Alex Panait on Current Trends and Possible Futures for AI

I had a thought-provoking conversation with Alex Panait of Mission Automate on current trends and possible futures for AI applications.

Alex Panait on Current Trends and Possible Futures for AI

I first met Alex Panait of Mission Automate at the last Bootstrappers Breakfast of 2023. It was an online event and I came away impressed that he was a serious entrepreneur with a strong team. He next dropped by an in-person breakfast at Voyager Coffee in San Jose in early 2024, and we ultimately invited him to offer a briefing on “AI in Action” at Lean Culture. He offered a thoughtful and practical briefing, clearly informed by his direct experience using AI tools and research into emerging capabilities.

We stayed in touch, and in early September, I asked him for 20 minutes of his time to get his assessment of an exciting talk on AI agent accuracy by Mehak Aggarwal. She explained how her team relied on a mix of tools to complement LLM-based models to improve outcomes. A week later, I added an article by Shelly Palmer on “Four Days with OpenAI’s o1 models” to our agenda. Palmer had concluded that while o1 was not ready for prime time yet, they were likely to represent a new class of AI applications that can reason through complex problems using “chain of thought” techniques. He advised his readers to start experimenting now, given the rapid evolution in the application category.

We had a great conversation that lasted a little over an hour and Alex suggested we reconvene and do a podcast on “Current trends and possible futures for AI.” This post is an edited transcript of that conversation.

Edited Transcript for Current Trends and Possible Futures for AI

What follows is a a transcript of an hour long conversation, edited for clarity with hyperlinks added.

Sean Murphy: Hey, it’s Sean Murphy. I’m sitting here today talking with Alex Panait of Mission Automate. We’re going to explore what Mission Automate does and then look at what we think is coming in AI, the impact of AI on our lives, and the future of work.

Alex Panait: That sounds great. Thank you, Sean. It’s great to be here.

Sean: Can you explain what Mission Automate does for people who are unfamiliar with it?

Alex: We are an applied AI R&D team. Our clients come to us with AI challenges: we survey the market to identify the best solutions for their needs. Sometimes, we take pieces that are off the shelf and connect them. Other times, we craft creative solutions at the back end, which may even mean developing custom algorithms or training custom models.

Sean: So you are broadly focused on artificial intelligence applications. You help startups and more established firms develop, enhance, or extend products they’re working on with elements or components of artificial intelligence technology.

Alex: Yes. We are an enabler of AI-related building blocks and applications. We research solutions in the market and help our customers understand the technologies better. We also develop tools for specific use cases or even products.

Sean: AI has attracted an enormous amount of investment in the last few years, driving significant innovation. I would assume that research is a substantial component of the value that you offer your clients.

Alex: Certainly, for many of them. The way that I see it, AI has become pretty expensive even as more and more companies are interested in adopting it. Use has expanded so quickly that, in many cases, companies are interested in adopting it without completely understanding how it can help them. This approach can lead to a bit of unease, which is one reason they approach us. We engage and understand their needs. We help them avoid problems and suggest how AI technology can help them meet their specific business needs.

What Developments in AI are the Most Interesting?

Sean: What developments in AI are you most interested in or most excited about?

Alex: I have two areas of particular interest where I see AI development heading.

First, big-name services like ChatGPT, Gemini, and Claude, which are widely used by both individuals and businesses, are in a race to stay competitive. They’re investing heavily in new features. For example, ChatGPT recently released an updated model that seems to be aimed at surpassing Claude, which used to have the largest context capacity. Now, ChatGPT has an even larger context, meaning it’s better at understanding and reasoning about user queries. This focus on expanding context size is a significant focus for development in companies like Anthropic, Google, and OpenAI.

