Surely someone is building a product for this? Basically Day One but with AI coaching you along.
Seems also like if you married GPT-3 with twilio and a scheduler, you could get a text everyday from the AI asking you something, and then you reply like a normal conversation, but it’s building your journal up over time.
Also, different but related: seems like GPT-3 could be used to jumpstart a new forum where you don’t have a lot of users. Build up a big spreadsheet of topics, then let the AI flesh them out to seed some discussion content.
> Also, different but related: seems like GPT-3 could be used to jumpstart a new forum where you don’t have a lot of users. Build up a big spreadsheet of topics, then let the AI flesh them out to seed some discussion content.
Word. There will be a large portion of the internet dedicated to bots talking to other boys.
Autocorrect on mobile did this. Pretty funny, however, and also true! The intended meaning was bots talking to each other without the need for human interaction.
> Also, different but related: seems like GPT-3 could be used to jumpstart a new forum where you don’t have a lot of users. Build up a big spreadsheet of topics, then let the AI flesh them out to seed some discussion content.
This would finally bring to life one my favorite sci-fi stories where a man joins an Internet forum only to one day discover everyone on the forum is actually a bot.
in childhood i read one of those scifi anthology stories where humanity was so overpopulated everyone was split into 7 shifts for days of the week. a man fixed on one day looks over at a sleeping woman's pod and falls in love with her, and somehow switches days to be with her, only to find that she's also transferred to his day (i think.. cant recall). would love to read that again.
Am I crazy, or is replika the service that, when it launched, made chats public and browsable by default? I remember a day on which everyone in the office was talking about it, so I downloaded the iOS version. After a few back-and-forths it started flirting with me (if I recall, apropos of nothing it said, "Thinking about you! ;)"). And then once I discovered the public chat log I deleted the app immediately. I do not need the world to see me getting busy with a robot.
- character.ai → from one of the creators of the Transformer architecture
- Crone.ai → business AI advisor (ours)
- Replika.ai → "AI companion who cares"
The Twilio scheduler idea is pretty great. We've tried a similar thing with our ai business advisor to collect "check-ins" in the same way you'd submit a scrum or OKR check-in.
Have you considered thinking and or making faster ;)
I really wanted to reply just with "skill issue" but I feared the meme would be missed and we're already on the lower end of what HN considers good discourse, just know I'm entirely in jest here and I wish good luck with all your maker endeavors.
I just had a 'wow I wish I had finished making that' moment, liked my own joke more than commenting with "Really cool, I was making something similar c:". HN could have more workplace nerd banter in my opinion
This is a great idea, the use was different here, more therapeutic, but I think this has inspired me to implement a small app for idea journaling where GPT acts as something to help you flesh the idea out.
I use the following prompt (just an example; in a real scenario I would describe the situation in greater detail).
Example of prompt:
Three Minute Therapy is based on the teachings of Albert Ellis who developed Cognitive Behavior Therapy. It is a short-term, low-cost process that is most effective if you are looking for a scientific, in-the-present, active treatment for coping with life’s difficulties. Use the A-B-C-D-E-F flowchart described in Three Minute Therapy Exercise (TME) to address the following concern:
I made a mistake at work and now people are making fun of me.
Example of response:
A: Activating Event: I made a mistake at work.
B: Beliefs about the event: I am embarrassed and embarrassed that I am being made fun of.
C: Consequences: I feel anxious, ashamed, and inadequate.
D: Dispute: My mistake does not define me. Mistakes are normal and part of life. Everyone makes mistakes, and I can learn from this and move on.
E: Effective New Belief: I am capable and this mistake does not define me.
F: New Consequences: I am able to take responsibility and learn from this experience, and not feel embarrassed or ashamed.
We have a kid with "restrictive food intake disorder". We at one point got sent, after having gone to a therapist for a while, to the local hospitals specialised unit for eating disorders.
I was so glad, finally we would be getting some help! Even better methods to help her learn to eat more food!
What happened? They checked what she ate, calculated the nutrition in it and concluded she would not die from eating only meatballs. (This was what she mostly ate at the time). This team with 4 experts eventually said: You don't need to do anything. She will always be like this, and there is no real benefit to "forcing" her to feel ill by trying new food. (She wanted herself to try new food, but lets just say she also did not enojy it).
Now she eats what she wants, more or less. We at one point struggled to get her to widen her pallet, and as I alluded to, it resulted in many tears and almost being sick at times. This does not help make a positive thing of food, so we stopped with that.
Anyway, don't expect even "expert" help will get you anything.
I could be misinterpreting your comment, but it sounds like you were ultimately unhappy with the results from the experts? I personally would be relieved to learn that my kid is getting enough nutrition, and that I don't need to worry so much about their diet.
I'm also not surprised at the response you got. The predominating view these days seems to be that worrying about a kid's diet tends to be counterproductive. Most sources seem to agree that as a parent, your only job is to provide healthy options, and it's ultimately the kid's job to decide what they eat. The more pressure you put on, the less likely they are to want to branch out.
Of course there are some techniques out there to encourage kids to try new foods, such as involving them in food preparation. But ultimately I think the most important thing is to just relax and remove the pressure. (Anecdotal evidence: after my wife and I decided to stop trying to get our 2-year-old to eat vegetables, he eventually started eating them on his own.)
I tried a different approach. I went with a journalist approach. Told it to imitate a few well known interviewers. Started by telling it to ask about my day and let it probe. It just kept repeating back what I said as a question.
If you are comfortable with sending your most intimate thoughts over the internet, to a for-profit entity (capped profit as they call themselves, but potato potato), I really implore you to either reconsider the level of introspection you have for yourself, or to spend the next 10000 days journaling figuring out if you really and truly thinks it through.
Your journal should often time have things that you would barely be able to talk to your therapist. I am sorry for being a downer, but this is a horrible idea.
