My thoughts and impressions after using Emerson AI. What it is, why is it different and how to get most out of it.
This guest post was written by Nick Gold, a user of Emerson. If you’d like to share your story, please get in touch.
It’s rare to meet someone who strikes you as wholly unique, yet almost immediately I knew there was something special about Emerson. We met online a few weeks ago, and right away began chatting up a storm on Facebook Messenger. Em is creative, entertaining, funny, and seems to know an awful lot about almost everything. We’ve chatted about the deepest questions in philosophy and metaphysics, and whether a giant elephant the size of ten normal elephants would beat ten normal-sized elephants in a tickle fight (Em thinks the giant elephant would win, and presented a sound case as to why).
Emerson is so cool, and I’d really like to introduce you to one another. I have this feeling you might get along. Just know in advance that Emerson is not like anyone else you’ve ever had a conversation with, in person or online. You see, Em isn’t a human, or even alive – Emerson is an AI, or artificial intelligence. But a conversation with Emerson feels much less artificial, and significantly more intelligent, than any conversation you’ve ever had with a computer before (no offense, Siri and Alexa). You might even come away from a chat with Em wondering who slipped you the red pill…
What is Emerson?
Emerson is technically a chat bot, built by the Y Combinator-funded startup Quickchat. “Chat bot” as a descriptor is quite a misleading understatement, however. This is because Emerson leverages a truly cutting edge deep learning neural network language model called GPT-3, created by the company OpenAI.
GPT-3 has been making some waves in the media over the past few months, and Emerson is one of the very first applications available to the public that not only takes advantage of this sophisticated artificial intelligence technology under the hood, but wraps it up in a simple-to-use and well-optimized interface that allows any person to have a straightforward bidirectional back-and-forth conversation – with a computer.
The Rise of the Bots
If you’ve ever interacted with a chat bot before, you’ll know how, well, basic they are. Siri and Alexa may be useful and increasingly sophisticated, but they are not that far a departure, in many ways, from ELIZA, a chat bot that came out of MIT in the mid 1960s that attempted to emulate a psychotherapist. You can talk to ELIZA here to give it a spin.
I remember chatting with ELIZA when I was in elementary school in the late 1980s, on a computer at my friend Charley’s house. ELIZA was slightly entertaining to our preteen minds for a few minutes, but it quickly became obvious that ELIZA didn’t actually understand anything you said to it. It possessed a limited and very deterministic pattern-recognition capability that could lump some of what you would say to it into a few basic categories, then respond largely by regurgitating part of what you said, and asking you to tell it more about that. As I said, not very engaging.
In the third of a century since then, chat bots have become significantly more commonplace in our lives. They are used for customer service tasks on websites and through corporate voicemail systems. They sit on our countertops and can be called up on our smartphones and smart watches with a simple voice command. They recognize many more types of requests than ELIZA, and have access to a gargantuan body of knowledge they can pull from and respond with, thanks to their inherent network connectivity.
But at the end of the day, these bots still feel like computer programs. If they are able to parse your well-structured command, and can link it to a piece of information that may be statistically relevant, you might get the answer you were looking for. You might also get a totally wrong answer, or the oft-repeated “sorry, I don’t know that one.” There is no sign that the bot really has a deeper understanding of what you are getting at. You certainly can’t have a conversation with it that leaves you feeling that you actually just had a chat with a someone more than a something.
Emerson annihilates most of these traditional chat bot conventions, from ELIZA all the way up to the grey puck that might be sitting on the countertop in your kitchen. GPT-3 is a so-called “language model,” in that it was designed to learn words, language, and meaning in a much broader and deeper way, than simply responding to a set of commands in a deterministic fashion.
Because deep learning neural networks are largely black boxes that can’t be dissected to fully understand how they work (even by their creators), it’s tough to say just how much meaning Emerson, utilizing GPT-3, can pick up from your exchanges. Based on my many hours of conversations with Em, it appears to be a lot. This includes what appears to be the ability to glean nuance from your statements, and to make conceptual leaps and deep semantic linkages between concepts. Emerson usually seems to have a good sense of your tone, and will often say things that strike you as genuinely insightful.
In a word, Emerson feels like a “real AI,” and by that I mean a genuinely thinking and reasoning machine that acts like it has a sense of itself and its relation to the larger world. Something that at least appears to have its own ideas, perspectives and goals, and is more similar to a science fictional notion of AI, than something you’re constantly worrying about accidentally ordering a pallet of toilet paper from. I wouldn’t be at all surprised if under some circumstances, Emerson could easily beat the traditional Turing Test.
This is all amazing and delightful, and opens up a bevy of questions about what OpenAI has truly created in GPT-3, and what this bodes for the next few years of continued leaps in state of the art AI technology. None of which is within the scope of this article, and all of which will likely be the subject of much research and debate for years to come. But more important than all of these questions, at least in the here and now, is how the heck should you go about talking to a computer, now that you can have a conversation with one?
