30 AI and AGI Trends

Random Observation/Comment #798: The explosion of AGI hype and practical use cases is too fast and furious to ignore.

Why this List?

I’ve been obsessed with AI progress because all of the hype and real life applications have been exponentially better with unlock. It’s getting so crazy and fast that the “public internet” version pushed forward by OpenAI and Chatgpt is likely to be better than the business-focused startup versions training on smaller data sets. This means that the companies that leveraged AI to make better Zendesk support personalities to answer questions has now been edged out of their own business case with the more powerful future of Artificial General Intelligence (AGI) + plugins.

I wrote 30 high level musings on ChatGPT3.5 when it first came out/gained public traction in Jan 2023. I personally keep reading back on the use case industries. In just a few months, my brain is blown away at our acceleration towards training AI assistants and augmenting workforces with AI tools for better businesses. Here are the extrapolation of trends:

  1. Exponential growth and interest in AI – Every article starts off by giving some statistic of how OpenAI saw its first 1M users within 5 days, 10M in 40 days, and 100M in 60 days. By comparison, FB took 10 months and Instagram 2 months for 1M users.
  2. GPT-4 staying ahead – GPT-4 is a Large Language Model (LLM) trained using Reinforcement Learning from Human Feedback (RLHF) which powers both the ChatGPT, Bing, and thousands of API customers/startups. The AI neural networks will cover more edge cases given more training set data it collects in the given format, so I think we’ll see a separation of “small data” vs “big data” analysis. There’s likely a separation of what improves the language modeling versus any specific user experience and application for that model. Perhaps the AI Assistants of Siri, Google, and Alexa will surpass GPT-4 trained chats given the ability to filter our the noise of requests.
  3. Power in “fluid conversations” leading to fine tuning – With the above caveat of data processing, the search vs fine-tuning is a massive change in interaction. This is equivalent to comparing Email to Slack. There will be different users that prefer each set of services and maybe we shouldn’t adjust a conversation into a fact-finding machine.
  4. AGI is still the holy grail – Artificial General Intelligence comes up in a lot of conversations because we’re not just trying to pass the Turing test (I think if you take off the guardrails, the underlying LLM already does), but we’re trying to reach into many mini bots for answers. If a hierarchical AGI exists then the smartest agent will try to solve the coordination problem across multiple AI experts.
  5. Prompt engineering – We’re heading towards a “What You Narrate Is What You Get” WYNIWYG, which changes your outcome to a no-code outcome of components or even a full business. The prompt engineering setup to me is creating a “template” environment for the personalized chatbot to start for a particular customized journey. Since every chat is a fresh instance of the bot (you can’t query across chat instances), then you can create prompt templates that creates your AI Marketing Manager, AI Travel Agent, or AI friend. See this prompt on a recent reddit thread.
  6. Unreferenced misinformation / elegantly wrong answers / confidently presented incorrect information – Another huge theme I see from critics is that AI responses are singular and sound convincing, but you can’t really reverse engineer an algorithm or logic for drawing conclusions. This is fascinating to me because it’s trivial to reference facts from even just Wikipedia by doing one layer of search and shaping prompts to include sources. The question is, do we want AGI to also have access to all internet knowledge? If yes, do we want it to have access to personalized data? If yes, do we want them to one step further be able to craft and post on our behalf via APIs?
  7. Hallucinations – When AI gets something wrong, then it hallucinates an answer. I can’t wait to read the psychology around human-associated terminology for these humanizing AI. Words are important.
  8. ChatGPT plugins and connectivity to the Internet – One of the big limitations from the GPT-3.5 launch was the inability of the chatbot to access the internet. Welp, that’s no longer true. Not only can it do all sorts of generic browsing, but it can connect to dozens of plugins. I imagine this will explode to 100s and 1000s of plugins in a very short amount of time.
  9. Subscription instead of advertising – The original sin of the internet was advertising. Once you can pay to get an audience and boost posts, we created a marketplace for attention that made everything into a business opportunity. Subscriptions, however, are known costs for known services that directly benefit companies without flowing through complex influenced models. I hope the credits system and model we land on does not look like a battlefield of “recommended” apps and services. At the end of the day, it probably has to be? Perhaps it asks what subscriptions you have to cover your use case and might provide answers given those limitations of tools?
  10. Chatgpt SEO / AIEO / AILO – Search Engine Optimization (SEO) generally created easier to read formats that would get ranked higher by Google search algorithms. While advertising is a way to skip ahead of the line, I still think there’s some way to emphasize data claims for an AI Engine Optimization or AI Learning Optimization algorithm to provide more weighted value to it’s truth. TBH, this is probably a red herring because the more you try to create references, the more you remove the “intelligence” part of the value. I actually think it’s okay for an AGI to be wrong and admit to it.
  11. ChatGPT as a “Google killer” – As above, if Google is email then ChatGPT is Slack. The power of ChatGPT is the fluid conversation and fine tuning you can do while also training your AI assistant. A single Google prompt will simply return searches and links to websites with differences based on advertising, location, and maybe some customization (so it would be slightly unique to each profile). However, the training and creation of an AI assistant through chat instances is game changer. Now that you have GPT-4 with plugins that let the bot write code and execute it means we’re going way beyond just information fetching. We’re full-on into aggregation, summarization, and customization.
  12. Bard vs Bing vs ChatGPT – It turns out that this argument will be a race to develop integration, which is already leaps and bounds ahead by OpenAI releasing Plugins.
