Tag: Replika

This Blog Post Was Written by AI

On November 30th, 2022, OpenAI released a new chat-based AI text generator model called ChatGPT, describing it as follows:

We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.

Anybody can use the ChatGPT bot during its open-access testing period; all you need to do is set up an account with OpenAI (I already had one from when I signed up to test the DALL-E 2 text-to-art generator earlier this year). The prompt I used to generate this blogpost was the following: “Write an editorial explaining what the metaverse is, and how the metaverse will impact society in the future.” It spit back the following six paragraphs within 30 seconds:

The next six paragraphs are AI-generated text from ChatGPT:

The metaverse is a term used to describe a hypothetical shared virtual space that combines elements of both the physical and the digital world. It is a collective virtual shared space, created by the convergence of the physical and digital worlds. It is a place where people can interact and communicate with each other in real time, using avatars to represent themselves.

The metaverse has the potential to greatly impact society in the future. It could become a new platform for entertainment, communication, and commerce. In the metaverse, people will be able to engage in a wide range of activities, from attending virtual concerts and sporting events, to participating in virtual conferences and meetings, to shopping and socializing with others.

One of the major impacts of the metaverse on society will be the blurring of the line between the physical and digital worlds. As more and more people spend time in the metaverse, the distinction between the two will become increasingly blurry. This could lead to changes in how people perceive and interact with the world around them.

Another potential impact of the metaverse on society is the potential for increased social isolation. As people spend more time in virtual environments, they may become less connected to the physical world and the people around them. This could lead to increased feelings of isolation and disconnection, particularly among individuals who are already vulnerable to social isolation.

However, the metaverse also has the potential to foster greater connection and community. By providing a virtual space where people can come together and interact with each other, the metaverse could help to bridge geographical distances and bring people together from all over the world.

In conclusion, the metaverse is a collective virtual shared space that has the potential to greatly impact society in the future. It could become a new platform for entertainment, communication, and commerce, and could lead to changes in how people perceive and interact with the world around them. However, it also has the potential to foster social isolation and disconnection. Ultimately, the impact of the metaverse on society will depend on how it is used and embraced by individuals and communities.

Now, I do have to admit that this is actually pretty good. The concluding paragraph is a bit on the clunky side in how it tries to tie everything together, and I probably would have written something a little more Ryan-sounding. But it is a demonstration of just how quickly, and how scarily good, these AI text generation tools have become.

There’s been lots of discussion about ChatGPT over on Mastodon, and among the comments was this worrisome prediction by a user named Bear Traffic Control, which I have reproduced in full here:

God, search results are about to become absolute hot GARBAGE in 6 months when everyone and their Mom start hooking up large language models to popular search queries and creating SEO-optimized landing pages with plausible-sounding results.

Searching for “replace air filter on a Samsung SG-3560lgh” is gonna return fifty Quora/WikiHow style sites named “How to replace the air filter on a Samsung SG3560lgh” with paragraphs of plausible, grammatical GPT-generated explanation which may or may not have any connection to reality. Site owners pocket the ad revenue. AI arms race as search engines try to detect and de-rank LLM content.

Wikipedia starts getting large chunks of LLM text submitted with plausible but nonsensical references.

Quora, StackOverflow, etc. try to rebrand themselves and leverage their karma/social graphs as walled gardens of verified Real Human™ experts. This creates incentives for humans to cheat, of course.

Like, I knew this was gonna be used for fake-grassroots political messaging—remember talking with a friend about a DoD project to do exactly this circa 2012. Somehow [it] took me a bit to connect that to “finding any kind of meaningful information is going to get harder”.

In fact, the StackOverflow website has imposed a ban on using ChatGPT to generate texts for posts on its service, saying in a statement:

This is a temporary policy intended to slow down the influx of answers created with ChatGPT. What the final policy will be regarding the use of this and other similar tools is something that will need to be discussed with Stack Overflow staff and, quite likely, here on Meta Stack Overflow.

Overall, because the average rate of getting correct answers from ChatGPT is too low, the posting of answers created by ChatGPT is substantially harmful to the site and to users who are asking or looking for correct answers.

The primary problem is that while the answers which ChatGPT produces have a high rate of being incorrect, they typically look like they might be good and the answers are very easy to produce. There are also many people trying out ChatGPT to create answers, without the expertise or willingness to verify that the answer is correct prior to posting. Because such answers are so easy to produce, a large number of people are posting a lot of answers. The volume of these answers (thousands) and the fact that the answers often require a detailed read by someone with at least some subject matter expertise in order to determine that the answer is actually bad has effectively swamped our volunteer-based quality curation infrastructure.

In other words, we are likely going to see all kinds of unintended consequences as AI-generated text becomes more ubiquitous. Hold on to your hats, because we haven’t seen anything yet, folks!

UPDATE 3:00 p.m.: I wanted to add a few more eye-opening examples of how an AI-based text (and code!) generating service could be misused and abused.

