Rich Heimann’s Post

📢 Calling All AI Readers! 📢 If you've already read (or are just curious about) "Generative Artificial Intelligence Revealed: Understanding AI from Technical Depth to Business Insights to Responsible Adoption," I'd love your thoughts! 📖✨ Whether you're interested in AI fundamentals, practical tips, or just the thrill of being a book tastemaker, your review could be the spark that influences other business leaders to find the insights they need. Please help me by sharing your feedback; click the link to BookSirens below to leave a review or download a review copy to get started reading and reviewing. 🙏 Thank you for helping spread the word (and playing a part in the algorithmic game). https://lnkd.in/e2Zvpafr

Dario Amodei, the CEO of Anthropic, says he "dislike(s) the term AGI" despite using it frequently over the past several years. He's created a new term: powerful AI. What is powerful AI? He imagines it as a "country of geniuses in a datacenter." Powerful AI is an example of moving the goalposts and aligning ad hoc goals with one's research. It is self-serving and fails to establish a shared standard for AI. Moreover, this definition fails to describe intelligence and doesn't rise above tautology. It merely redefines intelligence as something geniuses do in an oddly specific location. We can add powerful AI to a litany of names such as narrow, general, super, weak, and strong (all poorly defined). Yet, as a rule of thumb, the number of categories erected to explain an area of research is inversely proportional to the available knowledge about that research, thus anesthetizing so-called AI speedometers and AGI countdowns. That said, the rest of the article attempts to distance powerful AI from the x-risk in AGI. At nearly 15,000 words, it is not light reading, but I agree with Amodei that all the x-risk is overstated, and I suppose his attempt to codify risk is something.

On one level, Andriy Burkov's advice (below) makes sense. LLMs (e.g., ChatGPT) can tell, but it doesn’t know. As ChatGPT responded to me in one chat, “I will always do my best to provide a helpful and accurate response,” followed by, “I don’t have the ability to know what I don’t know.” This piece of Rumsfeldian wisdom is a reality for LLMs, which don’t "know" anything. The consequence for users of LLMs appears to be a paradox: either you know what you are looking for, meaning you don’t have to look, or you don’t and cannot expect to find it. For questions where a user knows the answer, ChatGPT is trivial, and for questions they don’t, users will be unable to recognize if it is correct, and neither will ChatGPT. The mysterious origins of knowledge can make ChatGPT’s value seem fickle. However, on another level, this advice underestimates human potential and mischaracterizes epistemology. After all, if you can ask a question, you have already demonstrated some knowledge and some desire to know the truth required to build an epistemological framework. Meno's paradox challenges how we come to know things we genuinely don’t know. It is not an either/or situation—that we must already possess the knowledge to ask questions or that without such knowledge, we cannot recognize the truth. This false dichotomy diminishes human inquiry's complex, iterative nature. The most charitable interpretation of Burkov’s advice is cautious. It’s a reminder that while LLMs are valuable tools, they are not infallible (though what is?). Another interpretation is gatekeeping, which calls those using the technology engaged in "self-sabotage." It's important to remember that LLM's limitations don’t preclude its value as a learning tool; they indicate that it shouldn’t be the only source (which, by the way, is impossible). When used alongside other resources like experience or checked against different sources, ChatGPT can be a helpful assistant. Of course, that is not the hot take that everyone wants on social media platforms.

Generative Artificial Intelligence Revealed is now available on Audible! 🎧 https://a.co/d/2NYe0AK Enjoy listening, and feel free to let me know which chapter made you nod along, nod off, or raise an eyebrow. Drop a comment or send me a direct message to receive a free promo code!

