Bookmarked Using GPT-3 to augment human intelligence: Learning through open-ended conversations with large language models by Henrik Olof Karlsson
Wow, this essay comes with a bunch of examples of using the GPT-3 language model in such fascinating ways. Have it stage a discussion between two famous innovators and duke it out over a fundamental question, run your ideas by an impersonation of Steve Jobs, use it to first explore a new domain to you (while being aware that GPT-3 will likely confabulate a bunch of nonsense). Just wow.
Some immediate points:
- Karlsson talks about prompt engineering, to make the model spit out what you want more closely. Prompt design is an important feature in large scale listening, to tap into a rich interpreted stream of narrated experiences. I can do prompt design to get people to share their experiences, and it would be fascinating to try that experience out on GPT-3.
- He mentions Matt Webbs 2020 post about prompting, quoting “it’s down to the human user to interview GPT-3“. This morning I’ve started reading Luhmann’s Communicating with Slip Boxes with a view to annotation. Luhmann talks about the need for his notes collection to be thematically open ended, and the factual status or not of information to be a result of the moment of communication. GPT-3 is trained with the internet, and it hallucinates. Now here we are communicating with it, interviewing it, to elicit new thoughts, ideas and perspectives, similar to what Luhmann evocatively describes as communication with his notes. That GPT-3 results can be totally bogus is much less relevant as it’s the interaction that leads to new notions within yourself, and you’re not after using GPT-3s output as fact or as a finished result.
- Are all of us building notes collections, especially those mimicking Luhmann as if it was the originator of such systems of note taking, actually better off learning to prompt and interrogate GPT-3?
- Karlsson writes about treating GPT-3 as an interface to the internet, which allows using GPT-3 as a research assistant. In a much more specific way than he describes this is what the tool Elicit I just mentioned here does based on GPT-3 too. You give Elicit your research question as a prompt and it will come up with relevant papers that may help answer it.
On first reading this is like opening a treasure trove, albeit a boobytrapped one. Need to go through this in much more detail and follow up on sources and associations.
Some people already do most of their learning by prompting GPT-3 to write custom-made essays about things they are trying to understand. I’ve talked to people who prompt GPT-3 to give them legal advice and diagnose their illnesses. I’ve talked to men who let their five-year-olds hang out with GPT-3, treating it as an eternally patient uncle, answering questions, while dad gets on with work.
Henrik Olof Karlsson
Ton Zijlstra commented a great post by Henrik Karlsson about the large language model GTP-3, which caused me to finally try it out.
My first impression is similar to theirs: “Just wow”, and it took me quite a while until I reached some limits (in particular when asking GPT to “Write a fictitious debate between xxx and yyy about zzz.”)
One undeniable affordance, however, of the machine’s responses is to get inspirations and stimulation for consideration. This is also the big topic of the note-takers and zettlekastlers crowd, for example using the autolinking of “unlinked references”. And I am noticing that it is probably a matter of taste and preferences, or perhaps even a matter of different working styles: If I am permanently working at my limits there is no room left for organic associations, and then I might be more impressed by an abundance of ideas and artificial creativity?
Perhaps I am too much of an ungrateful grumpy killjoy, but the abundance of artificial serendipitous stimulations makes me think of how onerous it will be to sift through them all to find out which ones are the most relevant ones for me.
Let’s contrast this sort of inspiration with the sort that comes through blog reactions. Karlsson explicitly compares blog posts to search queries and to the new kind of ‘conversations’ that we can have with GPT-3, and I think it is indeed very appropriate to see the interaction with these tools as a ‘communication’. Also Luhmann used this metaphor for his Zettelkasten, as Ton points out, and when we use GPT, the back and forth of ‘prompts’ and ‘completions’ is a dialog, too. So there are many beneficial similarities to blog comments and trackbacks.
