David Evans’ Post

In the early days of SMS, you'd think twice about sending an emoji (albeit they didn't exist yet) because each text had an implicit cost. That cost constrained usage, a pattern we are seeing repeat in AI. AI processes have two parts: training and inference. Training is the process of ingesting data and developing the model (aka software). Inference is using the model to create predictions, batches of text, images, etc. Training happens once over a very long, computationally expensive period of time. Inference happens millions of times a day as users interact with a model. It's lower cost, relatively speaking, but happens at a massive volume. So, while large players are looking to cut the energy costs of training, it's just as important to reign in the costs of inference to make the technology accessible, just like unlimited texting. https://lnkd.in/gCw2BZUe

Vertical software has gotten a bad rap because the total addressable market (i.e. how big the market is) can be constrained. In other words, investors love TAMs where your customer is fundamentally "everyone" because it's easy math: huge market = big company potential (QED). That said, companies like Procore and Toast have been very successful in scaling in vertical markets by delivering industry specific value. AI is only going to increase the need for this level of specialization. How do you optimize AI for restaurants? By providing it with as much restaurant-specific data as you can. Maybe generic web data from LLMs like ChatGPT can augment those insights, but the real core insight needs to be industry specific. It's no different than human counterparts. Who's better to run a restaurant turnaround show, Gordon Ramsay or Jim Cramer? https://lnkd.in/grcnbtnq

Inference is the process of "running" AI models. The challenge here, unlike most web systems, is that they can't be run on commodity hardware. While platforms like Google are running on cost optimized hardware, platforms like Character AI are running on high-end GPUs from NVIDIA. This is going to lead to a host of computational improvements that reduce the infrastructure costs of running AI, and this is not new. In the early 2000s, Google developed a technology called MapReduce to manage the vast volumes of data they needed to process, leading to an entire ecosystem of products and services. Some of the most interesting parts of the AI revolution won't have anything to do with generating text or images, but with efficiently providing the infrastructure to do it. https://lnkd.in/gkjmnFQ2

Vertical AI may or may not be the future of AI, but it's certainly the current cycle in AI. Controlling outputs from AI applications is much like a wooden labyrinth game. You know the path, but keeping the ball on the path isn't quite so straightforward. The smaller the game board, the easier it is to navigate the maze. That's the push for vertical AI. By controlling the range variables, we effectively keep the data game board smaller and easier to manage. Plus, by going vertical, it allows for getting deeper insights. The challenge long term will be scale. Most vertical organizations will be pushed to flatten to grab larger and larger market share as they scale. So don't be surprised when Bessemer starts talking about flattening the curve. https://lnkd.in/gcHeeuwU

Fine-tuned retrieval augmented generative systems (RAG) from companies like LexisNexis still hallucinate, or generate errors, in about 17% of cases. That's a far cry better than the 40%+ of generic use models like GPT, but still not good enough. AI does not deal well with subjectivity because, after all, at its core are still mathematical models. And at its core, the law is subjective. But there are a lot of things that are subjective, from medicine to business strategy. The key mental model is to stop trying to build systems that have the answers. Build systems that let users comb through more data and synthesize that into rational responses. https://lnkd.in/gQKsbT7R

Thin applications that solve the limitations of other software systems have been around forever. Things like antivirus or web browsers were all once unbundled, and in SaaS, everyone tries to make systems like Salesforce not quite so hard to use. AI has exacerbated this problem because for most of the last two years, AI-driven features just didn't exist in the underlying product. The challenge with "thin wrapper" products is that they are not truly innovative. Real innovation means you need to re-think the entire process, the data, and the outputs. If you don't, you're just one product management meeting away from becoming a part of the system you sit on top of. https://lnkd.in/gmkNuSna

When talking about AI, I'll generally use a broad umbrella that includes both generative AI (ex. ChatGPT) and machine learning (ex. predicting when a machine will fail based on historical data). In both cases, the machine is using data that it has seen in the past to extrapolate a future state. With machine learning, that prediction is generally a finite result: predicting a numerical value or the most likely option. So the outcomes have some level of predictability. With generative AI, it's a more amorphous result. Using a batch of text or image input to produce a variable output of text. Because of the breadth of inputs and outputs, generative AI can be less predictable. But the reason why they get used interchangeably at times is that, at their core, it's using predictions to determine an outcome. The only real difference is in the complexity of the process and the associated results. https://lnkd.in/g4V3jd64

Yes, churn is bad. But you also need to remember that your product, while the most important thing to you, is likely far from the number one priority of your customers. So, as much as having great onboarding and customer support functions are critical, it's also imperative to ensure that you're not becoming a chore to work with. Are you slowing down time to value with yet another training or walkthrough session? Do they really need to hear from you every 3 months with a feature update? At the end of the day, you're a piece of software that they use to complete their job. Enable that and get out of the way.

Want to know if your job is safe from AI? Ask yourself how much of your job requires deep thought. About 11/12th of a coders work is thought. I can attest to this firsthand after years of solutions devised everywhere except at my desk (in the show, in traffic, etc). While AI can generate code, it's can't replicate these processes. Thought isn't encoded in the source data for AI models. Code, documents, presentations...all of the written output we have fed to AI is a compressed version of all of the context and reasoning that went into that final work product. In other words, the models are missing 11/12th of the work. So, the aspects of your job that involve thought are safe. The rest, not so much.