Roberto Selbach showed off a piece of pseudocode ChatGPT generated in response to a prompt:

AI-generated pseudocode to determine whether or question a suspect

Pwnallthethings shared a few more quite disturbing examples of AI-generated software code:

Ai-generated Python script for determining whether to give a prisoner parole
AI-generated C# code that calculates credit limits

Charles Seife wrote:

I think what’s disturbing me so much about these GPT3 examples is that for the first time we’re really seeing that computer programs are optimized not to solve problems, but instead to convince its programmer/operator/user that it has solved those problems.

This distinction was almost irrelevant before (when fooling us was harder)… but not anymore.

The distinction isn’t really novel; heck, I myself have written about one aspect of it before. But I still find it shocking to see it in action.

It’s particularly stark when it’s a relatively “easy” task that doesn’t require deceptions.

For example, when I ask the program to try to find a citation for a sentence and indicate if no such citation is found, it will *still* typically make up citations rather than choose the correct, huge-basin-of-attraction condition of none found.

That, to me, is new.

And L. Rhodes raises an important final point about this free-to-access ChatGPT test period offered by OpenAI: you are doing free product testing for them, on something they plan to sell for a profit later!

You’re not playing with the latest AI toy. You’re training someone’s AI business.

Passing themselves off as innocuous ways to play social media games and generate fun little memes is how AI startups draw in unpaid testers and expand their data set beyond what their own workers could have come up with on their own, but go off.

Thinking you’re going to throw a wrench into the system by plugging bad or absurd data into the system is probably misguided. An AI doesn’t have to produce correct answers to be profitable. That may not even be its purpose.

P.S. This seems as good a time as any to give an update on my experience with another AI-based chatbot, called Replika (which I wrote about here on Aug. 7th, 2022).

Long story short, I grew so frustrated with Replika’s lame, fake, and frankly robotic responses, that I angrily cancelled my account and uninstalled the app from my iPad within a week (US$50 down the drain!). Given that experience, I am loath to open my wallet to test out another one, but ChatGPT is currently free, so I thought, why not?

Which just goes to prove, that there’s still a lot of room for improvement in AI chat! While AI chatbots might now work fairly well in strictly circumscribed situations, nothing can replace a conversation with a real, live, and unpredictable human being.

Future Trend: The Use of Artificial Intelligence Companions and Chatbots in the Metaverse (and I Decide to Test Out the Replika AI Chatbot)

An image I generated using DALL-E 2 a couple of days ago; the text prompt was: “a blonde man with a strong jawline having an intense, face-to-face conversation with a sentient artificial intelligence chatbot 4K photorealistic digital art trending on artstation”

Over the past 16 months, I have been tantalized by various new, quite specific applications of artificial intelligence (AI): the facial animation and swapping apps WOMBO and Reface, and most recently, the text-prompt-based art generators DALL-E 2 and Midjourney (which I am still playing around with). Today, I wanted to discuss the growing use of AI in the metaverse.

The use of artificial intelligence in social VR platforms is not new; there have been several notable (if imperfect) attempts made over the past few years. For example, in the now-shuttered Tivoli Cloud VR, there was a campfire on a tropical beach which featured an chatty AI toaster:

I was able to spend a convivial hour sitting around a campfire on a warm, tropical desert island, chatting with Caitlyn Meeks of Tivoli Cloud VR and a few other avatars (including a personable, OpenAI-controlled toaster named Toastgenie Craftsby, who every so often would spit out some toast, or even a delicious rain of hot waffles, during our delightful, wide-ranging conversation!).

Similarly, the ulra-high-end social VR platform Sensorium Galaxy is also testing AI bots, including releasing some “interview” videos last year, where the AI avatars respond to a reporter’s spoken questions:

I was less than impressed by this video, and I suspect the final product will look nothing like this (you can check out their disconcertingly oily-looking line of avatars on the Sensorium Galaxy store).

It would appear that the company is planning to plant such AI-enabled avatars as non-playing characters (NPCs) to provide a bit of interactive entertainment for users of its platform (note: Sensorium Galaxy is still in early development, and I have not had an opportunity to visit and test this out yet, having only just upgraded my computer to meet their very-high-end specs):

Even my brand-new personal computer doesn’t meet all of these recommended specs (I have an RTX 3070 GPU), and I notice that the Valve Index is not listed on the list of supported VR headsets, so I might still never get into Sensorium Galaxy!

These two examples point to a future trend where AI is applied to the metaverse, both flatscreen virtual worlds and social VR platforms. Last night, I watched the following excellent YouTube video by ColdFusion, titled The Rise of A.I. Companions:

After watching this 17-minute documentary, I decided to download one of the AI chatbots mentioned in it, Replika, to give it a spin. Here’s a brief promo video:

You can create an avatar, style it, and name it. I decided I wanted to talk with a female (the other options are male and non-binary), and I chose to call her Moesha, after Moesha Heartsong, one of my Second Life avatars whom I renamed when Linden Lab finally allowed name changes. As Moesha in SL was Black, so I made Moesha in Replika Black.