The second direction AI is moving toward is the creation of smaller models. These are designed to pack as much capability as possible into a more compact form. Smaller models are faster and cheaper to run, and many users don’t always need the most advanced, high-context models. Smaller models are especially valuable for high-volume applications that don’t require complex reasoning. What’s interesting about smaller models is that they are more accessible for open-source communities because they are easier to work with, unlike the larger, cutting-edge models developed by big companies, which are too costly and complex for open-source teams to keep up with.It’s

Sean: When you mention open-source, I have read that Facebook is taking an open-source approach but developing some big models

Alex: Yes, Facebook has been very friendly to the open-source side. Because they entered the market later than ChatGPT, their LLaMA offering has had a harder time becoming popular with the end users, who often think of ChatGPT, Gemini, or perhaps Claude. Google, due to their size, has been able to invest a tremendous amount into the marketing of their Gemini product. Claude is less well-known, but it is still a fantastic technology.

Possible Futures for AI: Technology Singularity, AI Winter, or Muddling Through?

Sean: It’s been hard for me to get too excited about AI because of these periodic waves of enthusiasm that promised breakthroughs and failed to deliver.

I studied computer science in the late 1970s and took some AI courses because my brighter friends were very excited about it. I wandered off into CAD for semiconductors, missing what my friends later called AI Winter.

In 1993, Vernor Vinge wrote “The Coming Technological Singularity.” He predicted the emergence of superhuman intelligence that would mark the end of the “human era” within 30 years. If he is correct, then right about now, our world is going to become unrecognizable. Hopefully, our new AI overlords will at least keep us as pets or put us in their equivalent of ant farms. While I know some very bright people who are worried about that possibility, it strikes me as very unlikely.

There is also a common complaint among AI researchers that whenever a new AI application starts to work, it’s no longer called AI–reserving AI for applications that don’t work–yet.

I worked in semiconductors and electronic systems for two decades and witnessed chip functionality increase by at least five orders of magnitude without obsoleting humanity. I think we will still keep people in the loop, which means the challenge is learning what tasks to delegate and how to do it effectively. The developments I am interested in are intelligence amplification or augmenting human intellect.

Alex: I agree that it’s highly unlikely that AI will turn humans into slaves and pets.

I think a more likely dystopian future would involve AI being integrated into hardware systems that cause damage. Not because someone intentionally caused harm but because AI is still somewhat unpredictable. This unpredictability could lead to issues since AI is trained on human knowledge, and humans are often unpredictable compared to perfect logic and rationality. For instance, an AI managing the global power grid could misinterpret a situation and shut down power to save energy. While unlikely, this is the kind of concern that AI safety advocates focus on.

A more realistic scenario, which I think is most likely, is that AI will become a commodity controlled by large organizations. This is already starting to happen: only a few major companies have access to the most advanced AI technology and decide how it’s distributed. Smaller companies can still get funding and innovate or get funding, but the most powerful AI will likely remain in the hands of a few. This scenario seems likely because training advanced AI models now costs around a billion dollars each iteration, which only the largest companies can afford.

Finally, there’s the optimistic or “utopian” scenario, which is more likely than the dystopian one but still not super likely. This would involve a technological singularity, where AI can improve itself and, with or without human help, lead to significant advancements in areas like food production, medicine, energy, and transportation. This would make life more affordable and accessible for everyone, improving access to basic needs.

Sean: The long-term trend since the Industrial Revolution, arguably the last “Singularity” we have experienced, is human flourishing.

I think the challenges AI presents are not intrinsic to AI but to the “crooked timber of humanity”–human frailty, ignorance, our propensity to selfishness, and our capacity for evil. It’s how we use this technology as individuals and societies. If we look at Norwegians, Brazilians, and Chinese–to pick three disparate cultures–we will likely see three different kinds of applications. Norwegians will probably construct some very interesting saunas and do an even better job of taking care of one another; the Brazilians will become even more amazing at soccer; the Chinese will have even more effective open-air prisons for the subpopulations they want to suppress. The outcomes are tied more to individual and cultural values and choices.