I've met several entrepreneurs who overemphasize AI's positive impact on their own business and product development while underestimating its broader, systemic effects on the competitive landscape and customer ecosystem. Entrepreneurs must consider these broader systemic effects, as they may focus on how AI will revolutionize their product offerings or create new market opportunities. However, they should also consider that competitors can quickly replicate or leapfrog the same innovations and that customers faced with similar AI-driven options may react unexpectedly, shifting their loyalty or changing their demands. The error arises when innovators attribute their success to their use of technology while downplaying situational factors like market dynamics, consumer behaviors, regulatory challenges, and competitor responses that will similarly affect others in their industry. This reflects a kind of myopia: they believe AI gives them an edge, but in reality, it gives everyone an edge. The result is a self-serving bias mixed with elements of an attribution error. Entrepreneurs might attribute the success of their product or the inherent superiority of their approach (self-serving bias) while downplaying the environmental factors that allow competitors and customers to catch up just as quickly (attribution error). This leads to overestimating advantages and underestimating how the same technology could empower others equally, if not more effectively, especially when external conditions like industry hostility are involved. To estimate both the impact technology will have for you and on you, it's essential to take a holistic approach. By examining both the positive (how technology can drive your success) and negative (how it can disrupt you) dimensions, you can better leverage the benefits while mitigating the risks. 𝗣𝗼𝘀𝗶𝘁𝗶𝘃𝗲 𝗦𝗰𝗲𝗻𝗮𝗿𝗶𝗼 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴: • Evaluate how technology can boost your strengths. • Identify how technology aligns with industry trends or unmet customer needs. • Map how the technology can improve your product, reduce costs, and differentiate your offerings in the short and long term. 𝗡𝗲𝗴𝗮𝘁𝗶𝘃𝗲 𝗦𝗰𝗲𝗻𝗮𝗿𝗶𝗼 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴: • Consider how others in your industry might leverage the same technology.  • Consider how this technology will impact customer expectations. • Will they remain loyal if it offers alternative solutions faster? • Could the technology commoditize your offering? • Assess how low barriers to entry with this technology could attract new competitors who may disrupt the market. 𝗗𝘂𝗮𝗹-𝗜𝗺𝗽𝗮𝗰𝘁: • Recognize that what benefits you will benefit competitors. • Map out where you might gain or lose ground in the competitive landscape. 𝗕𝗹𝗶𝗻𝗱 𝗦𝗽𝗼𝘁 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀: • Seek input from industry experts to uncover blind spots. • Entrepreneurs often miss the broader implications of technological changes because of cognitive biases or a narrow focus on their advantages.

The OpenAI founding team Only three of OpenAI's thirteen founding members remain. While this is a large founding team, the churn is high, especially for a $150b company that might be on track to a trillion-dollar valuation.

Language models evaluate what comes next in a sequence by weighing probabilities. This process shapes their behavior. It is not about complexity; instead, it is about frequency. A key implication of this probabilistic nature is how language models handle tasks of varying commonality. Regardless of model size, they inherently perform better on tasks they encounter more frequently in their training data partly because it is memorized. Even if two tasks are equally complex, the model performs more adeptly on the task it recognizes as more common. Even in deterministic tasks—where there’s an unequivocal correct answer—the model’s responses are influenced by the statistical likelihood of potential outputs. Think about Google Search. Intuitively, a language model will prefer prompts that correspond to more search results but struggle if a prompt yields few search results and will hallucinate when the prompt is orthogonal to the consensus of the training data. The image (below) summarizes the Monty Hall problem, a classic probability puzzle. It describes a scenario on a game show where you're given a choice of three doors, with one door hiding a car and the others hiding goats. The counterintuitive conclusion is to switch doors after Monty reveals a goat. This counterintuitive aspect of the problem is a well-known example of how our intuitions about probability and risk can often be misleading. The variant below by François Chollet changes the setup, reveals a car behind the first door, and then asks if you want to change your choice. The text asserts that it is advantageous to change your choice despite already winning the car. This advice is based on the classic setup predicated on someone picking a door goat. This means a model’s predictions are not just about crafting the killer prompt. Certainly, prompting and prompt engineering are essential. Studies have shown that a good prompt can significantly impact model performance, varying by more than 10 percent, even without changes to the model itself. However, a model’s response is also shaped by how probable the correct answer is within the context of its training. This distinction highlights a nuanced aspect of language models: their responses are not merely reactions to prompts. They are based on their training, which the statistical likelihood of possible outputs can significantly sway. This influences whether you add additional information to a prompt using prompting engineering or an external database or whether you need to fine-tune your model.