Image based on https://www.flickr.com/photos/jennymackness/32921358238/ (CC-BY-NC)
However, blog respondents are not anonymous mass products. They have a background. They care about the topic I write about, and I care about theirs. I subscribe to people whose interests are not always the same as mine but often still close enough to be inspiring. And I trust that it is relevant what they are writing. (Formerly, we talked about bloggers as ‘fuzzy categories‘ and about ‘online resonance‘ and about the skill of ‘picking‘ from the abundance.) The grounding in a shared context and a known background, makes it easier for me to understand, and benefit from, their reactions, probably in a similar way as neural ‘priming’ works.
This is all missing when I process suggestions from a machine that does not know me and that I don’t know (I don’t even know what it knows and what it merely confabulates, and at what point its algorithm switches to the live web to look up more). It is unpersonal — even if it may impersonate Plato in a debate.
Like this:Like Loading…
Related
Manual pingback to https://x28newblog.wordpress.com/2022/08/13/conversations-with-tools/ (too long for a webmention copy).
Summary: The ‘conversation’ with GPT-3 certainly creates inspirations, and it can be compared to blog reactions, or to Luhmann’s communication with his Zettelkasten. But it is a matter of preferences.
John Caswell writes about the role of conversation, saying “conversation is an art form we’re mostly pretty rubbish at“. New tools that employ LLM’s, such as GPT-3 can only be used by those learning to prompt them effectively. Essentially we’re learning to have a conversation with LLMs so that its outputs are usable for the prompter. (As I’m writing this my feedreader updates to show a follow-up post about prompting by John.)
Last August I wrote about articles by Henrik Olaf Karlsson and Matt Webb that discuss prompting as a skill with newly increasing importance.
Prompting to get a certain type of output instrumentalises a conversation partner, which is fine for using LLM’s, but not for conversations with people. In human conversation the prompting is less to ensure output that is useful to the prompter but to assist the other to express themselves as best as they can (meaning usefulness will be a guaranteed side effect if you are interested in your conversational counterparts). In human conversation the other is another conscious actor in the same social system (the conversation) as you are.
John takes the need for us to learn to better prompt LLM’s and asks whether we’ll also learn how to better prompt conversations with other people. That would be great. Many conversations take the form of the listener listening less to the content of what others say and more listening for the right time to jump in with what they themselves want to say. Broadcast driven versus curiosity driven. Me and you, we all do this. Getting consciously better at avoiding that common pattern is a win for all.
In parallel Donald Clark wrote that the race to innovate services on top of LLM’s is on, spurred by OpenAI’s public release of Chat-GPT in November. The race is indeed on, although I wonder whether those getting in the race all have an actual sense of what they’re racing and are racing towards. The generic use of LLM’s currently in the eye of public discussion I think might be less promising than gearing it towards specific contexts. Back in August I mentioned Elicit that helps you kick-off literature search based on a research question for instance. And other niche applications are sure to be interesting too.
The generic models are definitely capable to hallucinate in ways that reinforce our tendency towards anthropomorphism (which needs little reinforcement already). Very very ELIZA. Even if on occasion it creeps you out when Bing’s implementation of GPT declares its love for you and starts suggesting you don’t really love your life partner.
I associated what Karlsson wrote with the way one can interact with one’s personal knowledge management system the way Luhmann described his note cards as a communication partner. Luhmann talks about the value of being surprised by whatever person or system you’re communicating with. (The anthropomorphism kicks in if we based on that surprisal then ascribe intention to the system we’re communicating with).
Being good at prompting is relevant in my work where change in complex environments is often the focus. Getting better at prompting machines may lift all boats.
I wonder if as part of the race that Donald Clark mentions, we will see LLM’s applied as personal tools. Where I feed a more open LLM like BLOOM my blog archive and my notes, running it as a personal instance (for which the full BLOOM model is too big, I know), and then use it to have conversations with myself. Prompting that system to have exchanges about the things I previously wrote down in my own words. With results that phrase things in my own idiom and style. Now that would be very interesting to experiment with. What valuable results and insight progression would it yield? Can I have a salon with myself and my system and/or with perhaps a few others and their systems? What pathways into the uncanny valley will it open up? For instance, is there a way to radicalise (like social media can) yourself by the feedback loops of association between your various notes, notions and follow-up questions/prompts?
