GPT-3 is the best journal I’ve used
every.to
every.to
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
E.g. https://huggingface.co/models?sort=downloads&search=GPT
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.
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.
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.
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.
Also as another user has mentioned, this was also included in Emacs as M-x doctor.
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.
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.
However, if that's what it takes to help him journal and reflect on his thoughts and emotions more power to him.
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.
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.
Human therapists don't usually operate on a "walk in and tell me the thing you want to talk about" model. There is a whole onboarding process and an expectation that you'll discuss a range of issues. Personally, I don't consider many of my problems to be worth all of that. Even calling a close friend seems like something I would only do if I felt the issue was really, really important.
Talking to ChatGPT, on the other hand, is as low stakes as it gets. It'll offer incredibly generic advice without passing judgement or burdening another human with my problems. I can be as trivial and repetitive as a I want. It's just a language model.
The privacy considerations are real. OpenAI has my name, email and phone number. But if a reputable organization that I trusted with my privacy offered a similar service (ie, a national psychiatric organization), I would almost consider it.
Can't speak for every country, but in the UK you can call 116 123 and speak to the Samaritans. They're not professional therapists but they are trained to be nice and to listen to people unburdening themselves and trying to make decisions. And yes, they're there at 2am. Also, they're people, so they understand you as a fellow person with emotions, not as a correlation between this text input and that text output. I realise some people might actually prefer the lower stakes of the generic machine output. But the people service does - at least in some places - exist.
However, having it run as a centralized service instead of a firewalled local application is quite simply dystopian.
There is nothing "Socratic" about the Socrates example. There is nothing therapeutic about the therapist examples.
"It sounds like you're experiencing a range of emotions. Is there one that stands out more than the others?" is really a more wordy version of "That sounds really hard, say more about that."
Which is fine! This is a Rubber Duck on Steroids. But at least in the examples listed here, this is hardly a step beyond what ELIZA could have done 40 years ago, and the evidence in no way supports the author's original conclusion that GPT-3 is ALMOST as good as an "Empathetic friend".
It sounds like the author needs better empathetic friends, which to be fair....IS a real problem: https://nypost.com/2021/07/27/americans-have-fewer-friends-t...
The dream interpretation text generation is the most impressive. Unfortunately, there is zero evidence that dreams have deeper metaphorical meaning, as opposed to just a brain defragging itself.
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.
I sometimes just think stuff in my head, conceptual stuff about life, science etc. anything that interests me.
Every now and then, I come up with a fruitful position, or a question.
ChatGPT helped me document these thoughts through a singular propositional in a dense sentence. ChatGPT breaks it down to simpler sentences that is more understandable If I read it in the future.
Basically. For me ChatGPT has been;
Zip(Thoughts) --(6-10 words)--(ChatGPT)--> ./Thoughts (50 words; more future proof)
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.
Crazy privacy implications though.
> 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.
Surely GPT-3 doesn't "want" anything. It doesn't actually have desires, does it?
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 seems like a platitude of the type when people say “there is nothing wrong with capitalism, but”.
Why won’t it be a replacement in a few years? Seeems therapist could be one of thousands of jobs replaced by this.
Are chess engines an adjunct to grandmasters? Kasparov espoused a “centaur” theory for a couple yearsthat a human plus a computer beats a computer alone, but that is now way outdated. Computers beat humans period.
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.