It’s hard to recapture what it was like before cell phones and Google, which arguably has, at least in the United States, tremendously transformed the quality of our lives. Now, the other side of it is that cell phones also look a little bit like carrying a digital heroin dispensary in your pocket., making it something of a mixed blessing.

It’s interesting that in terms of electric power, for example, the cost to build an electric power plant or nuclear power plant is the order of magnitude, a billion dollars or more, right? And yet, all that means is I have power available to run all of the appliances in my house for perhaps 15 to 30 cents an hour, depending upon what I need.

We’ve been able to decouple the need for massive investment with some kind of complete control of my appliance usage by the power companies.

AI Technology Will Likely Diffuse Beyond the Control of a Few Entities

Alex: I take your point. I think that, ultimately, the technology will no longer be under the control of a few entities. That’s what usually happens with advances; they diffuse.

Sean: The Russian military, US military, and Chinese military will hold access back to a few technologies that you and I are not going to have in our homes or offices. Some separations will be maintained, but I hope it’s not the corporate stratification you outlined.

Alex: I agree that such a possibility is more related to culture than technology. As a society, I think we have the responsibility of deciding how we will conduct ourselves in the values we promote. I think AI is a lot more applicable and accessible than other types of advanced technology, such as military technology.

Sean: I think your point about accessibility is a good one. I have seen information access advance from author, title, and subject card indexes in library book catalogs to menus and files in directory tree structures in computers to keyword-based search tools for file systems and web pages. AI now enables asking questions and getting answers in a conversation with a computer.

Alex: Conversational interaction is certainly one aspect of user accessibility. Another aspect that I think is important is creative accessibility, which is the ease of using AI as a building block. We’ve reached the point where the average user can ask ChatGPT questions, and developers and product creators can build their own tools and products using AI technology.

They can’t make another ChatGPT because that’s not accessible, but they can make offshoots: AI assistants focused on specific use cases. And I think that’s going to continue to stay that way. The risk when you have very advanced technology controlled by a small number of people is that other people won’t be able to use it to transform it into something else.

For example, let’s say that ChatGPT, Google, or another major player in control of AI technology creates a model so advanced that it is completely indistinguishable from a human being. And it’s also instantaneous. It can answer any question instantaneously, access the latest data, and do Google searches to understand the difference between relevant and up-to-date information and what’s not. It understands hallucinations and can prevent them.

So, we are talking about a perfectly benevolent and all-knowing assistant. The company that created this model could say you can access it for free or at minimal cost, but only through our app. You cannot recreate anything with it. It hasn’t happened yet, and I’m not saying it will happen. However, a privately owned or publicly traded company may create such a technology and restrict access to approved websites or apps.

There is a real risk that this very advanced technology is controlled by a small number of very large players. I don’t think it’s a huge risk because I think there will still be alternatives, that companies will have an open policy because that’s another revenue source, right? They can license the API for that product and let other people use it as a backend.

Sean: It’s interesting to compare AI applications to how tools for websites evolved. Web pages were open-source; you could look at the HTML source and see how someone else built an interesting page. There was a mix of open source and COTS tools that triggered a massive transformation.

Can we talk next about the challenges that your team faces in navigating this rapidly evolving AE landscape? It’s full of promise, but it’s also unlocking other challenges. What are you wrestling with?

The Biggest Challenge for End Users of AI Tools Remains Hallucinations

Alex: Our biggest challenge as end users of AI tools is still hallucinations. They are not always easy to notice because they are very believable. We recently asked Gemini about a category of AI advancements. The list it came back with was very detailed and believable, but when we investigated them, most of them were made up.

Another challenge is inaccuracies or mistakes. For example, we use Read.ai to record meetings and provide transcripts and summaries–it’s pretty fantastic. It also provides action items as part of the summary. However, I’ve noticed that in meetings where we’re just exchanging ideas, Read AI will still come up with some action items based on the conversation, even though there aren’t any action items, right? But it’s still very capable: sometimes we converse in Romania, and Read.ai automatically translates it and gives us a summary in English.

