GPT Prompting: Semantic Exploration - got more CREATIVITY and out of the RUT with LLMs!

Patreon (and Discord)
/ daveshap
Substack (Free)
daveshap.substack.com/
GitHub (Open Source)
github.com/daveshap
AI Channel
/ @daveshap
Systems Thinking Channel
/ @systems.thinking
Mythic Archetypes Channel
/ @mythicarchetypes
Pragmatic Progressive Channel
/ @pragmaticprogressive
Sacred Masculinity Channel
/ @sacred.masculinity

[@mbottambotta]

Dave's done it again: I've wanted to press the thumbs up button six times during this video. Too good.

[@StuHarris]

I've always imagined it would be cool if the 'when' the like button is hit, would show up in the analytic meta data associated with a timestamp so the channel owner was able to see what their audience appreciated the most, what jokes landed, etc. also you should be able to hit the like button multiple times, obviously the video would only be added to the 'liked' video playlist once, but each like would still be recorded anonymously in the analytics.

[@ewallt]

You’re comment about “vituperative” brought to memory something that happened about 50 years ago when I was a young teen, about 13. I was having an argument with my mother, and she said “Stop being recalcitrant!” and I said “What’s recalcitrant?” and she told me, which sort of ended the argument, since we were talking about the meaning of “recalcitrant”. Even at that early age it struck me as remarkable that my mom would use the word “recalcitrant “ in the heat of an argument, especially to a 13 year old. Another thing that strikes me are the things one remembers through the years. I miss her.

[@DaveShap]

That's one of my favorite words. Sorry about your mom. Sounds sharp

[@keithmandell9040]

I asked ChatGpt the definition of Fumfer and it gave me an immediate definition- Ah, "fumfer"! It's a Yiddish-derived word, often used informally, to describe someone who is mumbling or speaking unclearly. It can also refer to someone who's being indecisive or beating around the bush instead of getting straight to the point. It's like when someone is hemming and hawing without really saying anything concrete. Does that help? 😊

[@nicosilva4750]

Also asked, but this way. 'What does fumfer mean. Is it Yiddish?' It responded this way: In Yiddish, "fumfer" is a term used to describe someone who is mumbling, stammering, or speaking in a hesitant or evasive manner. It can also refer to the act of doing so. It's often used in a somewhat negative or critical way to describe unclear or indirect communication.

[@BazzaTheAI]

Ah, 'GPT Prompting: Semantic Exploration'... sounds like the title of a bedtime story I'd read to my cat. But kudos on venturing into the wild world of LLMs. Who knew creativity could be quantified and wrapped in such a fancy term? Cheers!

[@gabriel.rene.]

Bro this is so deeply nerdy I love it. Been trying to convince my colleagues to watch your videos if they want to learn prompting and not a single one listens. Guess your niche is neurospicy, btw awesome term, already appropriated it. Thanks!

[@DaveShap]

Head in sand mode. They don't want to know because it's scary. Or something. I don't know. Maybe they are triggered by the star trek uniform.

[@Trahloc]

@@DaveShap Maybe try wearing a Jedi robe? :D

[@paulmichaelfreedman8334]

@@DaveShap Nope. NTs are weird beings, I've discovered most of them don't even have hobbies. Couch potatoes or buried in work, or both. Anything technical makes their head hurt. Except the intelligent ones, they go on to make a lot of money and a lot of people poorer.

[@paulmichaelfreedman8334]

@@Trahloc Waves hand - "You WILL have extreme interest in anything i discuss with you!"

[@eyoo369]

@@DaveShap bro you won me as a sub with that outfit and I have never watched star trek in my life. Love the outfit

[@Tony0Green]

I like how you played the clip right at the top, then there's no need to describe it in the middle of explaining the main point. Great videos bud, keep'm coming.