The clip below is by François Chollet (Google), who recently gave the keynote at AGI-24. In it, he discusses the distinction between intelligence and its result of intelligence (i.e., task-specific skills). He says, "Intelligence is a process, and skill is the output of that process... skill is not intelligence." An artifact of the cognitive revolution was functionalism. Functionalism views intelligence (i.e., mental phenomenon) as the brain’s functional organization where individuated functions like language and vision are understood by their causal roles. Functionalism focuses not on how something works or if it is made of the same material. It doesn’t care if the thing that thinks is a brain or if that brain has a body. If it appears to function like intelligence, it is intelligent like anything that tells time is a clock. As long as it keeps time, it doesn’t matter what the clock is made of (or if it is a simulation of a clock). This all reminds me of the so-called AI Effect, which friends of AI will say occurs when skeptics discount the behavior of an AI as not "real" intelligence. Edward Geist credits John McCarthy for coining the term "AI effect" to describe this phenomenon.[1] The author Pamela McCorduck writes: "It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking.'[2] McCorduck calls it an "odd paradox" that "practical AI successes, computational programs that actually achieved intelligent behavior were soon assimilated into whatever application domain they were found to be useful in, and became silent partners alongside other problem-solving approaches, which left AI researchers to deal only with the 'failures,' the tough nuts that couldn't yet be cracked."[3] Friends of AI will call this moving the goalposts. But, it's not goalpost moving if you don't know your goal. AI research is simultaneously lost and astoundingly vibrant—making tools that work, yet not toward the goal friends of AI believe they are pursuing. It's a kind of futile progression. Like all problem solvers, friends of AI create shortcuts to solve problems. These shortcuts are required and often take careers to discover. However, these shortcuts subvert the ostentatious goals because solving a problem means you are not solving intelligence. These efforts produce problem-specific solutions that, at best, simulate intelligent tasks that do not represent human cognition and lack cognitive plausibility. We may be clever enough to ponder the mysteries of intelligence but not yet wise enough to unravel them. So far, our ability to imagine artificial minds far exceeds our capacity to create them, yet these arguments do not immediately subvert the usefulness or economic utility of these tools.

I just gave Google’s NotebookLM my entire book (co-authored with Clayton Pummill), and what came back was a surprisingly good way to analyze a document. If you're unaware, NotebookLM turns books and documents into an AI-powered NPR-style podcast plus a study guide, FAQ, timeline, and a surprisingly accurate chat experience—all in one. The podcast version is fun and provides a compact dive into the book’s themes. That said, it skims the surface and doesn't completely capture the book's nuance. Still, neither do most podcasts due to the temporal constraints and the limitations of an author's ability to cover the same information verbally as in writing. It’s tough to match the depth and precision of writing reflected in painstaking rewrites—after all, that’s what writing’s for, right? What also stood out was how the “hosts” used clever tricks to keep the conversation going. It all feels very organic. I thought it would feel contrived and give me flashbacks to the Rogerian tricks ELIZA used. Perhaps this format will become grating after more iterations, but my first impression is that NotebookLM entertainingly captures the core ideas and does so in an engaging manner. I will certainly be trying it on some ARVIX papers that I have not had the time to read. You can check it out here:

𝗜𝘀 𝗼𝟭 𝗮 𝗴𝗼𝗼𝗱 𝗰𝗮𝗹𝗰𝘂𝗹𝗮𝘁𝗼𝗿? Yuntian Deng (X: @yuntiandeng) recently conducted an interesting test involving the performance of o1 in solving up to 20x20 digit multiplication problems. The results indicated that o1, depicted in the top image (below), can effectively handle multiplication up to 6x6 digits and 9x9 with decent accuracy. In contrast, GPT-4o, displayed in the middle image (below), struggles with multiplication beyond 4x4. The reason why 4o specifically struggles is that these large multiplication problems are underrepresented on the Internet, and 1o shows that CoT unlocks greater generalization. Deng also shows that for 20x20 multiplication, o1 utilizes approximately 3,600 reasoning tokens. In comparison, human reasoning requires around 400 reasoning tokens for partial products and a similar number for sums. This shows that CoT is computationally expensive, which accounts for the extra time. It is not system two thinking. As an aside, Deng has done some exciting research using implicit CoT. Deng et al. start with a model trained for explicit CoT reasoning, gradually remove the intermediate steps, and fine-tune the model. This process allows the model to internalize the intermediate reasoning steps. This approach does not work as well as explicit CoT, but it is close and performs at the same speed as models without CoT, thereby aligning with the intrinsic computational strengths of machine learning. 𝗩𝗲𝗿𝗱𝗶𝗰𝘁: o1 is not a good calculator nor great at reasoners, though CoT does help these models perform various tasks and generalize better from their training data. Reasoning, much like chess in the early days of computational research, is emerging as the frontier that will attract the focus of researchers. I think we will continue to see small incremental gains in these areas. However, I believe Mike Knoop (co-founder of the ARC Prize) put it best (link below): “o1 represents a paradigm shift from ‘memorize the answers’ to ‘memorize the reasoning’ but is not a departure from the broader paradigm of fitting a curve to a distribution in order to boost performance by making everything in-distribution.”