Sean: I agree with you on hallucinations. We’ve tried two different chatbot implementations of chatbots for our website, and we had to take them both down because the answers did not always match the content on our site.

When I look at our business, I am concerned that AI will dramatically lower the cost of content generation. I think we’ll see an order of magnitude or maybe even two orders of magnitude more content on the web. So there’s a challenge: how do we stand out in the face of a flood of content?

There are two additional wrinkles: what does SEO look like when a keyword-based search is replaced by AI tools answering questions, and how do we deal with AI-enabled plagiarism?

Some of our content is starting to get referral traffic from Perplexity; we assume this is coming from reference links provided with an answer to provide more context. We are unsure how to make this happen more often; what is the equivalent of SEO best practices for AI answer tools?

A few weeks ago, I published an article on LinkedIn that was plagiarized two hours after I posted it. They left a comment that they had written about the same topic. I had laid out a five-step approach to a particular problem, and they recycled the description of the five steps and changed the domain slightly. I can see how they wrote a prompt, “Take this five-step approach and apply it to this problem.” I see more “same-day plagiarism” cropping up in the next few years.

You mentioned Read.ai, which triggers a final observation. In addition to adding summaries and action item detection and trying to extract other useful insights from listening in to the conversation, the transcription services are now enabling folks to send their “notetaker” to meetings they don’t attend. I think this is a terrible idea on many levels; it’s a surveillance model that disempowers the team that is meeting.

I would like a service that pipes the transcript–with a few seconds of delay–into a Google Doc or other shared edit tool that lets the team pull from it as part of a shared note-taking model. We do a lot of shared note-taking in our meetings. Shared note-taking makes brainstorming more effective, at least in a small team with high trust. It would be like adding a court reporter or stenographer who captures people’s words so they can be refined.

How to Present and Manage a Spectrum of Perspectives or Answers

Alex: I understand your chatbot challenges. You cannot control how they train the models and what biases the models have. I agree that there is a tremendous amount of new content out there, but it’s not clear whether it’s accurate or meaningful. I think we will see more footnotes with links and other approaches to offering independent substantiation of claims in an answer.

It would be useful to have a search tool like Google or Bing that automatically filters for the best quality content, which it should be doing already.

Sean: The Ground News site allows you to pick an issue, event, or news article and get other perspectives. They explain how it’s being reported in various news entities that lean various ways. I would like to see more sites like this. From an early age, we learn how to listen to different arguments and sort out what to believe. As adults, we learn to keep an open mind and update our understanding based on new information.

What AI tools have you found most helpful or easy to adopt and deploy?

Alex: I think Read.AI is a quintessential example of an extremely useful tool that’s easy to adopt because all you do is sign up for it, configure it to get added to Zoom, and it just works. If you don’t want it to record, you just kick it out of the meeting, and that’s it. That’s one clear example.

I have been using Claude a lot recently. One reason is a feature called Artifacts. We published a video today where I’m having a conversation with one of my team members and showing her artifacts on Claude. The main idea here is that responses from an LLM are quite verbose. In a longer conversation, you create a very large conversation history, which requires a lot of scrolling to understand. So Claude added the concept of artifacts, documents that capture responses that are embedded as icons on the left and right side of the main conversation. You can have versions for those artifacts. So if the response is almost what you want but not quite, you can say, update this artifact to do blah, blah, blah. And it’ll give you another version, which is very useful, especially in more extended, more involved conversations.

A third very useful tool is Gamma, which creates presentations. I showcased it in the webinar I did for Lean Culture.

Sean: Has your team experimented with tools for software development like GitHub Copilot?

Alex: Yeah, I do use GitHub Copilot. I find it very useful. Recently, I’ve been using it a lot because it can generate not only code but also entire files. So you can say, write all the database models, and it will generate a complete set. You can open the chat interface and ask, “if I wanted to do X, how should I organize my project?” It will explain exactly what files to create and what to put in them. So it doesn’t help you just with just the code generation but also with the project organization.