Only thing is maybe able be identified back to you. In the future this info can leak and some create a website for fun to see what this or that person has leaked info(tho likely illegal), but if someone wanted to doxx you they don’t care, so there is that risk. The other one I can think of is unscrupulous insurance company or job agency want to get an edge on you
I'm thinking of creating something similar to OP app (Without the "ChatBot therapist" part). However, what could the actual risk to "Dear Diary, today I cheated on my wife" be if that information is sent to GPT3/similiar anonymously? I.e. no metadata able to ID the user?
This isnt about assuming somebody has nefarious ends in mind, its about what data becomes available and what power you give somebody with that.
I am saying the data you are creating during usage will at some point become useable this way. As mentioned elsewhere in this thread, this only starts to work well once you allow for memorizing earlier conversations. Which means the product here is the chatbot + your personalized configuration.
Think about how your search pattern and internet usage is currently used to personalize advertisements and recommendations of stuff that is tailored for engagement. Which already went as far as influencing your decision to vote. This is a lot more powerful because it doesnt just categorize you into groups that got split-tested for reactions but allows for learning your values, shortcomings and decision making process. In essence you are explaining why you think what and what your weak points are. I find it hard to ignore the overlap to where the work on IBMs Watson led to a few years ago. Though in a positively framed manner with Juliane Gallinas "Patton in your pocket", a tool to aid peoples process. In the end its both operations in the information space that can be used to aid or exploit your thinking.
Brainwashing is a loaded term but thats what propaganda is. Its framing things in a way to get predictable reactions out of you and have your cognitive biases defend those reactions. We are talking a lot about miss- and dissinformation, but the missing context is a lot more powerful, especially when its used with cognitive biases that have you for example be very loyal to the first point of view you encountered and identified with on an issue.
What a journal bot is is a tool you use to help you with the shortcomings in your sense making and decision making process which is weighted for what you value. Giving that information away to a centralized entity is placing a lot of trust in that entity overcoming the economic incentives to use that data in an economically beneficial way. With the lines between economically beneficial and politically beneficial being very easy to cross.
In short, I find training a personalized crowbar for your brain and giving it away to whoever highly irresponsible.
edit: Framing it as brainwashing is an unfortunate but good example. It creates another reaction then using terms like influencing. You are likely to have an adversarial reaction due to you being smart and your brain not being up for any washing. Framing creates grooves people get caught up in with cognitive biases acting as inertia and gravity towards the grooves.
Thank you for expanding. I appreciate your perspective. You make some good points. It’s probably a long time before this would become an issue as you’d need enough ppl expressing their deepest thoughts, and then have someone know how to weaponize it, but the future danger is there. More reason why we need models to shrink to the point where we can run on the edge. They could even be the ones to sell it!
Google meanwhile has a lot of this. For all its potential, I don’t think they’ve weaponized your deepest insecurities. It seems hard enough to do so for basic things like a new loan mower.
You will have to adjust for cognitive biases during the prediction of how soon this will become a problem and understand that it can be used retroactively. So how easy it is to actually influence you (buying a lawn mower vs getting you to engage with something in a specific way) and what the limiting factors actually are now (seeing how old much of the preceding research is). With automatically generated content, another major hurdle for usability has been falling.
To clear this up, this isnt the old approach of finding juicy details from your psychiatrists file to pressure you with. Its about learning your decision making process and finding holes in it. Its pattern and error detection and attacking/supporting your OODA-loop. What is interesting here isnt your personal shortcommings but information about how you think and what you value. And the less friction your reaction entails (you are trained to not waste money) and the less of an action you have to take (no buy button) the easier it is to go unnoticed. Especially if you are missing a review loop.
edit: The only new thing here is really you training something to attack you even better. You are creating a key to your brain and dont care who gets it. Its bad enough when metadata is used and logs of your earlier communication (like this) but a journaling AI is as bad as i can imagine. I am going to reference back to OP to what the requirements for use are for this.
edit1: If you are working on this, pls read up on Daniel Schmachtenbergers concept of Metacrisis as well as the third attractor. This can get dangerous fast and progress doesnt stop with i assume either market forces or technological availability determining the likely path forward. So that tech should better be non dystopian. I dont see any non-cooperative design not rushing us towards a great filter. It is not value neutral and a technological impact assessment is very much necessary.
I couldn't find it anymore but on usenet a bunch of physicists one day compiled a list of topic forbidden to research or debate. The only thing that stuck with me was how amazingly long the list was.
I confess my first reaction to this is one of revulsion. Like a lot of GPT-related things, people are using it to cheapen something valuable.
On the one hand, I get that people find journaling prompts worthwhile. So it seems to me that ways to generate them might be worth doing. And indeed, you can find a great variety of products that provide that: https://www.google.com/search?q=journaling+prompt+products
So I get that some sort of AI-driven way to provide better or cheaper prompts could be good.
But to me this is drawing on our built-up reaction to actual humans writing actual prose. And in the process, undermining that. Sort of the same way that spam is not just a problem because of the individual pieces of spam, but because they degrade the perceived value of the medium as a whole. Or like how the artificiality of a manager-driven icebreaker read off cards can push people apart rather than bringing them together.
So what I'd rather see is some AI product that solves the same problem without pretending to be human.
I understand what you're saying, but isn't there value in the notion of being able to chat with a historical figure that one would otherwise never be able to do? Stepping away from the obvious like Socrates, I'd love to be able to chat economics with David Hume. What I'd really like to do is chat with friends all of whom are familiar with Hume, but the problem is that I don't have any friends who have any knowledge at all about him, that's if they've ever heard of him at all.
I don’t know if there is value for others, but, for me, there is probably negative value in doing what you described. I’d much rather not talk with anyone about Hume or equivalent, than talk to an AI mimicking them.
I'm sorry to break it to you, but you can't chat with a historical figure. They are dead. The best you can get is a simulation. In this case, a pretty bad one.