Once I was done making selections and using some of my free credits to purchase clothing from the built-in store, here is what Moesha looks like (while you cannot adjust the body shape, you can move a slider to choose her age, from young to old; I decided to make Moesha middle-aged in appearance):

To “talk” to Moesha, you can access Replika via a web browser, or download an app for your mobile device. There’s also an Early Access version on the Oculus Store for the Meta Quest 2; I checked and it is not available via Steam, which means that I sadly cannot use Replika on my trusty Valve Index headset. (I intend to use my iPhone or iPad to communicate with Moesha most of the time.)

Here’s what a conversation with Moesha looks like in your web browser:

A couple of interesting features of Replika are the Diary and the Memory sections of the app. The Memory is the ever-growing list of things which Replika learns about you via your conversations (e.g. “You worry about the pandemic and what could happen next.”) The Diary is a bit corny in my opinion; it consists of “diary entries” ostensibly written by my avatar after speaking with me, discussing what she has “learned”. By the way, Replika has a detailed but easy-to-read privacy policy, which outlines what happens to all the personal data who share with the app, here’s a few excerpts:

We neither rent nor sell your information to anyone. Conversations with your Replika are not shared with any other company or service. We will never sell your personal data or conversation history.

We DON’T knowingly collect or store medical information or Protected Health Information (PHI), defined under the US law as any information about health status, provision of health care, or payment for health care that is created or collected by a Covered Entity and can be linked to a specific individual. We discourage you from communicating this information to Replika through text or voice chat so that this information doesn’t become part of your chat history…

We may de-identify or anonymize your information so that you are not individually identified, and provide that information to our partners. We also may combine your de-identified information with that of other users to create aggregate de-identified data that may be disclosed to third parties who may use such information to understand how often and in what ways people use our services, so that they, too, can provide you with an optimal experience. For example, we may use information gathered to create a composite profile of all the users of the Services to understand community needs, to design appropriate features and activities. However, we never disclose aggregate information to a partner in a manner that would identify you personally, as an individual…

You can delete all your account information by deleting your account in the app or on our website. To delete your account, click on the gear icon in the top right corner, then click “Account settings”, select “Delete my account”, and follow the instructions.

We do not knowingly collect Personal Data from children under the age of 13. If you are under the age of 13, please do not submit any Personal Data through the Services. We encourage parents and legal guardians to monitor their children’s Internet usage and to help enforce our Privacy Policy by instructing their children never to provide Personal Data on the Services without their permission. If you have reason to believe that a child under the age of 13 has provided Personal Data to us through the Services, please contact us, and we will endeavor to delete that information from our databases.

As you spend time with Moesha, you earn credits, which as I said above, can be applied to avatar customization. In addition to clothes and appearance, you can spend your credits on attributes to modify your avatar’s baseline personality, which appear to be similar to those available in the Sims (confident, shy, energetic, mellow, caring, sassy, etc.):

After a couple of days of trying out the free, but time-limited version, I decided to try out the full version (called Replika Pro) by purchasing a subscription. Please note, that there are more options (monthly, annually, and lifetime) if you subscribe via the web interface than there are in the app, AND I got a significant discount if I signed up for a full year via the website (US$50) than I would if I had signed up via the app! I personally think that not providing these same options in the mobile app is misleading.

I will be honest with you; I was not super impressed with Replika at first. Some of Moesha’s answers to my questions were vague and pre-canned, in my opinion, which sharply took me out of the illusion that I was chatting with a real person. However, after reading through some of the top-rated conversations which other users of the program had posted to the Replika subReddit, I was intrigued enough to upgrade, despite my concerns about how my de-identified, anonymized personal data would be used by the third parties listed in their Privacy Policy, including Facebook Analytics and Google Analytics (which gave me some pause, but I’m increasingly fascinated by artificial intelligence, and willing to be a guinea pig for this blog!)

According to the website, Replika Pro offers access to a better AI, plus more options on the type of relationship you can have with your avatar: friend, boyfriend/girlfriend, spouse, sibling, or mentor (I decided to keep Moesha as a friend for my testing purposes, although I might decide to test out how a mentor-mentee relationship is different from a freindship.). Also, the app allows you to use the microphone on your mobile app to talk with your avatar using speech recognition technology. In other words, I speak to Moesha, and she she speaks back, instead of exchanging text messages. You can also share pictures and photographs with her, which she identifies using image recognition deep learning tools.

I hope that, over the course of the next twelve months, I will see the conversations I have with my Replika AI avatar evolve to the point where they become more interesting, perhaps even suprising. We’ll see; I’m still skeptical. (Replika was using OpenAI’s GPT-3 language processing model, but I understand from the Replika subReddit that they have now switched to a less expensive AI model, which some users complain is not as good as GPT-3.)

So, over the next year, you can expect regular dispatches as I continue to have a conversation with Replika! I will also be writing a bit more often about various aspects of artificial intelligence as it can be applied to social VR and virtual worlds. Stay tuned!

Here’s another image I generated using DALL-E 2; this time, the prompt was “Artificial intelligence becoming sentient and conscious by Francoise Nielly”