LLM Tools useful for Brainstorming, But Don’t Run Open Loop To Customer

Sean: We use Grammarly for most of our writing. It’s now so vanilla that we no longer think of it as AI–it passes the test: it’s not AI because it’s useful. I’ve been impressed with Perplexity because it did two things early that every AI app will do: providing footnotes with links to provide context and suggesting natural follow-on questions. The latter is handy for exploring new topics or concepts.

We do a lot with ChatGPT. We treat it as a source of common knowledge. You won’t get original ideas, but it will suggest ideas we have overlooked or needed to familiarize ourselves with. You can ask for ten topics an article on “X” should cover: seven or eight are ones we’ve already covered, or they’re nonsense. But we give serious consideration to two or three. It’s organized common sense.

Alex: We have found it very useful as a brainstorming partner as well.

Sean: That’s a good metaphor.

Much of my work in the 80s and 90s involved supporting electronic design automation tools. There was a co-evolution between the tools and the systems that could be designed and an ongoing tension between what a system designer could delegate to a tool and what they needed detailed control over. As systems became more complex, designers had to focus on higher levels of architecture or abstraction and, for the most part, let go of fine-grain control. However, they would rely on multiple tools in a sequence to create a specification, verify it, generate an implementation, and then verify and validate it against the specification and other well-understood requirements.

I think there may be some analogies to your use of AI tools. What do you view as the higher-level decisions that define the requirements or specifications, what do you delegate to the tools, and how do you verify the outputs or results?

For example, you are not going to generate a proposal for a $100,000 project and mail it to a prospect without reviewing it in detail.

Alex: No, of course not. We take professional communication, written and verbal, very seriously. Whenever I write an email, I take my time and review it.

I will also use two different LLMs and have them generate options that help me clarify what I want to express. I will take sentences that say exactly what I want to say and splice them together like a movie. I have not delegated the role of PR representative as much as I’ve delegated the legwork of a PR representative assistant. I can do my own PR with help from some assistants who are doing the legwork.

Another example is this one from marketing. We have a team member who handles marketing. When we need an image for one of our LinkedIn posts, she is quite adept at using Gemini to make some very good-looking and original images. She avoids many of the common pitfalls people run into when they use an image generator. Hers avoid that “AI look” and are a bit more original.

One thing I’ve done recently for one of the groups I work with needed an introductory music sting for their “future of work” podcast. I used an AI tool called Suno that generates music to create something they liked.

Generating multimedia requires giving a tool the right prompt, which is very important, but having the patience to regenerate new versions until you get something that looks decent or sounds decent. Too many people give a prompt, look at an unsatisfactory output, and assume that’s the best the tool can do. It’s really more like a roll of the dice. If you didn’t get what you wanted, try a few more times.

Blending Deterministic Tools with LLM-based Applications

Sean: I have not done much with image generation, but one of my partners uses generators quite often. Her frustration is that everything looks great except for one area. She would like to circle the problem and provide some guidance on how to improve it. She can do that when working with a freelance artist, but there does not seem to be an easy way to explain specific shortcomings to an image generator.

Alex: I don’t think we’re quite there yet. Some open-source tools will let you chain different models together and achieve the desired results. But then, instead of just giving it the prompt, you actually put together all the different components. For example, one generates text, and another generates an image in a manga style. You can chain those together to get exactly the kind of image that you’re looking for. But that requires a lot more technical sophistication than writing a prompt.

One challenge when dealing with generative models is that they don’t really understand what the image contains–this is also true for text generators. They don’t have the ability to change components of their output.

Sean: It’s interesting that in your example of taking two different streams of output from two different LLM models because the output is text if it’s 93% of what you want, you can manually edit it to get to the finish line. You don’t keep re-prompting or regenerating; you just do a few edits, and you are done.