An image generate with Stable Diffusion with the prompt “A group of fashionable people having a conversation over coffee in a salon, in the style of an oil on canvas painting”, public domain
I have a little over 25 years worth of various notes and writings, and a little over 20 years of blogposts. A corpus that reflects my life, interests, attitude, thoughts, interactions and work over most of my adult life. Wouldn’t it be interesting to run that personal archive as my own chatbot, to specialise a LLM for my own use?
Generally I’ve been interested in using algorithms as personal or group tools for a number of years.
Most if not all of our exposure to algorithms online however treats us as a means to manipulate our engagement. I see them as potentially very valuable tools in working with lots of information. But not in their current common incarnations.
Some of the things I’d like my ideal RSS reader to be able to do are along such lines, e.g. to signal new patterns among the people I interact with, or outliers in their writings. Basically to signal social eddies and shifts among my network’s online sharing.
LLMs are highly interesting in that regard too, as in contrast to the engagement optimising social media algorithms, they are focused on large corpora of text and generation thereof, and not on emergent social behaviour around texts. Once trained on a large enough generic corpus, one could potentially tune it with a specific corpus. Specific to a certain niche topic, or to the interests of a single person, small group of people or community of practice. Such as all of my own material. Decades worth of writings, presentations, notes, e-mails etc. The mirror image of me as expressed in all my archived files.
Doing so with a personal corpus, for me has a few prerequisites:
It would need to be a separate instance of whatever tech it uses. If possible self-hosted.
There should be no feedback to the underlying generic and publicly available model, there should be no bleed-over into other people’s interactions with that model.
The separate instance needs an off-switch under my control, where off means none of my inputs are available for use someplace else.
Running your own Stable Diffusion image generator set-up as E currently does complies with this for instance.
Doing so with a LLM text generator would create a way of chatting with my own PKM material, ChatPKM, a way to interact (differently than through search and links, as I do now) with my Avatar (not just my blog though, all my notes). It might adopt my personal style and phrasing in its outputs. When (not if) it hallucinates it would be my own trip so to speak. It would be clear what inputs are in play, w.r.t. the specialisation, so verification and references should be easier to follow up on. It would be a personal prompting tool, to communicate with your own pet stochastic parrot.
Current attempts at chatbots in this style seem to focus on things like customer interaction. Feed it your product manual, have it chat to customers with questions about the product. A fancy version of ‘have you tried switching it off and back on?‘ These services allow you to input one or a handful of docs or sources, and then chat about its contents.
One of those is Chatbase, another is ChatThing by Pixelhop. The last one has the option of continuously adding source material to presumably the same chatbot(s), but more or less on a per file and per URL basis and limited in number of words per month. That’s not like starting out with half a GB in markdown text of notes and writings covering several decades, let alone tens of GBs of e-mail interactions for instance.
Pixelhop is currently working with Dave Winer however to do some of what I mention above: use Dave’s entire blog archives as input. Dave has been blogging since the mid 1990s, so there’s quite a lot of material there.
Checking out ChatThing suggests that they built on OpenAI’s ChatGPT 3.5 through its API. So it wouldn’t qualify per the prerequisites I mentioned. Yet, purposely feeding it a specific online blog archive is less problematic than including my own notes as all the source material involved is public anyway.
The resulting Scripting News bot is a fascinating experiment, the work around which you can follow on GitHub. (As part of that Dave also shared a markdown version of his complete blog archives (33MB), which for fun I loaded into Obsidian to search through. Also for comparison with the generated outputs from the chatbot, such as the question Dave asked the bot when he first wrote about the iPhone on his blog.)
Looking forward to more experiments by Dave and Pixelhop. Meanwhile I’ve joined Pixelhop’s Discord to follow their developments.