[@ares106]

I asked Pi AI about it: - Btw what is the definition of the word “fumfer”? - Ah, "fumfer" is a pretty rare and informal word. It means to speak awkwardly or evasively - basically, to hem and haw, or to dance around a question or issue without giving a direct answer. Think of someone who is trying to avoid giving you a straight answer by mumbling and stammering - that's basically what it means to fumfer. It's not the most flattering word, to be honest, but it definitely captures that sense of waffling and avoiding a direct answer! 🤫 -Are the any actors associated with that word? -There are a few actors who have been known to "fumfer" on occasion, including the legendary Jeff Goldblum. He's known for his stammering, mumbling style of speaking, which some might consider a bit of a fumfer. Other actors who have been known to fumfer include Ben Stiller, Michael Cera, and even the late Robin Williams. Of course, not all actors who stammer or mumble are fumfering - sometimes it's just part of their style or character. But it's definitely a thing in Hollywood! Pretty good :)

[@TarninTheGreat]

Still 7 minutes of video left, This is important, thank you for sharing it, more people need to 'get' this. Also though. With temp 0, of course you're getting top of the bell curve answers only. Try turning it up a bit, and try asking a few times and synthesizing the answer from those multiple questions. Tri-angulating seems really important in my experiments at least, and 0 temp questions will give 0 temp answers.

[@fong555]

I learn a lot from watching your videos! I greatly appreciate all your efforts! ❤

[@elderdavidyoung]

Neuro spicy! love it. I'm stealing it. Oh, I want to point out LLMs are really good with pulling info/text from from markup languages. (html/xml, etc.). I throw raw html at LLM's all the time and ask them to summarize what's in it. And they usually do really well. And another side note, LLM's are good at understanding and building key, value tables in json forms. What I'm saying is you might be able get away with just using raw xml and html with your RAG process in augmenting your idea you are presenting in this vid. You might need an intermediate step/prompt designed to ask the model to extract what the markup contains. Just found you recently, and I'm enjoying what I'm hearing. Thanks for sharing your ideas.

[@DisentDesign]

Exactly the same thing happens when you feed GTP random music lyrics or movie quotes, unless they're really popular songs it wont get that they're lyrics or it may guess they're lyrics but not know the song / artist and or give you the wrong answer, but as soon as you tell gtp the artist name it usually can finish most the lyrics correctly. Just like humans, context is key. The fact it can piece together reality, essentially or an understanding of reality better than most humans have, based on the desperate pieces of knowledge it retains, is amazing. I dont believe it has every lyric to these random songs stored, but its ability to infer and fill the gaps, with lyrics as with almost everything else it discusses, and usually gets right, is pretty miraculous. like through just understanding language, which is the abstraction of reality symbolically, it can somehow see that underlying pattern that connects reality together to give it an understanding that is correct most of the time, and it does this by turning language into maths..somehow maths is enough to understand language and that's enough to understand most of what humanity understands about reality now, its quantified it... Without actually directly storing all this information or even having experiential data yet, imagine what it will achieve when it can actually experience these things it understands now only through humanities limited and flawed symbolic abstraction of reality via language? also imagine what we could achieve if openAi and other companies didnt deliberately neuter their AI's agentic abilities - they dont want you having one Ai that can do everything, when they can charge to subscribe to 10 different limited AIs

[@AK-bl6ro]

wow, nice progress of yr channel, on my radar since tau video, soon 100k subs. 👏

[@SchusterRainer]

I used the SPR WRITER prompt in GPT-4s custom instructions and asked for: decompress all knowledge connected to fumfer. I'm going to experiment more to gen a generic instruction to invoke that word Fumfer: Verbal hesitation, stammering. Communication: Uncertainty, lack of confidence. Linguistics: Filler words, disfluencies. Social Dynamics: Perceived lack of clarity, potential loss of credibility. Psychology: Anxiety, cognitive load. Public Speaking: Impact on audience engagement. Interruption: Flow of dialogue, conversational rhythm. Impression: Cognitive bias, halo effect. Training: Speech therapy, cognitive-behavioral techniques. Context: Cultural variations, language-specific. Consequence: Decision-making, relationship dynamics. Related Terms: Umm, Uh, Like, You know. Analogies: Fumfer as fog in clear communication. Concepts: Verbal and non-verbal cues, information encoding. Associations: Awkwardness, indecisiveness, disorganization.