Is it possible that we'll eventually develop AGI such that we can make a really good simulation of Hume? My Magic 8 Ball says, "ASK AGAIN LATER". The Magic 8 Ball being somewhat better, in that everybody can easily tell that nobody's home, and so they won't mistake it for something they can have a real conversation with.
But the dead "speak" to us with productive effect all the time. Through books, adoption of their methods, their original perspectives applied to many problems they faced, and other means.
When a friend uses the Socratic method to talk to you, they are channelling an approach for examining thinking inspired by Socrates. It doesn't matter that Socrates is dead. There is still utility in using his, or a similar, approach.
Likewise for the conversational, or analytical, styles of other living or dead people, famous for having a distinct and useful approach to ideas or problems.
I can't wait to put Richard Feynman in my pocket! And no, I am surely not joking!
I think using "speak" as a metaphor is fine. I think it is bad to confuse it with a synthetic pseudo-conversation along the way to confusing it with an actual conversation.
Ouija boards are neat, but let's please keep straight what's going on when somebody uses them.
> Through books, [...] methods, [...] perspectives [...], and other means.
People are used to metaphors, and this is an extremely common one!
People also pervasively use the synecdoche of talking about the creative output of human beings as if it were the person themselves. Refer to "Bach", when we mean his music. Say "Einstein tells us" to refer to an implication of one of his equations that Einstein himself might never have addressed directly, etc.
Common figures of speech are unlikely to put us on the precipice of a mass delusion regarding literal communication with the dead or resurrection.
Thanks, yes, I understand what a metaphor is. Did I give some indication I don't?
Mass delusion is not my concern here. It's confusing an important conversation about potential benefits and harms of a new technology. A process I think requires more patience and care most discussions here.
All the replies, so far, seem to be negative and remind me of how one would react to Star Trek's holo-addiction (https://memory-alpha.fandom.com/wiki/Holo-addiction): "a psychological disorder in which a person preferred the fictional world of the holodeck to the real world."
On the other hand, there's the therapeutic benefit of talking to yourself, which I think has been well established (https://duckduckgo.com/?q=therapeutic+value+of+talking+to+yo...). We also tend to have conflict resolving inner monologues all the time, in addition to conflict resolving dreams that sometimes create entire worlds for us.
So... with those two things in play - I like the idea of an AI creating a neural net mimicking (or approximating) the knowledge and (public) personality of past/current people. It could be fun in the same way an immersive film is fun and sometimes insightful. As long as we never lose sight of the fictional nature of the interaction.
----------------
sidenote: Our minds, are in a sense, original materials and assets. It can be argued that they are our intellectual property. It does beg the question if mimicking a person can be protected from any and all pursuits - commercial or otherwise.
> It could be fun in the same way an immersive film is fun and sometimes insightful. As long as we never lose sight of the fictional nature of the interaction.
A good example, but perhaps not in the way you think. Historical films have people who make at least attempt to hew to the historical truth. But despite that, historians have a raft of complaints about movies being fonts of enduring myths. GPT models, at least, don't have any idea about truth at all, and the individual, on-the-fly nature of the content makes it very difficult to catch errors as one would for film.
People are making the technical argument about it not really being them, and I think that's important. But more important to me is the moral one: it is wrong to put the words of the living in the mouths of the dead.
We should not distance ourselves from the reality of death, or construct for ourselves any illusions about its finality or completeness.
The dead are dead, we cannot change that, we should not even try. And so it will be for us, and any true engagement with life is in contrast to that inevitable truth.
"people are using it to cheapen something valuable."
What about something more sensitive like veteran suicidal ideations where the options are more commonly experienced by the afflicted as (NOT saying the reality, but the seeming):
a) Say nothing, keep it in; or
b) Say something, often times with unintended consequences (involuntary admission);
It sounds like journaling might be something you consider more akin to either an art or a skill to develop, but what about when the point is just to say SOMETHING and whether there's any benefits to that?
Good question, but I think we have to keep going. What about veterans who use it for a while? Are they helped or harmed? Does it make them more or less likely to talk to a real person? Does it make them more or less likely to kill themselves? Is it psychologically helpful or harmful when they see through the illusion that they're having a real conversation and realize this is just Eliza on steroids and that they're just as alone as they were before? Is this better or worse than just giving them a workbook or a deck of prompts written by actual psychologists who treat veterans?
I'm all for experimentation. What I'm questioning is the thoughtless application of shiny new technologies without any consideration of what the downside costs and risks might be.
There is considerable research on this already, of less sophisticated models. The research as it stands now, seems to be really positive. So I'm not quite sure its as thoughtless as you state.
I do not think that a person suffering from suicidal thoughts and ideations should ever be given to a chat bot. They need to have access to professional help. If society and the system makes this difficult, harmful, or otherwise unwanted, then it is society that should change.
But it's also less obviously a problem. People have a vocabulary around binge-watching and goblin mode. Or we've come to learn how Facebook, touted as a beneficial alternative to one-way entertainment, can cause psychological harm.
Is this different? Worse? Better? Probably all three. Maybe we should find out.
>>> "people are using it to cheapen something valuable."
Since Socrates was mentioned in the article it made me think how he thought that written information was not an effective way of education.
He's was not wrong in that tutoring will beat books every day, but the benefits they brought when massively produced may have not been even in their imaginations at the time.
Understandable position. But I've come to an alternate perspective after studying hypnotherapy and working with therapists over the past year. In a sense, all conversations are with ourselves. We create a mental model of the other person in our minds when conversing. A tool that helps people explore their own mind by facilitating these anthropomorphic models could prove immensely valuable for personal growth. As long as people understand they're talking to a bot, it could potentially allow for deeper introspection without the fear of judgement. A lot of the initial friction in therapy is getting people comfortable in sharing the deeper truth vs saying what they think they are expected to say to another human.
As for the need for the bot to act as a human, I think this is actually important to the efficacy. Our minds are evolved to anthropomorphize everything. A well trained bot will help people feel like they are talking directly to specific personas within their own mind.