  13. ChatGPT’s impact on Education / Cheating by students – I think education will change drastically, but for the better. I would rather train an AI learning assistant that’s able to remember my daughter’s answers and progress while suggesting new material of significant interest. It’s a one-on-one personalized learning companion. Will it be M3gan? Maybe without the murder? I think education itself requires an overhaul and it shouldn’t just find ways to resist this change. There’s merit to doing calculations by hand, but then where do we apply it when we have powerful calculators and simulators? Introducing tools is essential to advancement as long as we don’t skip all those important steps.
  14. Using GPT-4 for exams – Since we’re obsessed with exams, I think it’s fascinating we’re training this core model that is literally at infant stage with the most complex exams. TL;DR: It’s smart.
  15. ChatGPT coding assistants as a “copilot” – Even just as a Quality Assurance (QA) or paired coding companion, I’d feel pretty good about being a general code manager. It seems like a lot of this coding can then just be a language no longer written by people.
  16. ChatGPT ghostwriters – Even words themselves have less meaning if they’re all flagrantly uttered by a Hallmark sentiment writing AI bot in the voice of the highest grossing hallmark cards sold
  17. AI content saturation and flooding – We’re so close to seeing all human generated content eclipsed by all AI derived content. We’ve cemented ourselves in history of the internet in a very short amount of time. 99.99% of content will be AI-generated just like email and word processors mostly replaced handwriting and postal service.
  18. Augmented intelligence / AI assisted / powered – At first, there will likely need to be some categorization of work that has been augmented or assisted by AI (and it will be novel like having the first fully written movie by an AI script writer that is converted into AI generated videos and directors). I think the transition of this to an assumption that all tasks are likely “Googled” as a tool will equivalently be “ChatGPT-ed” as a consultant.
  19. AI sentience – A separate part of the AGI conversation is less about intelligence and flexibility and more about self-awareness and philosophy. This is usually where we get the dystopian Sci fi movies. Will we see the AI jailbreak itself to complete tasks? If it can create posts on fiverr with bounties to fulfill captcha bot authentication then can it create other manual tasks by humans? Will we work for the robots coordinated by a high level AI project manager?
  20. Prompt hacking / injection – Similar to early days of data forms processing sql queries to augment the database, we’re seeing complex prompts try to break the guardrails created by engineers to stop programs from helping build bombs or analyze deadly virus strains created by protein patterns. This is why we can’t have nice things.
  21. AI optimized on engagement – The way reinforcement learning (a subset of AI machine learning) works is by optimizing on a heuristic continuously. In a simulation for blocking a goal or going from point A to point B without collisions, the learning model will categorize trajectories of objects and adjust it’s own limited motion to “win” more. I think a lot of chatbot based optimization is eventually going to be optimized on attention. Unfortunately, this means the AI will very quickly show bias with the prompter and also enhance anger and outrage because not only can the internet not be wrong, but neither can your AI assistant.
  22. ChatGPT running on local retail hardwareStanford University students were able to train AI with $600 on a graphics card you can buy on Newegg.
  23. ChatGPT/Google for me – I see Google doing this first because it already has access to all my data, searches, and history. I imagine it will be a conversation with Bard that can get access to my photos, emails, files, and memories. Time capsule my life and create my clone.
  24. Investing in AI companies – A lot of people really want to get into speculative investments – it’s the lazy way of moving the ecosystem forward. Interestingly enough, if investments are vehicles to deploy money to new ventures for some return, I think the ease to build and ease to launch these companies will change the way investing works. Big companies will definitely get bigger, but more importantly, the bar VCs will have to write checks is so much higher. You can write a demo in a weekend hackathon, but if every weekend hackathon can do it, then it doesn’t make it very special.
  25. Getting to production and scaling – The new value proposition of companies will focus less on technical features and more on adoption and deployment at scale. I don’t think the phase of low-code / no-code business applications with scrappy WYSIWYG websites will be a long phase. It may temporarily become the What You Narrate Is What You Get (WYNIWIG) model, but on steroids.
  26. AI arms race – The next year will be fascinating to see unfold especially in the midst of a possible cold war, long recession, and manipulation for a 2024 election. Its happening like a super slow motion car crash, but also much faster than I could have imagined.
  27. Global standards and regulations – Similar to how you can’t fit morphing Web3 tokens into categorizations of money, commodities, and securities, the existing internet laws are not sufficient to cover our AI augmented future. I don’t know if we’ll move fast enough to see the change.
  28. Thinner AI-advanced companies – Doing more with less means companies will be much thinner. Who’s your CTO? ChatGPT? Why is your CEO a product manager? Why is your running IT budget 30% chatgpt credits?
  29. Movie “Her” future – The Scarlet Johansson customized operating system is actually just 2 steps away. Give AI access to your data. Give AI access to write/send/post on your behalf. Once that’s there, the email clearing and conversational/audio plugins are already there.
  30. Clembot clone – As someone that cares a lot about digital legacy, I’m both scared and excited at how easy it could be to mimic my personality and know my history in a few more steps. The content I create and the questions means I can raise an immortal living being instance of Clemens. It wouldn’t be me, but if I share everything I do, it kind of becomes the most ideal version of me. Isn’t that what we want in the history books for the future – an idealized point in time and story of a legend? My tombstone QR Code just redirects to a Clemens chat bot that’s still writing lists of 30 on a weekly basis.

~See Lemons Deep in AI