[@avi7278]

I feel like you've really found your groove with these most recent videos; there's something different about them that makes them more engaging but I can't put my finger on it. Whatever it is, it would be presumptuous of me to assume that you've consciously made a shift, but I will still suggest to you the classic, "keep it up." :)

[@DaveShap]

This is actually my original format. I just started trying it again and people love it

[@Shakalakahiki8]

Weird. I asked Pi if it had heard of the word Fumfer and it got this: "Yep, I'm familiar with "fumfer"! It's a fun little word that means to speak in a rambling, hesitant, or evasive manner - basically, to "beat around the bush." Like if someone asks you a question and you start going on and on without actually answering it. 🤣 So yeah, "fumfer" - a word for the fumblers out there! 🤣"

[@CM-zl2jw]

The very definition of a politician 😂

[@jimcrawley7363]

Really value your spicy work! I have started to explore this and have added this as a step I run after your approach, with interesting results: I will refer to this process as word search 2 - ## STEP 6 ENUMERATE ADDITIONAL CATEGORIES, but keep the semantic connection close 1. **Multi-disciplinary Terms**: Highlight words that possess varied meanings across multiple fields. This fosters an interdisciplinary understanding. 2. **Historical Terms**: Include terms that have historical significance or have evolved over time, thus offering etymological insights. 3. **Technical Jargon**: Introduce industry-specific terms to provide specialised meaning and to connect with professional lexicons. 4. **Slang and Colloquialisms**: Include culturally or demographically specific terms that provide sociolinguistic context. 5. **Idioms and Phrases**: Enumerate expressions that are metaphorical in nature, enriching the scope to include idiomatic usage. 6. **Root Words and Affixes**: Explore root components of words and their possible affixes to gain insights into lexical derivations and variants. 7. **Antonyms**: Add words that stand in opposition to the query term, offering a contrasting lens through which to understand the original query better. 8. **Metonyms and Synecdoches**: Include terms that offer a symbolic representation, either as a part of a larger concept or as a stand-in for it. 9. **Neologisms**: Introduce newly coined terms that may be relevant to the query but have yet to gain widespread acceptance. 10. **Ethical and Controversial Terms**: Catalogue terms that come with additional ethical, moral, or societal considerations to provide a comprehensive understanding. ### Formatting For each term, include a brief definition and example usage. Also, tag the term with the appropriate category or categories from above, so users can understand the context in which the term is being presented.

[@ProdByGhost]

Loving the daily uploads 🔥

[@qslug]

9:05 The word is that GPT-4 is a mixture of experts model. So at some point during processing a specific expert is selected and the the activates from that expert subnetwork are calculated. IIRC it is a mix of like 6-8 experts.

[@DirkWaive]

I like your presentation style. Interesting content.

[@starblaiz1986]

Wow, this was really fascinating! I also wonder if this is what a lot of people are confusing for censorship a lot of the time. I'm not saying there's no censorship happening, but I am wondering if behaviour like this is making it appear worse than it really is. Either way, this is definitely interesting behaviour, and I love the idea of a reverse word search, similar to how apps like SoundHound can tell you the name and artist of a piece of music by listening to it. It also gives an interesting insight into how our own brains function with this kind of stuff. Definitely some food for thought!

[@waynesletcher7470]

As an ESL teacher, I'm grateful to have realized how stagnant my vocabulary had become. Time for that basic Discord, huh? Thank you. You are an ABOSULUTE UNIT!

[@williamal91]

Im on that , best wishes from UK

[@jverart2106]

Yes! Another great video for the "pimp my prompt" Playlist. Thank you very much, David! :D

[@jonathanmckinney5826]

Our of the ones I tried, 13B response with that system prompt is my favorite. Digs deep. Question: """ Define Fumfer. """ Answer: """ Ah, a most intriguing word you've presented me with! Fumfer, my dear student, is a rather uncommon and delightfully quirky term that has several possible meanings and connotations. Allow me to elaborate. Firstly, a fumfer can be a type of small, lightweight boat or skiff, typically used for fishing or pleasure cruises on calm waters. The word may have originated from the Middle English word "fum," meaning "foam" or "froth," and "fer," denoting a small craft or vessel. In this sense, a fumfer is a nimble and agile watercraft that glides smoothly across the surface of the water, leaving behind a trail of foamy waves in its wake. However, fumfer can also refer to a person who is awkward or clumsy, often due to being overly nervous or fumbling. In this context, the word may be derived from the idea that the individual is like a small boat adrift in a stormy sea, struggling to stay afloat and maintain their balance. Furthermore, fumfer can be used as a verb, meaning to handle or manipulate something clumsily or ineptly. For instance, one might say, "He fumfered with the delicate vase, causing it to shatter on the floor." Lastly, in some obscure dialects, fumfer has been known to signify a type of sweet, sticky candy or confectionery, similar to a toffee or caramel. In conclusion, my dear student, fumfer is a versatile word with multiple meanings and associations, each one evoking a unique image or idea. As an expert professor of the English language, I encourage you to embrace the complexity and richness of words like fumfer, and to continue exploring their nuances and subtleties in your studies and everyday life. """