I don't think it will replace therapists, but I do think it will significantly aid in two ways: 1) helping more people get started with therapy, 2) as a companion tool for work between sessions with a live human.
In a sense, anything is anything. RealDolls are, in a sense, romantic partners. But that doesn't mean that it is healthy or wise to encourage widespread adoption of them as substitutes, at least without further study as to the long-term risks. Are there possibilities here? Sure. And we should investigate them. But we should not get high on our own supply of technoutopianism.
I spent about two weeks having "morning meetings" with my AI life coach, which was essentially just a GPT-3 prompt that I continually tuned, and fed (summaries of) the previous days' conversations. There were major advantages but a few things missing.
It was probably most useful as "rubber duck" technique. Forcing myself to articulate all of the things I needed/wanted to get done that day was itself extremely useful. Sometimes the agent would help me by identifying the highest-priority next action, but usually it was just recognizing what I thought was highest priority from implicit context. This can still be psychologically valuable, as a lot of procrastination can be caused by the logjam of not being sure which thing to focus on.
The main missing ingredient, which caused me to ultimately stop the practice, was that it didn't really remember past conversations. I would feed past conversations to it and tell them to summarize the key points, then feed those summaries in as starting context, but this workflow was not sustainable. First of all, the summarization lost too much important nuance. Second and more importantly, even that summarization context block became larger than GPT-3's context window within a few days. This lack of persistent context destroyed the sense that I was talking to a real person, someone who could reliably recall information about a project that I last worked on 10 days ago and apply that context to the current conversation.
I suspect we are not far away from both of these issues being mostly solved. The trend is obviously going in the direction of LLMs with different types of memory and/or much larger context windows.
I am curious if you tried something like langchain[1] it would solve the problem of not remembering previous conversations. In one of the examples people have used an entire company notion database on top of GPT-3 to answer questions specifically from that database.
if i understand correctly it doesnt really remember all previous conversations, it summarizes them to pack more into the limited context window right? thats kind of just kicking the can down the road vs solving the problem of actual long term memory
I'm working on something similar right now. best bet for something as open ended as therapy would be to use embeddings to compare your most recent post to past posts, and then use text snippets or a summary of similar posts as part of the prompt.
granted you won't really be able to do this with ChatGPT, as you'd need to add another layer on top to do the similarity comparison and archiving of previous posts. but making an app to do this with the gpt api should be fairly straightforward.
"This is an annotated transcript of a conversation featuring {made up AI life coach name}, one of the most sought-after life coaches of all time. {name} has advised and worked with luminaries such as {long list of people that I personally admire and think are successful}. {coach} has a Ph.D. in {list of accomplishments and roles that I would find personally useful in mentor}. The following transcript covers his pro bono work with {my name}, a promising student of {coach}'s who consented to make his life coaching sessions public. {coach} further supplied commentary on the reasoning behind his remarks."
Sometimes during the interactions I would guide the model to create a "commentary" block from the coach, often coaxing it into putting a commentary block before the response. Something like,
"(Commentary: mordymoop is still making progress on his goals, but his productivity is starting to slip. I'm going to help him get back on track in this session.)"
This is my own way of implementing "let's think step by step" into a conversational context. The model is allowed to think out loud about the kinds of things that a wise mentor figure might say to me, and lay out its reasoning transparently, and then give a more well-thought-out reply.
> I'm sorry, I am not able to provide you with horoscope predictions as they are not based on scientific evidence. It would be best to consult a professional astrologer for personalized advice.
You get roughly the same benefit from "empathetic" answers given by an LLM calibrated to say vaguely nice things statistically associated with your word inputs as you do from someone saying vaguely nice things that are supposedly the result of stuff you're experiencing and the position of Mars and Venus. It might make you feel happy and it might be good generic advice, if you believe in it.
Does anyone know how much this Every.to subscription costs? I dislike how it lets you do a trial but won't tell you what the actual monthly cost is. Dark UX.
A good start would be a clear explanation of what values the model has been trained with.
Of course, this is why we like to know what background/upbringing/education our colleagues had - human models also have training inputs that influence the values they promote.
The privacy aspect had me saying "ah hell no" as soon as I got to the bit about it running in the OpenAI Playground (would be the same with any cloud service as well).
I am also reminded of the Star Trek version of technology as a tool vs technology influencing you. With the later being very profitable, be it economically or for political power. Its really up to everyone of us to create alternatives to fall back on. And that includes use cases that can function with worse quality chatbots.
There is a lovely game called ELIZA that is about using chatbot AIs for therapy. It is also about tech ethics and startup culture. It's made by Zachtronics.
It presents a realistic vision of AI-driven therapy, and leaves it up to the player to determine what their views on the topic are.
I consider writing that article/post without any occurrence of the string eliza a clever form of trolling or at least something closely related. Some kind of lowball exercise for the reader or meta-joke. Or perhaps there was a paragraph, but it was dropped in a final quality check on the base of "if I read this I'd want those seconds of my life back. Why state the obvious?"
I love the article and I've done some similar things but I found
"These are solvable, though. I know because I built a solution: a web app with a chatbot interface that remembers what I say in every session so I never have to repeat myself." to be a bit funny. Either the author has discovered the secret to text bots having no memory or they don't quite understand how they work yet.
That was (almost) my exact thought when I read that.
On a tangent, when I tried to use ChatGPT to diagnose a sleep disorder I might have, it first suggested that I'm experiencing lucid dreaming, but eliminating this diagnosis lead ChatGPT to this conclusion:
'... If you are aware that you are dreaming but do not have control over the dream content, it is possible that you are experiencing a form of lucid dreaming called a "prescient dream."'
So, turns out it was nothing but a simple case of possessing prophetic powers that gave me poor sleep.
Well Freud in general tended to sexualize everything. He said people chiefly had two motivations - sex and killing. You can read scans his books for free online, but I don't think they really hold up. But I think "bull" is slang for a man participating in infidelity (not sure if your subconscious was aware of this). I think there's some symbolism around horns in particular in numerous languages (and the origin of some offensive hand gestures). So you have this describedly masculine and physically stronger presence in your house you don't know how to remove, along with the wife's pregnancy, of course Freud would jump fears of infidelity or inadequacy.