Alex: Yeah, it’s more straightforward with text because it is easy to edit, right? So there’s that. Strictly speaking, somebody with Photoshop skills could take a generated image with one area that needs to be changed and either edit it or generate a new image just for that area and composite them together.

Possible Futures for AI and Co-Evolution with Work Practices

Sean: Our final topic: what do you foresee as the impact of AI on people’s lives and the future of work? What do you see coming? How do we get ready?

Alex: I think prompt engineering is the most important skill for anybody in any career to learn right now. Prompt engineering is a new discipline created by the Chat GPT input format. It requires you to understand how to describe what you want very specifically.

People are flexible, and when you tell someone what you want, they often understand it. They can glean it from whatever words you use. LLMs don’t have that power. You have to tell them precisely what you want, right? So it’s a conversational skill to some degree, a communication skill, to be accurate and detailed.

The other aspect of prompt engineering is understanding how the words in your description will be interpreted by the specific tool you’re using. Now, there are strategies here that will work for any tool. However, certain tools have specific responses to specific problems. For example, if you’re generating images, it’s helpful to describe the style of image you’re looking for.

Sometimes, you can describe the style of the image in terms of a particular artist: I’d like an image that looks like it’s been painted by Picasso. You have arbitrary degrees of precision in your description, which you can use as a technique for constructing or engineering your prompt.

Sean: My understanding of your description of prompt engineering is that we will have to become very good at effective delegation to these generation technologies: asking the right questions and framing the outputs we want.

However, when you say people are flexible, many of my clients wrestle with the challenges of effective delegation. One blind spot is the “I’ll know it when I see it,” which matches your iterative approach to finding an acceptable sting. To be clear, “I’ll know it when I see it,” is reasonable when you are in an exploratory mode or working on more artistic outcomes. But it’s challenging to delegate to algorithms or people.

I like your framing of prompt engineering as “understanding how the words in your description will be interpreted by the specific tool you’re using.” Too often, I see this “book of magic spells approach” where you can buy my magic incantations and use them without understanding why. Given how fast AI tools continue to evolve, I suspect this will lead to tears in the medium term.

Alex: I agree that there is a lot of overlap between delegating to a person and delegating to a large language model. I think becoming good at prompt engineering will also help you become better at collaborating with other people. I want to be clear: when I say prompt engineering, it’s not enough to find a list of prompts and use them. Good examples are helpful, just as when you are learning to write.

But the ultimate goal you want to reach is to start writing your own prompts as soon as possible. That’s what I mean by prompt engineering. It’s not enough just to find the right prompt and reuse it. You can do that in the beginning as you’re getting used to an AI tool as part of the learning process.

Ultimately, you want to know how to write your own prompts for two reasons. No book of spells, no matter how extensive, will cover all of your needs. And tools change. Your old prompts may work differently in new versions. It’s the same challenge you face when a coworker or employee moves on. Suddenly, you are working with someone else and need to make adjustments because the way you are used to does not work as well.

Final Remarks on Possible Futures for AI

Sean: Any final remarks, observations, insights, or bold predictions?

Alex: We are still in the Wild West of AI–or the Gold Rush. There are few rules and less enforcement, which enables a maximum amount of creativity.

Over time, we will have more regulation and cultural and ethical constraints as we understand the negative impacts of specific actions. And probably most importantly, we’re going to start defining best practices. We will do fewer things in fewer ways because we will coalesce around the best ways of doing things.

But we should not be scared of AI; it’s fundamentally safe, and you should be experimenting and exploring with it. Those are my final words.

Sean: thanks for taking part, I have found it very thought provoking and learned a lot!

Alex: Thanks. Thanks a lot, Sean. Pleasure talking to you.

M	T	W	T	F	S	S
	1	2	3	4	5	6
7	8	9	10	11	12	13
14	15	16	17	18	19	20
21	22	23	24	25	26	27
28	29	30	31

Alex Panait on Current Trends and Possible Futures for AI