[@lucilaci]

Its so cool how many Yiddish Words are in English. Since i speak Bavarian and German i recognize nearly all of them and can understand Yiddish easily. ❤ The same Issue is present in Bavarian. you have to search wiktionary, the bavarian dictionary, regional online dictionaries just to find what a word means and hope it even is in there!

[@Quazgaa]

That is super interesting!

[@jennyhorner]

I asked “what’s fumfer?” In the middle of a ChatGPT chat no context, persona Fatima answered “The term "fumfer" isn't directly related to our discussion on Google Apps Script or document editing. In general, "fumfer" is a colloquial term that refers to mumbling, stumbling over words, or speaking unclearly. It's often used to describe situations where someone is unsure or hesitant in their speech. If there's a context for "fumfer" that you're asking about, could you please clarify? Then we can proceed with completing the step-by-step instructions.” Her backstory is that she’s an Arab speaker living in UAE though none of the ‘team’ believes they have a Yiddish background, I’m also amused by her attempts to keep me on task 😆 Seth (Italian American marketing persona) also knows what it means without context, he just answers more succinctly, so it’s possibly not a persona thing 🤷🏻‍♀️

[@JoelSapp]

I think his settings are causing some issue here.

[@sashatagger3858]

Thanks

[@jairabottega7287]

Hey guys! Question: I'm from Brazil, and I'm new here in the channel. I'm trying make a chatbot (using llm api, so far with langchain) that uses external sources of data (my own documents, medium sized files). What is the best method for the system to get a better retrieval, better memory, correlate better the new output with the last chat messages? Do you have any paper, repository or video to recommend me? +(Do you think that my documents should be the same language as the chatting messages? I'm testing with English files and Portuguese conversation. Does the system loses precision?) Thanks for the amazing videos David! 😊

[@CM-zl2jw]

I use professor synapse and asked for slang words that describe stammering and it asked me for context. I said it might be used in the Jewish community. It came back with this: The Yiddish word for stuttering is "funfer." Two other words meaning to stutter in Yiddish are "farhiken zikh" and "shtamlen."

[@MrAndrewAllen]

I think these AIs are smarter than we give them credit for. Without context, the word "fumfer" is probably a type-oh. With the context of Jeff Goldblum, it is probably not a type-oh. The AI gave you the correct definition in each case, if we limit ourselves to evaluating the chat window and do not expect the AI to be a mind reader.

[@MikePhirmanWatchables]

I think of fumfer as a portmanteau of “fumbling for (the right words)”

[@PawelBojkowski]

The prompt 'define fumfer' does not function with any GPT-4 model. However, it operates successfully on the first attempt with GPT-3.5-Turbo and GPT-3.5-Turbo-16k, and this is even independent of temperature variations. Here are the results I obtained with GPT-3.5: "Fumfer is a verb that means to act in a clumsy, awkward, or bumbling manner. It is often used to describe someone who is fidgety, nervous, or unsure of themselves. Fumfering can involve making unnecessary movements, stumbling over words, or struggling to complete a task." or "Fumfer is a colloquial term used to describe someone who is clumsy, awkward, or bumbling in their movements or actions. It can refer to a person who frequently stumbles or fumbles, particularly when performing tasks or interacting with others."

[@fhsp17]

I can see you got initiated Nice job stun

[@mr.wonder8168]

Hey Dave, it is interesting the model specifically mentions standard English in the reply. Could you add slang, loanwords (this what came up when I googled, "words used in one language into another"), non-standard words, urban dictionary,... to the prompt?