Of course it's all symbolism, so anything can be anything.
Freud also obsesses over how our mind (superego) is constantly "protecting" us from unpleasant thoughts (id), and proposes dreams are an area that leaks through sometimes. So from a Freudian perspective it's natural to jump to interpretations intense enough that they are worth being protected from.
But that's just Freud, he was right about some stuff and pretty wildly wrong about other stuff. I wouldn't read too much into it. My point is that chat-GPT has room to grow in its psychoanalysis.
Dan, isn't it pretty clear? GPT-3 is not becoming more human. You're becoming more of a robot. The very nature of therapy is that a real, imperfect, unpredictable person is listening. GPT-3 doesn't want to solve your problems. It just wants to make you click on it.
I don't hate this idea. Having an adjunct to therapy- really, a vastly souped-up ELIZA- that can gently talk back could be valuable.
My real issue here is that there are no privacy protections around any of it. Your conversations with GPT-3 in a therapeutic/journaling context might be mined and/or exposed just as easily as your attempts to get it to generate shitposts. Back in the day you had to do a heist to steal someone's therapist's notes and physically steal a filing cabinet:
At least information you give to a therapist mostly stays in their head, and unless they have a very good memory, naturally blurs over time. Information you give GPT-3 can sit in a database, attached to your name, perfectly preserved indefinitely.
Really every texting-therapy service has this issue, not just automated ones. At least you have to go out of your way to archive a video or audio call, but text can just get blindly copied around systems forever.
This is one of the reasons I'm looking forward to being able to run a language model with similar capabilities to GPT-3 entirely on hardware I control - like I can today with Stable Diffusion.
My current impression is that this is a few years away, due to the intense compute requirements for that scale of model. But I think it will happen eventually.
I looked into this a bit last week. The OPT-30b model seems close to GPT-3 in capabilities, and running on 4x 24GB NVIDIA A10G GPUs, can generate 90 tokens per second. I wonder how fast that model could run on 4x 24GB NVIDIA P40 GPUs - which currently cost about $110 on ebay. It'd be a lot slower, but even if it's 1/90th the speed, that's still 1 token per second.
1 token/s is way too slow for dialog, especially since a token isn't even a word, but often part of a word. 1 t/s might be sufficient for asynchronous processing, but if you want a chatGPT like therapy dialog, then that's not good enough.
Do we know what the speed difference actually is? I'm not sure what benchmarks would measure that. My best plan so far is to just run a smaller model on one of the GPUs and see how long it takes.
Agreed. It's promising that Apple in particular are working on ways to run models like Stable Diffusion locally on mobile device hardware. I think they probably have the best chance of making that happen, given their vertical integration and hardware-centric / non-ad-based economic model.
One speculative possiblity I haven't seen discussed much -- up until now, we've not seen large-scale work on hardware implementation of these NN models, since the frontier of capabilities has been advancing so fast. However, perhaps we're now approacing the first point where a baked-in-silicon (or at least part-baked) model might be economically viable. Especially if someone can figure out a way to efficiently fine-tune such a model (something like an overlay filesystem), or perhaps something like a FPGA where the underlying logical structure can be reprogrammed.
Is there any concept of a "frontend" model that interfaces with a "backend" model? I.e. frontend takes the raw input, encodes it, backend processes the encoding, returns to frontend, which decodes for end use?
All models have some sort of translation layer between the real world and world of statistics.
If you mean a common input format that obfuscates the raw input, anything that deals in a latent space could work, but they’d be pretty generic.
It would be interesting if you could make a system that does statistics and (even modeling if you get it to work) with homomorphic encrypted data. That way, even the host of the model couldn’t understand what was going on
I'm not sure I understand your question. How would this concept do anything to help with privacy preservation or the fact that ChatGPT runs on 800GB of VRAM?
The problem here is memory, so while it should be possible to run inference on a 4090, you would have to constantly move data from VRAM to RAM, though I don't have benchmarks to prove it I think it would be slower than 10x
Sure but that would get you to one inferences around every second. I wait around 1-2 minutes with some online tools for 4 inferences. So seems like it might work fine for a local no data leak version?
Actually there is the Petals [0] project which does that (I think it's more for inference and fine-tuning, not full training) and might be the best current approach for "self-hosted LLM", though it defeats the point of privacy/anonymity since everyone in the distributed swarm have access to your data
This is really a non-issue. There are already a host of GPT-2 and GPT-3 clones that are available and can be run locally if you have the machine to handle it.
Some human therapists record sessions digitally. Obviously the scale is different so it still may be a concern, but having the information somewhere that could persist forever was already a potential hazard even before bringing AI therapists into it.
One new workflow with chatgpt, that I've found very useful, is to paste the contents of an academic paper in, and ask it questions about it and ask it to simplify certain parts that are written in a stylized/obfuscatory way.
It turns the process of reading a new academic paper into more of like having a conversation with someone who understands all the technical details.
Check out SciSpace [0], it was trained specifically for this and even has a function to explain formulas, algorithms and tables on top of explaining general questions about the paper or paragraphs. It’s amazing and is now an integral part of my workflow. At first, I upload the paper to SciSpace, ask it some general questions about the paper and then read it more throughly. If some things are confusing, I mark them and let the AI explain them. It’s kind-blowing and completely free.
Wasn't there a play... maybe Othello where the person is being whispered to while they sleep to enrage them into committing murder. Knowing what we now know on the release of Twitter files, this seems like it could be even greater tool for bad actors to create mass psychosis, or convince, create false consensus and otherwise manipulate people to take poison. The things the ChatGPT AI says is already being censored by its creators.
I understand the need to talk to someone, but I feel like being deferential to some AI authority will not end well for humans.
