Introduction
This inaugural article for my newsletter here on Substack is focused on separating fact from fiction in AI, and to cut through the hype machine, which has been full of misinformation. Readers who opted in to my previous newsletter from KYield have been subscribed to this newsletter. For those who have not yet read our executive briefings, I highly recommend the most recent: “What is an EAI OS (KOS)? and “Introduction to the GenAI function in the KOS”.
I also publish a monthly newsletter on Enterprise AI over on LinkedIn, and frequently post articles with commentary on AI, EAI, business, and economics if you’d like to follow me on LI. Although most of my articles on Substack will be on AI and related business and economic issues, I may occasionally write about something else.
The financial stars aligned long ago between LLMs and big tech due to the vast cloud resources required, hence the multibillion dollar strategic investments in LLM startups by big tech firms, and the focus of their massive lobbying machines. Rarely have the stakes been so high or ignorance been exploited in so many. My hope is that 2024 will usher in a new phase that includes much more evidence-based reporting, and less spin.
Myth: Large language models (LLMs) are the only type of AI models.
Truth: Many AI models exist. Until the fall of 2022 very few in the media or public were aware of LLMs. They’ve been around a long time in research labs, but so too have many other types of models. LLMs are among the least accurate or safe, but they are good at reconstructing full content reproduction of text, code, and images in response to prompts. LLMs are favored by big tech cloud providers due to the vast amount of cloud computing resources necessary, hence the massive investments by Microsoft, Google, and AWS. LLMs have been the focus of many researchers primarily due to budget—about 90% of the research budgets have been invested by big tech in self-generating algorithms and related, for obvious reasons. (For an example of a different type of model, see my EAI newsletter edition on neurosymbolic AI).
Myth: GenAI provides a competitive advantage.
Truth: GenAI run on LLMs was among the most rapidly commoditized technology in history, and is already well along in the process of being integrated into the most widely available apps in the world, including Microsoft Office and Google Docs.
While GenAI raises the bar on competitiveness due to the increase in productivity for some functions (particularly coding and writing), commoditization provides no advantage. Some individuals may derive an advantage by becoming more skilled than others in prompting tasks for LLM bots, and some app developers can create an advantage with proprietary high-quality data, software, or machine learning, but GenAI won’t provide an advantage—just higher costs. It can provide an attractive ROI for companies due to higher productivity, though it varies greatly depending on the industry and way it’s applied.
Myth: GenAI is more profitable than SaaS.
Truth: The cost of machine learning has been dropping rapidly while performance has been increasing, but the cost of GenAI is very high, and margins are much lower than SaaS (20-40% lower). GitHub co-pilot for example has been charging $10 per month and cost Microsoft $20 per month for each user (see WSJ article: “Big tech struggles to turn AI hype into profit”). While costs are expected to fall over time, the problem is LLMs are the most wasteful and costly model available (see my recent EAI newsletter: “What investors are getting wrong about AI”). What we need are AI systems like our synthetic genius machine (SGM) that compresses knowledge and data, doesn’t rely on copyrighted data, and is much more accurate, secure, and efficient (Our SGM is still in R&D but part of the technology is currently viable).
Myth: LLM chatbots can be made safe with so-called “guardrails”.
Truth: LLMs and other self-generating algorithms are inherently unsafe and should only be employed within larger AI systems like our KOS that has strong governance and precision data management. We’ve long designed-in these algorithms for specific functions, but they are compartmentalized and only run on high-quality data. The use of the term guardrails to describe attempts to improve safety and accuracy in LLM bots is intellectually dishonest for competent scientists and management of LLM companies. The word guardrail is a false equivalent as it implies a physical barrier when it’s really the use of text and wording, which can be easily overcome (see research paper).
Myth: The greatest risk from AI are rogue bots that will take over the world and kill humans.
Truth: The doomsdayers have also had a record year of hyperbole, and it includes some of the most famous and celebrated AI scientists. Although someday a rogue world killer bot will become a serious risk, and precautions will be necessary, that day is a long way off. The immediate risks are significant, including bioweapons aided by LLM chatbots, and potential economic collapse caused by the transference of the world’s knowledge base to a small group of companies. We’ve studied most major catastrophes over the last three decades and have developed methods to prevent human-caused catastrophes (see Humcat).
Myth: Compensation for use of content by LLM chatbots will protect against a mass extinction event for knowledge workers, writers, artists and creative companies.
Truth: I don’t expect compensation agreements to be recurring for long. New methods are being developed that will likely work around this expense for chatbot companies (if they survive copyright infringement lawsuits). Everyone needs control of their data, which is precisely why we offer control in DANA (Digital Assistant with Neuroanatomical Analytics). DANA is included in every enterprise license for our KOS and can be extended to partners and even individual customers. We protect knowledge work and IP within a system-wide governance system with strong security and precision data management. Data represents all of our assets today and many of our human rights. Without control of our own data, it’s difficult to see pathway that can avoid a dystopian future.
Myth: LLM bots are showing signs of emergent behavior.
Truth: LLMs perform random acts based on inputs and datasets, some of which resemble emergent behavior, and they are unpredictable, but that doesn’t make them emergent. I do agree they are very good at mimicking behavior, and seem life-like at times (see article on topic at HAI).
These myths and truths only scratch the surface, representing the top few on my radar at the moment, based on 26 years of R&D in AI.