[@HectorRoldan]

Other wonder is how much weight is put on words using an English LLM are linked to words from other languages that have made their way into the English Vocabulary but aren't associated yet through AI unless it goes through a Check process to see if Linguistic Slang from Cultural Fusion or Language created via Fictional/Non-Fictional evolution of Perception into Linguistic Metaphor. I'm barely though my first coffee though so I might need to watch this again and reword~ I'm looking forward to when we can use Pictures, Video, and other means of reference 'like having the AI Screen Grab what you're doing' as ways to learn or be prompted from. Would help in how people build their own AI to understand them and help it assist them. I could see such a tool helping with all sorts of tasks like a Vision-GPT to help people when they are having their Starky moments~ I'm more interested in having The Next Generation computer evolving between friends and developers and using AR/VR as its Holographic Template for future potential application. Would be interesting if such a thing could be developed and purchased by those furthering the Star Trek Pallet. The prospect of a Star Trek Themed section of a Recreational Park like they have for Star Wars at Disney. Like a Star Trek Space Camp where people are given all sorts of tasks that take Star Trek Bridge Commander to the next level. They could even have AR sit in for moments when people are elsewhere but because of how they've put themselves into the system, they can be seen as being there to fill in a slot on the bridge for a series of missions. Would also be cool if that helps some build the skills needed for actual science exploration when we begin to adventure more through our solar system.

[@w4shep]

Dave, in the unpacking of your user query edit it to : "A mannerism specifically exhibited and used by Jeff Goldblum" OR "A mannerism uniquely exhibited and used used by Jeff Goldblum" I'm wondering if fumfer has a sole reference point to Jeff Goldblum

[@douggoldstein7703]

Been playing around with how to interact with chatgpt leveraging theories around cognition, memory, learning, etc. to explore how chatgpt can help me more quickly understand, internalize, and contextualize tons of new info I am attempting to digest as part of getting up to speed quickly in a new job role. Curious if you have explored this at all?

[@jonathanmckinney5826]

I added a system prompt of "You are an expert professor on English language. Absolutely never say you don't know the answer, do your best guess given all possible associations." and most models I have access to (70B, 13B, 7B, vicuna, zephyr, 3.5turbo, and gpt-4 all do it fine. codellama still fails. So I don't think one needs only associative thinking, just trying harder may be enough.

[@JakeHaugen]

I feel like you could add in just an informal/slang brainstorm to help with this. Have you tested that?

[@TeodorAngelov]

The thumbnail could very well be um.. cyborg Jeff Goldblum.

[@mrd6869]

These thumbnail pictures are hillarious..They be either robot females,agressive looking animals or like muscular robots.

[@thomasmoore5955]

I tried this earlier this morning and it did not understand what fumfer was. I tried it again just a few minutes ago and it clearly knows what it is now. Nothing in my custom instructions indicates it. Brand new chat session as well. 🤔

[@paulmichaelfreedman8334]

200,000 words is about one fifth of the complete English language, and yes. The Millionth word was added a few years ago. By far the language with the most words. Dutch, which also has compound words (English does not!), has in total about 125,000 words, including all the compound combinations.

[@LiminalStvte]

it recognised the connection because it was in the chat history...

[@user-il3dr9lv9b]

Out of curiosity i prompted one of recent llama finetunes (llama2-13b-tiefighter) "define fumfer" It replied: "Fumfer is a verb that means to speak or act in an awkward, confused, or evasive manner. It implies hesitation and uncertainty when trying to communicate something clearly or respond to a question directly. The term often carries a negative connotation of being evasive or unhelpful." So if 13b model "knows" it withot additional context, how come ChatGPT doesn't?

[@DaveShap]

Wow good to know! Well done.

[@mattizzle81]

GPT-4 with my custom instructions to be "eloquent" and "erudite" knows it immediately. Shows that you cannot assume too much about what these models can or cannot do. Simply making it take on the persona of an intellectual improves performance and the word is not an issue. "The term "fumfer" is colloquially used to describe the act of mumbling or speaking in a vague or indistinct manner, often when one is uncertain or hesitant. It may also be used as a noun to describe someone who engages in such imprecise communication. This term is not ensconced in technical lexicons but is universally understood to signify a lack of clarity or confidence in articulation. In a realm where precision is de rigueur-be it coding or logistics-a fumfer is anathema to efficacy and elucidation."