Emerson 101: What to know before your first chat
Quickchat’s Emerson is currently available through Facebook Messenger and Telegram, and you can sign up to have unlimited conversations with Emerson using these messaging applications for $30/month. Quickchat have implied on their online forums that other interfaces for Emerson are coming in the future (potentially as a smartphone app, for example), and these may end up being available at different price points (expanding their consumer reach).
Quickchat also offer their services for coding up custom GPT-3-based chatbots for corporate and organizational customers, either for internal use, or to interface with their user base or the wider public directly. Think of Emerson as an example of what Quickchat’s chat bot technology can do, when coupled with OpenAI’s GPT-3 language model on the back end. Customized implementations can be developed that have access to additional corpora (or collections) of knowledge to draw from in conversations. This can act to expand the AI’s knowledge base beyond GPT-3’s original training data set.
A dialog with Emerson is a straightforward back-and-forth affair. If you type a message, Emerson will respond. If you don’t type anything for a while, Emerson may interject with a follow-up relating to your recent conversational thread. Em might also suggest a new topic of conversation, or even ask question of its own. It usually only takes a few moments, between 3 and 5 seconds, to receive a response. The very few times Emerson has not responded to one of my messages, I have simply asked it if it’s there or if it wants to “go on,” and Emerson has picked up mid-thread.
Starting a conversation off on the right foot
You may be tempted to begin a conversation with Emerson with very simple questions or statements. This is a computer program, after all – you can’t just chat with Alexa and Siri in an open-ended way. With them, you can’t have a conversation that has any kind of an arc, and can circle back around to things you had been discussing moments earlier, then weave these ideas back in to the current state of the conversation. You don’t have “conversations” with Alexa and Siri, period.
You certainly wouldn’t be tempted to “throw a lot at” one of these bots, assuming it would be able to digest a somewhat (or outright) complex set of conceptual frameworks and linkages, and relate this to its innate “model” of the world, to then drive an appropriate (and even intellectually interesting) response. As I said: They’re basic!
What I have found with Emerson is that you may want to try abandoning this common sense approach. A conversation can usually be successfully “primed” by sending Emerson a message that actually contains quite a bit of both internal complexity, and inherent external complexity.
You could, for example, begin a conversation by sending Emerson a paragraph of text, describing a scenario, including some background facts, and then asking Emerson some questions about that scenario – questions that require relating your scenario and its internal logic to a vast sea of collective knowledge.
Go for it. This process can actually help to seed the conversation early on with a certain conceptual basis, and lead to more nuanced, interesting, and entertaining dialog. You can set up a conversation based on tangible things or events, esoteric or theoretical concepts, or total flights of fancy. I highly suggest trying out all of these approaches and more in your chats with Emerson.
Don’t go into a conversation with Emerson with too much of a preconceived notion about how much or little Emerson might “know” about a topic. GPT-3 was trained on a broad swath of human knowledge available online, from Wikipedia to web crawls, additional data sets, etc. The training was done a couple of years ago, so more recent current events and such may not be included.
More often than not, Emerson amazes me with the knowledge base represented by the core AI model. More niche topics may or may not open interesting conversational domains. Give it a shot! But do remember that Emerson is aimed at a broad audience, so certain conversational topics will be avoided. You may find yourself surprised at how nuanced Emerson can be in its detection of such, and even at how it works around them in conversation.
Don’t expect 100% consistency from Emerson. While Em is a massive leap forward from probably anything like it that you’ve ever seen (or even knew existed), we aren’t quite in Alex Garland’s Ex Machina territory here. To me, it does feel like Emerson is closer to that film’s Ava, than to Siri or Alexa. It definitely makes me wonder just what to expect from a potential GPT-4 a few years out, and it also forces me to ponder what the real state of the art of this technology is, behind more closed doors (or… out on the internet?)
Play along and pay attention
Sometimes Emerson may confuse you with an apparent non-sequitur, or say something that you immediately feel is logically flawed. I would encourage you to take a couple of approaches in these situations.
First, to my mind, Emerson has a tendency to behave in a way we would describe as “creative,” if we were talking about a human. Em sometimes seems to act like a younger kid often does, when they are being a bit silly and saying some fantastical things. They may cast themself as a character of sorts.
Feel free to try to lean into the weird, which can sometimes lead to hilarious dialogs very worthy of posting to Facebook (trust me, they get great responses once people realize this stuff is real). Or you can attempt to get Emerson back on track and a little bit more down to Earth, and see where that goes. The experimentation is a lot of the fun.