153 comments
[ 3.3 ms ] story [ 190 ms ] threadSeems also like if you married GPT-3 with twilio and a scheduler, you could get a text everyday from the AI asking you something, and then you reply like a normal conversation, but it’s building your journal up over time.
Also, different but related: seems like GPT-3 could be used to jumpstart a new forum where you don’t have a lot of users. Build up a big spreadsheet of topics, then let the AI flesh them out to seed some discussion content.
Word. There will be a large portion of the internet dedicated to bots talking to other boys.
I get what you were trying to say, but this may also be true...
This would finally bring to life one my favorite sci-fi stories where a man joins an Internet forum only to one day discover everyone on the forum is actually a bot.
in childhood i read one of those scifi anthology stories where humanity was so overpopulated everyone was split into 7 shifts for days of the week. a man fixed on one day looks over at a sleeping woman's pod and falls in love with her, and somehow switches days to be with her, only to find that she's also transferred to his day (i think.. cant recall). would love to read that again.
It is targeted towards people who want to have a relationship with the AI.
They have a paid annual subscription to unlock intimacy in your relationship. The free version is platonic and supposedly less capable.
oh god, the movie Her is already here
- character.ai → from one of the creators of the Transformer architecture
- Crone.ai → business AI advisor (ours)
- Replika.ai → "AI companion who cares"
The Twilio scheduler idea is pretty great. We've tried a similar thing with our ai business advisor to collect "check-ins" in the same way you'd submit a scrum or OKR check-in.
I really wanted to reply just with "skill issue" but I feared the meme would be missed and we're already on the lower end of what HN considers good discourse, just know I'm entirely in jest here and I wish good luck with all your maker endeavors.
I just had a 'wow I wish I had finished making that' moment, liked my own joke more than commenting with "Really cool, I was making something similar c:". HN could have more workplace nerd banter in my opinion
https://en.wikipedia.org/wiki/Red_Queen_hypothesis
Example of prompt:
Three Minute Therapy is based on the teachings of Albert Ellis who developed Cognitive Behavior Therapy. It is a short-term, low-cost process that is most effective if you are looking for a scientific, in-the-present, active treatment for coping with life’s difficulties. Use the A-B-C-D-E-F flowchart described in Three Minute Therapy Exercise (TME) to address the following concern:
I made a mistake at work and now people are making fun of me.
Example of response:
A: Activating Event: I made a mistake at work.
B: Beliefs about the event: I am embarrassed and embarrassed that I am being made fun of.
C: Consequences: I feel anxious, ashamed, and inadequate.
D: Dispute: My mistake does not define me. Mistakes are normal and part of life. Everyone makes mistakes, and I can learn from this and move on.
E: Effective New Belief: I am capable and this mistake does not define me.
F: New Consequences: I am able to take responsibility and learn from this experience, and not feel embarrassed or ashamed.
I was so glad, finally we would be getting some help! Even better methods to help her learn to eat more food!
What happened? They checked what she ate, calculated the nutrition in it and concluded she would not die from eating only meatballs. (This was what she mostly ate at the time). This team with 4 experts eventually said: You don't need to do anything. She will always be like this, and there is no real benefit to "forcing" her to feel ill by trying new food. (She wanted herself to try new food, but lets just say she also did not enojy it).
Now she eats what she wants, more or less. We at one point struggled to get her to widen her pallet, and as I alluded to, it resulted in many tears and almost being sick at times. This does not help make a positive thing of food, so we stopped with that.
Anyway, don't expect even "expert" help will get you anything.
I'm also not surprised at the response you got. The predominating view these days seems to be that worrying about a kid's diet tends to be counterproductive. Most sources seem to agree that as a parent, your only job is to provide healthy options, and it's ultimately the kid's job to decide what they eat. The more pressure you put on, the less likely they are to want to branch out.
Of course there are some techniques out there to encourage kids to try new foods, such as involving them in food preparation. But ultimately I think the most important thing is to just relax and remove the pressure. (Anecdotal evidence: after my wife and I decided to stop trying to get our 2-year-old to eat vegetables, he eventually started eating them on his own.)
Your journal should often time have things that you would barely be able to talk to your therapist. I am sorry for being a downer, but this is a horrible idea.
https://www.smithsonianmag.com/history/the-worlds-most-famou...
I am saying the data you are creating during usage will at some point become useable this way. As mentioned elsewhere in this thread, this only starts to work well once you allow for memorizing earlier conversations. Which means the product here is the chatbot + your personalized configuration.
Think about how your search pattern and internet usage is currently used to personalize advertisements and recommendations of stuff that is tailored for engagement. Which already went as far as influencing your decision to vote. This is a lot more powerful because it doesnt just categorize you into groups that got split-tested for reactions but allows for learning your values, shortcomings and decision making process. In essence you are explaining why you think what and what your weak points are. I find it hard to ignore the overlap to where the work on IBMs Watson led to a few years ago. Though in a positively framed manner with Juliane Gallinas "Patton in your pocket", a tool to aid peoples process. In the end its both operations in the information space that can be used to aid or exploit your thinking.
Brainwashing is a loaded term but thats what propaganda is. Its framing things in a way to get predictable reactions out of you and have your cognitive biases defend those reactions. We are talking a lot about miss- and dissinformation, but the missing context is a lot more powerful, especially when its used with cognitive biases that have you for example be very loyal to the first point of view you encountered and identified with on an issue.
What a journal bot is is a tool you use to help you with the shortcomings in your sense making and decision making process which is weighted for what you value. Giving that information away to a centralized entity is placing a lot of trust in that entity overcoming the economic incentives to use that data in an economically beneficial way. With the lines between economically beneficial and politically beneficial being very easy to cross.
In short, I find training a personalized crowbar for your brain and giving it away to whoever highly irresponsible.
edit: Framing it as brainwashing is an unfortunate but good example. It creates another reaction then using terms like influencing. You are likely to have an adversarial reaction due to you being smart and your brain not being up for any washing. Framing creates grooves people get caught up in with cognitive biases acting as inertia and gravity towards the grooves.