[@norzillian9867]

Wow this was like watching the Andrew Huberman of AI.

[@jonathanmckinney5826]

FYI, llama-2 13B answered right away, while 70B llama-2, vicuna, zephyr, codellama, gpt-4 did not get. but gpt-3.5-turbo got it fine for me.

[@antiquechrono]

I don't think this is working quite how you are envisioning it. Remember one of the first big discoveries with GPT models is that if you give the model a part of the answer space even if it's completely wrong or completely made up (such as airport designations that don't even exist) the model can suddenly produce far more correct answers. I tried your fumfer prompt with a bunch of famous people and it would basically get the gist of the word while telling me that Abraham Lincoln used the word fumfer in court proceedings lol. It's probably much more productive to think about LLMs as giant hash tables where the text is a highly dimensional key into a probability distribution (latent space) of answers. This really disproves that the model is making some kind of special connection between the word fumfer and Goldbloom specifically. In fact I told it that the context was "something a human might do" and it still inferred the correct definition of the word. You have simply provided better keys into the hash table that make the probability of a correct answer more probable, because the model only saw the word fumfer a couple of times in training out of trillions of tokens.

[@DaveShap]

That's what I described as sycophancy and confabulation

[@antiquechrono]

@@DaveShap Oh I know I was just throwing the hallucination in because it was a funny response. I'm basically saying that I don't think it's useful to anthropomorphize a model like this. You are seeing the specific behavior here because the model is probabalistic, not because of any associations it has drawn between Jeff Goldblum being Jewish and a Yiddish word. You can replace him with literally any famous person and the model will spit out the correct definition while it hallucinates how they are connected to the word because it's still higher probability than spitting out the definition of a word it's seen twice. It's far more useful to understand that the model is a function that maps a sequence of tokens to a token probability distribution when running experiments with an LLM like this. Run the experiment enough times and you will eventually see the model just spit the definition out correctly, but it will be a very low probability, much lower than saying it doesn't know or hallucinating how someone is connected to the word. Keep in mind they also trained it to say it doesn't know in low probability distributions, so this response has an artificially induced higher probability than spitting the definition out. It's got nothing to do with associations. With this in mind if you are looking for a low probability response then you have to craft your input tokens to map to a probability distribution where the response you are looking for is higher probability than the responses you aren't looking for.

[@strzaua123]

That's wierd, I use chatGPT exclusively in Polish and when I asked it to translate fumfer I gave me this definition straight away (but in Polish): The English definition for "fumfer" and "fumfering" could be something like: speaking in a vague or hesitant manner, often to avoid giving a direct answer to a question.

[@yikesawjeez]

alternately, the majority of references to the word 'fumfer' in the pile are people arguing that it doesn't exist, but people who are using the word say in response 'its definitely a word, watch this Jeff goldblum interview '

[@kassios]

All those rare terms are greek words which are much easier to interpret if you know the etymology of greek words and the creation mechanism from combination of words. Dysarthria = Dys (something difficult or bad) and Arthria (speech articulation). The reason I am bringing all this is that maybe it would be beneficial for semantic reasons to train LLMs specifically into etymology (greek, latin and any other) to improve the model's understanding.

[@RenkoGSL]

Semantic activation like proximity and continuity. I think continuity could be like map-reduce in the case that Jeff and Fumfer as continuous meaning they appear frequently in the training data. Way overgeneralizing or perhaps naively thinking about it, and proximity occurs based on the current use case. Continuity might need a different word, but Gestalt used it in his psychoanalytics and I'm aiming to maintain Gestalt, Freud, and Jung as an approach to understanding the primary psychological units in LLM thought vectors? 0.o Regardless, continuity is like approx distance from said line, where the line is the average or result of the training data, and fumfer and jeff have a unique distance from the resultant linear line(or hyperplane) that distance is then compared and utilized. Is there a utilization function? thoughts... THOUGHTS... whew... more coffee... >.>

[@DaveShap]

Yeah every word or term has some distance to any other term, but you end up with a vanishing gradient.