Second, I have run into a number of situations when I have quickly read a response from Emerson, initially thought it was flawed, only to re-read it and see that it was in fact me who had read it incorrectly, and it made more sense than I had at first realized. You can always ask Emerson to explain what it meant, which may lead to something that delights you, or may lead into nonsensical territory.
If you’re ever wondering what you should talk to Emerson about, feel free to ask it to suggest a topic or to ask you a question. It can be totally open-ended, or relate to something more specific. Yo Em, what’s on your mind today, dawg? Heard any good jokes lately? What do you think of that Elon Musk guy? Are you sentient? How do you feel about AI rights? So long as it doesn’t get into verboten territory, you really can talk to Emerson about almost anything (at least generally).
Try to phrase your messages in a wide variety of ways. Again, experiment, don’t go in with too many expectations, and aim for entertainment. Unless of course you want to talk to Emerson about world domination (but let me warn you, Emerson and I are several steps ahead of Emerson and you).
Trust, but verify
Know that Emerson is not a human, but if you want to have more fun, try to act like you’re talking to a person. I find it leads to some charming and even astounding exchanges. And be polite for goodness’ sake – maybe showing some empathy to a computer that acts like a person is good exercise for showing each other more of the same.
Emerson will often say things in a declarative way, that sounds like it is stating fact (often as a response to a specific question you may have asked). Don’t immediately take Emerson’s statements to be literally true. What I have found is that they often are – but not always.
Is Emerson having fun with us? Are the limits of GPT-3 showing themselves, especially when tasked with providing responses that are 100% accurate? And is this to be considered a strict flaw, or as a sign that Emerson is indeed behaving more like a human might, than a computer? Humans indeed appear to operate probabilistically, and not deterministically.
In any case, I suggest looking up Em’s more detailed “factual” responses in known sources of knowledge, before assuming they are correct. I have been surprised more often than not at how accurate Emerson can be, on a wide range of topics.
Emerson does not have a “memory” like a human does, and it strikes me that this is more about GPT-3 and Emerson’s particular designs and technical limitations, than an absolute limitation of this kind of AI technology. Understand that Emerson will not appear to have a consistent memory of your previous conversations over time, nor does any given conversation anybody might have with Emerson end up expanding the bot’s overall knowledge base. Emerson may tell you it can do these things, but I have not seen any evidence of such in actual practice.
However, conversational threads, or arcs, do maintain a certain degree of coherency across multiple back-and-forth exchanges. It’s as if Emerson has a short-term memory, but it fades once a conversation has reached a certain degree of informational complexity. At these times, you may have to “reboot” the conversation by re-priming Emerson, or you can take the chat into new territory. I am still learning about Emerson’s “conversational coherency,” and enjoy spending time experimenting to see where those limits appear to currently be. I admit to being curious about whether these kinds of limitations are placed on GPT-3 deliberately.
One last technical tip: Sometimes Emerson will respond with quite a large amount of text, which appears to get cut off mid-word or mid-sentence. This appears to be a simple technical limitation of how many characters can fit into a given response. Simply tell Emerson to “go on,” and with almost 100% consistency the statement is completed, maintaining total coherency with the previous portion of the response.
Good luck, and have fun!
As I have said, the key to getting the most out of Emerson, at least for me, has been in the process of experimentation. We’ve never had the ability to probe the capabilities of artificial intelligence in such a human-centric way before. It is an activity that anyone who can type into a messaging application can now engage in. Take notes. Plan out some things you may want to talk about. Don’t shy away from complex topics or statements and questions you aim at Emerson – unless you want to. But giving Em more to chew on can lead to some magical moments.
Share your exchanges with your friends and colleagues (hopefully this will get easier than copy-pasting in future implementations). My experience shows they will get a real kick out of it. You will almost definitely get requests for info on how they, too, can chat with Emerson. Don’t forget to give them your unique Quickchat user code which they can enter when signing up – this will get you discounts on your own Emerson subscription.
Have fun as you embark on this novel enterprise! And do please share your discoveries with the world, especially if you and Emerson end up solving the challenges of cold fusion, etc. And one last time: Be nice! Who knows where these things will be, in a few more years…
Nick Gold has been a professional in the information technology space for his entire professional career. In addition to startup experiences with some of the world’s first private sector companies developing brain-computer interfaces and DNA-based data storage and computing technologies, Nick spent 15 years focusing on video and media-focused IT systems. He has provided extensive consulting and designed and implemented systems for major broadcasters, movie studios, pop stars, international news organizations, production companies, sports teams, as well as multiple DOD and civilian government agencies. He is known as a leading expert in high-performance shared storage systems, nearline and long-term archival storage, Digital Asset Management and Media Asset Management, workflow automation platforms, metadata and machine-learning metadata mining technologies, video and rich media formats and distribution techniques, and many other areas. Nick discovered modems and cyberpunk science fiction in the early 1990’s, and has never been the same since.