-source model development. There are now hundreds of Gen AI models. "Small" language models are proliferating. Bigger models like ChatGPT are not necessarily what companies need. Smaller models are surging, they are cheaper to run, but are more task-specific. One model does not fit all. Most companies are finding they will use more than one model, depending on what work they want done. There won't be one dominant model.
• Compute. AI workloads are getting smaller, cheaper. Why? Because enterprises are leveraging smaller more efficient models and data centers are starting to use purpose-built AI chips.
• Use Cases. iPhone AppStore Analogy. We are in the very earliest of stages for Generative AI use cases. In thinking about the App Store, the earliest use cases were hit and miss and tended towards non-serious, amusing applications. Over time, enterprises and developers found more pragmatic, useful use cases. The same will apply to Generative AI. "Use LLMs as a reasoning engine to process information, rather than using it as a source of memorized information." -- Andrew Ng. Many users appear to trust the answers an LLM provide too much. The issue is LLMs, without proper grounding and other fixes, can be significantly inaccurate. Promising use cases -- Drug discovery, company search, code development, text summarization. Sketchy use cases -- writing assistants.
• Build vs. Buy. Unlike legacy AI, most enterprises are experimenting with building in-house Gen AI capabilities. One of the reasons why this is so is because there is a lack of end to end Gen AI application specialists providing solutions to the marketplace (for now). It is likely that DIY momentum will slow when many enterprises return to the wisdom of focusing your company's efforts on what you do best, and farm out the rest. As part of that concept, it's encouraging that the most advanced Gen AI applications in the market today come from a growing number of AI-experienced SaaS players, like Adobe , Salesforce , Zoom , 微软 , ServiceNow and others.
• Risk governance, Responsible AI. Lawsuits around copyright and IP. Deepfakes and disinformation. Minimal adoption of best practices. Early stages and efforts in standards, industry groups and advocacy. The EU's AI Act will likely have a GDPR effect on any company doing business in the EU.

Generative AI Roadmap for business

• Start with "what problem are we trying to solve?" Never do technology, any technology, for technology's sake. When you start with a more pragmatic issue -- what problem are we trying to solve -- the goal is focused and you have an endpoint in mind.
• C-Suite commitment. The companies that have succeeded with AI all have something in common -- commitment from the very top of the organization. Gen AI requires patience, resources and budget.
• AI risk management/guardrails. The companies that have succeeded with AI develop detailed AI risk management systems and guardrails. Don't ask "Can we do this AI?" Rather ask "Should we do this AI?" Develop a AI risk management framework for the lifecycle of AI -- not just the ideation phase, but through launch into ongoing operations. Establish an AI governance board or committee to be act as the gatekeeper for every AI use to check transparency, traceability, accuracy and bias. Establish guardrails to protect AI from causing harm or the misuse of AI.

Conclusions

• Generative AI is nascent, the market is fluid
• Generative AI potential is high – democratized ability to leverage DL, ML models, expanding use cases
• Generative AI has inherent flaws that present significant organizational risk
• Generative AI use cases will clarify, inherent challenges will recede by end of 2025
• AI models will commoditize, the key to success with Generative AI will be leveraging proprietary data
• Success with Generative AI – companies with deep legacy AI experience, proving that foundational learnings, processes and organization of legacy AI apply to Generative AI