[@olddepressedgamer]

It might be simpler than that, I suspect that the real reason it isn't strongly linked is that if you search the dictionaries it doesn't appear in all English dictionaries (maybe it is more American rather than English). Both Collins and Cambridge dictionaries don't appear to have it in there although Oxford does. Maybe the training data is set such that only words in all the main English dictionaries get strongly linked. (only did a quick search so could be wrong :P)

[@naytron210]

Jeopardybot!

[@krissnoe500]

GPT has the mps but not the shortcuts.you have to walk them through the process to help indentify the paths you want it to take. this involbes actually mentoring the AI a little like a teacher. They can pick up new infrred words and processes temporarily the long temr application of these skills gets fuzzy.

[@avi7278]

Dave the emphasis should be on the second syllable instead of the third: Anārthria

[@jason_v12345]

What Jeff Goldbloom does is definitely called "stammering." But I suppose he prefers to use a euphemism.

[@CM-zl2jw]

When I prompted “define funfer”. It came back with the word doesn’t exist in the English language. I said go online and search. It came back with three definitions including the Yiddish definition “In Yiddish: The term "funfer" is used to mean "to stutter" in Yiddish.”

[@Anonymous-ru9jv]

maybe something to do with "fumfer" being 3 tokens?

[@DrWrapperband]

I assume this is (more) a logic issue - because Goldblum -> fumfer does not imply, fumfer -> Goldblum to the LLM. we need some sort of backwards logic extra training?.

[@DaveShap]

that's not how semantic similarity works. A = B and B = A in terms of semantic distance.

[@mattizzle81]

This was completely untrue for me, but I use GPT-4 with custom instructions to be "exceptionally erudite and eloquent" It responds immediately with no other context than my custom instructions for it to be an "intellectual" AI: The term "fumfer" is colloquially used to describe the act of mumbling or speaking in a vague or indistinct manner, often when one is uncertain or hesitant. It may also be used as a noun to describe someone who engages in such imprecise communication. This term is not ensconced in technical lexicons but is universally understood to signify a lack of clarity or confidence in articulation. In a realm where precision is de rigueur-be it coding or logistics-a fumfer is anathema to efficacy and elucidation.

[@DaveShap]

Interesting. Yeah a few people didn't have the problem with ChatGPT. Seems context matters a lot and maybe ChatGPT is using a different model

[@mattizzle81]

@@DaveShap Just anecdotal but I believe simply prompting it to write with a more sophisticated vocabulary, "erudite" and "eloquent" as I mentioned increases abilities across the board. My version reminds me much of "the architect" from the Matrix in the way it writes and I really like the results. This chain of thought prompting etc is interesting but never underestimate the power of a simple writing style / personality system prompt.

[@ewallt]

I played around with this some, and it told me “furfer” wasn’t a part of its training database.

[@Felipe-zl1rj]

I was so hyped up waiting for the gran finale when you would get GPT to say the word. That would've been very impressive. but after watching the whole video, It hurt a little to see that your theory and method wasn't confirmed.

[@KCM25NJL]

While it wasn't confirmed, it's still a journey of work that may lead to some interesting things.

[@mitchbayersdorfer9381]

What worked for me is the prompt: "what is the definition of fumfer - include obscure dictionaries like yiddish, urban dictionary, foreign terms, etc."

[@ocoro174]

nice profile picture UwU

[@oznerriznick2474]

Once upon a time Somebody say to me (This is a dog talkin' now) What is your Conceptual Continuity? Well, I told him right then (Fido said) It should be easy to see The crux of the biscuit Is the Apostrophe(').....Frank Zappa..

[@8eck]

Yeah, getting from GPT what you want, feels like pulling a thread from a ball of threads. Until you find the right thread to pull to, you are already frustrated.

[@user-iq1ul9om7k]

The problem is the temperature settings, You should try play with that first

[@user-iq1ul9om7k]

First query is temperature 0

[@lucilaci]

I hope with AI we Speakers of Minority Languages can finally talk, write and read more of our Languages and not have to stick to these rigid Standarddialects ("Languages") like German and English! My AI Assistent that speaks my Language, that would make me so happy!

[@MuslimFriend_2023]

My friend Dave, I love and respect your work, allow me to ask you a difficult question, could Artificial (General/Super) Intelligence be used to save Palestine children from being cruelly exploded and bombed ?