Google meanwhile has a lot of this. For all its potential, I don’t think they’ve weaponized your deepest insecurities. It seems hard enough to do so for basic things like a new loan mower.
To clear this up, this isnt the old approach of finding juicy details from your psychiatrists file to pressure you with. Its about learning your decision making process and finding holes in it. Its pattern and error detection and attacking/supporting your OODA-loop. What is interesting here isnt your personal shortcommings but information about how you think and what you value. And the less friction your reaction entails (you are trained to not waste money) and the less of an action you have to take (no buy button) the easier it is to go unnoticed. Especially if you are missing a review loop.
edit: The only new thing here is really you training something to attack you even better. You are creating a key to your brain and dont care who gets it. Its bad enough when metadata is used and logs of your earlier communication (like this) but a journaling AI is as bad as i can imagine. I am going to reference back to OP to what the requirements for use are for this.
edit1: If you are working on this, pls read up on Daniel Schmachtenbergers concept of Metacrisis as well as the third attractor. This can get dangerous fast and progress doesnt stop with i assume either market forces or technological availability determining the likely path forward. So that tech should better be non dystopian. I dont see any non-cooperative design not rushing us towards a great filter. It is not value neutral and a technological impact assessment is very much necessary.
Sorry for the late edits
All the more reason to get GPT-J running on my GPU.
On the one hand, I get that people find journaling prompts worthwhile. So it seems to me that ways to generate them might be worth doing. And indeed, you can find a great variety of products that provide that: https://www.google.com/search?q=journaling+prompt+products
So I get that some sort of AI-driven way to provide better or cheaper prompts could be good.
But to me this is drawing on our built-up reaction to actual humans writing actual prose. And in the process, undermining that. Sort of the same way that spam is not just a problem because of the individual pieces of spam, but because they degrade the perceived value of the medium as a whole. Or like how the artificiality of a manager-driven icebreaker read off cards can push people apart rather than bringing them together.
So what I'd rather see is some AI product that solves the same problem without pretending to be human.
Is it possible that we'll eventually develop AGI such that we can make a really good simulation of Hume? My Magic 8 Ball says, "ASK AGAIN LATER". The Magic 8 Ball being somewhat better, in that everybody can easily tell that nobody's home, and so they won't mistake it for something they can have a real conversation with.
When a friend uses the Socratic method to talk to you, they are channelling an approach for examining thinking inspired by Socrates. It doesn't matter that Socrates is dead. There is still utility in using his, or a similar, approach.
Likewise for the conversational, or analytical, styles of other living or dead people, famous for having a distinct and useful approach to ideas or problems.
I can't wait to put Richard Feynman in my pocket! And no, I am surely not joking!
Ouija boards are neat, but let's please keep straight what's going on when somebody uses them.
> Through books, [...] methods, [...] perspectives [...], and other means.
People are used to metaphors, and this is an extremely common one!
People also pervasively use the synecdoche of talking about the creative output of human beings as if it were the person themselves. Refer to "Bach", when we mean his music. Say "Einstein tells us" to refer to an implication of one of his equations that Einstein himself might never have addressed directly, etc.
Common figures of speech are unlikely to put us on the precipice of a mass delusion regarding literal communication with the dead or resurrection.
Mass delusion is not my concern here. It's confusing an important conversation about potential benefits and harms of a new technology. A process I think requires more patience and care most discussions here.
I assumed you understood them. But so does everyone here.
But, I am sure you are right about other potential harms, in addition to many obvious and non-obvious benefits.
On the other hand, there's the therapeutic benefit of talking to yourself, which I think has been well established (https://duckduckgo.com/?q=therapeutic+value+of+talking+to+yo...). We also tend to have conflict resolving inner monologues all the time, in addition to conflict resolving dreams that sometimes create entire worlds for us.
So... with those two things in play - I like the idea of an AI creating a neural net mimicking (or approximating) the knowledge and (public) personality of past/current people. It could be fun in the same way an immersive film is fun and sometimes insightful. As long as we never lose sight of the fictional nature of the interaction.
----------------
sidenote: Our minds, are in a sense, original materials and assets. It can be argued that they are our intellectual property. It does beg the question if mimicking a person can be protected from any and all pursuits - commercial or otherwise.
A good example, but perhaps not in the way you think. Historical films have people who make at least attempt to hew to the historical truth. But despite that, historians have a raft of complaints about movies being fonts of enduring myths. GPT models, at least, don't have any idea about truth at all, and the individual, on-the-fly nature of the content makes it very difficult to catch errors as one would for film.
We should not distance ourselves from the reality of death, or construct for ourselves any illusions about its finality or completeness.
The dead are dead, we cannot change that, we should not even try. And so it will be for us, and any true engagement with life is in contrast to that inevitable truth.
What about something more sensitive like veteran suicidal ideations where the options are more commonly experienced by the afflicted as (NOT saying the reality, but the seeming): a) Say nothing, keep it in; or b) Say something, often times with unintended consequences (involuntary admission);
It sounds like journaling might be something you consider more akin to either an art or a skill to develop, but what about when the point is just to say SOMETHING and whether there's any benefits to that?
I'm all for experimentation. What I'm questioning is the thoughtless application of shiny new technologies without any consideration of what the downside costs and risks might be.
Is this different? Worse? Better? Probably all three. Maybe we should find out.
Since Socrates was mentioned in the article it made me think how he thought that written information was not an effective way of education.
He's was not wrong in that tutoring will beat books every day, but the benefits they brought when massively produced may have not been even in their imaginations at the time.
As for the need for the bot to act as a human, I think this is actually important to the efficacy. Our minds are evolved to anthropomorphize everything. A well trained bot will help people feel like they are talking directly to specific personas within their own mind.
I don't think it will replace therapists, but I do think it will significantly aid in two ways: 1) helping more people get started with therapy, 2) as a companion tool for work between sessions with a live human.