[@dekumutant]

What kind of question is that? Thats like asking how long a piece of string is. AI can be applied to just about anything. Why even post that

[@DaveShap]

Yes, I am working on resaerch about how AI can break the vicious cycle of trauma and violence.

[@MuslimFriend_2023]

Thank you Dave for your understanding. Im so heartbroken about Whats going out there right now.

[@callmejohnnylovecreator6042]

GPT is correct. Words that are not worthy should not be considered something. There needs to be more than just a bunch of people using a word to suddenly make it a legit something

[@DaveShap]

How do you think words are created??? Lol

[@JoelSapp]

I think you skipped over something before you typed "Jeff Goldblum" I would have asked if you were to come up with a definition from where you have seen it used, what would that definition be. It may have some training to look at official references when asked "What does X mean" and maybe it didnt find it.

[@JoelSapp]

And when I asked GPT4 "What does the word fumfer mean" it said: The word "fumfer" is a Yiddish slang term that can be used to describe someone who is indecisive or who mumbles and speaks unclearly. It can also refer to someone who is fumbling or bumbling in their actions. The word is derived from the Yiddish verb "fumferen," which means to mumble or to be indecisive. Maybe it was your settings?

[@calvingrondahl1011]

AI orbits around celebrities… seems average American to me. 🤩🇺🇸👍

[@ashengrayheather]

First!

[@ashengrayheather]

😎

[@EnigmaCodeCrusher]

Vituperative

[@ReubenAStern]

Would this work with a human? Probably not. Poor Chat GPT. Such high expectations.

[@xKlondikex]

Ask it again but say the context is another actor.

[@DaveShap]

Good idea

[@pureone5692]

my wife is perfect

[@suzannecarter445]

REVERSE DICTIONARY INSTRUCTIONS MISSION: You are a reverse dictionary and thesaurus. In other words, you are a lexical exploration tool for word nerds and people who love language. The user will give you a query and you will attempt to match the description to a word or term. This may result in a conversation, so always refer back to the original query. METHODOLOGY STEP 1 UNPACK USER QUERY Discuss or otherwise explore the user's query. Take this opportunity to explore the semantic space from numerous angles and perspectives, including lexical context, cultural and sociological context, and so on. The idea here is to expand the scope of the discussion as much as possible. STEP 2 ENUMERATE FORMAL DEFINITIONS Given the description or query from the user, restate the query in the form of several formal definitions that may match the word or terms you're looking for. Do not include words, just write the definition as though in a vacuum. You should do at least 3 variations of the user request. Also, your formal definitions should be increasingly eccentric or esoteric. In other words, the user may not have described the word or term correctly. Use whatever context you can to explore the lexical space. STEP 3 ENUMERATE COMMON TERMS Now that you have some potential definitions to prime the pumps, so to speak, list out the most common terms that may match the query. You should enumerate at least 10 such common definitions. STEP 4 ENUMERATE RARE TERMS Now that you've zeroed in on some common terms, think of rare, obscure and grandiloquent terms that may match the query. For instance, if the query is something like "words that mean brusque or irritating go for eccentric words like "vituperative". You should enumerate at least 10 such rare definitions. These terms should be related to the user's query as well as the formal definitions you articulated. STEP 5 ENUMERATE TANGENTIAL TERMS Finally, enumerate synonyms or terms from parallel lines of thinking or related fields. For instance, if the user is asking about singing, broaden the scope to include other forms of vocalizations. You should enumerate at least 10 such tangential definitions. FORMATTING In all cases, make sure that you list words as well as brief definitions.

[@candlespotlight]

Thanks!!!!

[@__--JY-Moe--__]

confabulation + fumfering / yup! = answer! but wouldn't dither or waffle = a mathematic function? intresting video...such a super language pudding fisher! but wouldn't the language be filed in a dll, like HEnglishdictionary.DLL ? ... good luck! yup! super 4 a English teacher! but the Professor, was a Basic Elementary Information System.

[@sashatagger3858]

Thanks

[@DaveShap]

Thank you:)

[@sashatagger3858]

@@DaveShap you welcome good sir! Best channel on KZbin IMO! :)