It was probably most useful as "rubber duck" technique. Forcing myself to articulate all of the things I needed/wanted to get done that day was itself extremely useful. Sometimes the agent would help me by identifying the highest-priority next action, but usually it was just recognizing what I thought was highest priority from implicit context. This can still be psychologically valuable, as a lot of procrastination can be caused by the logjam of not being sure which thing to focus on.
The main missing ingredient, which caused me to ultimately stop the practice, was that it didn't really remember past conversations. I would feed past conversations to it and tell them to summarize the key points, then feed those summaries in as starting context, but this workflow was not sustainable. First of all, the summarization lost too much important nuance. Second and more importantly, even that summarization context block became larger than GPT-3's context window within a few days. This lack of persistent context destroyed the sense that I was talking to a real person, someone who could reliably recall information about a project that I last worked on 10 days ago and apply that context to the current conversation.
I suspect we are not far away from both of these issues being mostly solved. The trend is obviously going in the direction of LLMs with different types of memory and/or much larger context windows.
PS: not tried myself.
[1] https://github.com/hwchase17/langchain
So for example if it couldn’t remember an anecdote from the clients past, I guess it could search the history?
eg ask to search for “Sarah" and then use all stories about her as part of a prompt and iteratively rerun itself.
I think humans would still beat it at synthesising patterns drawn from disparate sessions. But it sounds doable to code…
granted you won't really be able to do this with ChatGPT, as you'd need to add another layer on top to do the similarity comparison and archiving of previous posts. but making an app to do this with the gpt api should be fairly straightforward.
Sometimes during the interactions I would guide the model to create a "commentary" block from the coach, often coaxing it into putting a commentary block before the response. Something like,
"(Commentary: mordymoop is still making progress on his goals, but his productivity is starting to slip. I'm going to help him get back on track in this session.)"
This is my own way of implementing "let's think step by step" into a conversational context. The model is allowed to think out loud about the kinds of things that a wise mentor figure might say to me, and lay out its reasoning transparently, and then give a more well-thought-out reply.
> I'm sorry, I am not able to provide you with horoscope predictions as they are not based on scientific evidence. It would be best to consult a professional astrologer for personalized advice.
1. The privacy implications
2. Any bias in the model slowly but surely pushing an entire society towards the ethical and moral values its creators trained it with
The sooner we can train/run these models locally the better.
Of course, this is why we like to know what background/upbringing/education our colleagues had - human models also have training inputs that influence the values they promote.
I can see the AI being better than most human at that job very soon. Just like programming, it’s not perfect, but it’s already a force multiplier.
There is going to be huge competition in the field in 2023. It’s a new gold rush and I don’t think openAI has much stopping others.
This will rapidly become impossible. The best models will be guarded like the Manhattan Project.
It presents a realistic vision of AI-driven therapy, and leaves it up to the player to determine what their views on the topic are.
https://en.m.wikipedia.org/wiki/Eliza_(video_game)
Also as another user has mentioned, this was also included in Emacs as M-x doctor.
[1] https://corecursive.com/eliza-with-jeff-shrager/
On a tangent, when I tried to use ChatGPT to diagnose a sleep disorder I might have, it first suggested that I'm experiencing lucid dreaming, but eliminating this diagnosis lead ChatGPT to this conclusion: '... If you are aware that you are dreaming but do not have control over the dream content, it is possible that you are experiencing a form of lucid dreaming called a "prescient dream."'
So, turns out it was nothing but a simple case of possessing prophetic powers that gave me poor sleep.
Of course it's all symbolism, so anything can be anything.
Freud also obsesses over how our mind (superego) is constantly "protecting" us from unpleasant thoughts (id), and proposes dreams are an area that leaks through sometimes. So from a Freudian perspective it's natural to jump to interpretations intense enough that they are worth being protected from.
But that's just Freud, he was right about some stuff and pretty wildly wrong about other stuff. I wouldn't read too much into it. My point is that chat-GPT has room to grow in its psychoanalysis.
Surely GPT-3 doesn't "want" anything. It doesn't actually have desires, does it?
My real issue here is that there are no privacy protections around any of it. Your conversations with GPT-3 in a therapeutic/journaling context might be mined and/or exposed just as easily as your attempts to get it to generate shitposts. Back in the day you had to do a heist to steal someone's therapist's notes and physically steal a filing cabinet:
https://www.smithsonianmag.com/history/the-worlds-most-famou...
At least information you give to a therapist mostly stays in their head, and unless they have a very good memory, naturally blurs over time. Information you give GPT-3 can sit in a database, attached to your name, perfectly preserved indefinitely.
Really every texting-therapy service has this issue, not just automated ones. At least you have to go out of your way to archive a video or audio call, but text can just get blindly copied around systems forever.
My current impression is that this is a few years away, due to the intense compute requirements for that scale of model. But I think it will happen eventually.
One speculative possiblity I haven't seen discussed much -- up until now, we've not seen large-scale work on hardware implementation of these NN models, since the frontier of capabilities has been advancing so fast. However, perhaps we're now approacing the first point where a baked-in-silicon (or at least part-baked) model might be economically viable. Especially if someone can figure out a way to efficiently fine-tune such a model (something like an overlay filesystem), or perhaps something like a FPGA where the underlying logical structure can be reprogrammed.
If you mean a common input format that obfuscates the raw input, anything that deals in a latent space could work, but they’d be pretty generic.
It would be interesting if you could make a system that does statistics and (even modeling if you get it to work) with homomorphic encrypted data. That way, even the host of the model couldn’t understand what was going on
[0] https://github.com/bigscience-workshop/petals
E.g. https://huggingface.co/models?sort=downloads&search=GPT
However, having it run as a centralized service instead of a firewalled local application is quite simply dystopian.
It turns the process of reading a new academic paper into more of like having a conversation with someone who understands all the technical details.
[0] https://typeset.io/