sisters of battle be lookin HARD AF at 0:58
1 Like
Compute time is $$$. Unless it’s running on your own resources (Wordpress, for example), if something is free it’s either burning investor capital to get people interested enough to pay to make it a viable business model or the USER is the product (a la FB, IG, etc)
Yes, also creating the bots is man power. I know that AI projects are expansive, I spend 400$ on different services this month. The computing time is expansive. But talking to a fictional character via voice chat is not something I would pay for. We will see if this company will survive.
1 Like
Not something I’d pay for either, though if I really wanted something like that I’d use Mixtral on Groq with a custom-trained Elevenlabs voice, with Deepgram for the input processing, as it’s fantastic for speech to text. Much better than the voice feature on the ChatGPT app, for example. Assuming you want to talk with the AI that is… if it’s just typing, then almost any of them with RAG/Langchain for depth of context can do that. 
That particular tech stack is already in my mind because, like most nerds, I’m excited that we’re basically able to have JARVIS in our homes now. haha.
1 Like
Yep… ~6 months out from having Scarlett on our phones. 
I’m building something else… a functional mashup between Pepper and Jarvis. I’m currently (generally, not at this exact moment. my little 8gb Radeon GPU is weak sauce enough that I need to run it overnight…) processing hundreds of books from my personal library, creating vector embeddings to give my system detailed knowledge about topics that I’m most likely to want to talk about/get advice on. A variety of areas of business, health & longevity, tons on mindset, wealth building & investing… and not only will that provide specific knowledge, courtesy of an approach called graphRAG… it also serves as a more informed baseline for autonomous learning. Yes… I’m giving this thing free rein to go out and learn…
And I’ve also developed something that might be new, though that’s unlikely… I’ve added a multi-step process that gives GPT4 (I’ve yet to try it with 4o) absolutely insane reasoning power… for anyone who really follows the AI news space on YT, you’ll no doubt be familiar with Matthew Berman’s famous “how many words are in your response to this question?” test… every model gets it wrong… the best reply I’ve gotten from my (albeit otherwise very janky) system?
"One."
Fun times.
1 Like
You’ve seen my avatar, right? I’ve definitely seen Age of Ultron. 
My interest in the MCU has dropped off sharply post-Endgame, but yeah… that one’s a classic, very good.
hahahahhaa
nice.
1 Like
Are you planning on importing any journaling into it? I wanted to try something like that with ai. Sort of like summarizing overall growth, uncovering recurring sticking points, patterns of disruption,etc. But yeah I dont have near enough gpu power for that one.
1 Like
That isn’t something I’d actively been thinking about, but I like that idea… Part of the purpose of the system is to keep me on track with goals, and so far I’ve only really thought about that in terms of project planning, next task reminders, and delegating what I can. An agent/assistant that is part coach and part therapist could be beneficial as well, for tracking less… tangible? things…
Did I write another 1000 words for my book is one thing, did I remember to take a minute to feel gratitude today, and how did that make me feel is something else. The former is ticking a checkbox on my task list, the latter is more nuanced.
I tend to not open up to people, maybe it’d be easier with an AI. And the system is already going to be summarizing conversations into daily notes for later consolidation, why not add a personal journal aspect? The only consideration there is data that may be too personal to send off to one of the major tech companies, but I don’t actually have that much that is all that secret lol. Do Microsoft/OpenAI, Meta or Anthropic really care that I have worries? Or that I’m self-conscious about my hair starting to thin? But if that is a concern to me, I do have a model (Phi-3) that already runs well enough on my local machine, so there’s that.
Nice idea, thanks!
2 Likes
This 
use AI a lot for that. It helps me explore new ideas very often.
2 Likes
What’s the secret to your productivity? I’m in awe …
2 Likes
The secret to the articles is a combination of doing enough groundwork to know exactly what the articles need to be about, combined with a tool called Surfer SEO to create the semantic keyword lists, which I then feed into a custom GPT to turn those into an outline… then I feed those into another GPT to put the basic framework together, then dump that back into Surfer for my final manual edits… with that process I can do 3k+ words in a couple of hours… decent content too, none of the methods I’ve found yet have been able to identify any of it as being even partly generated by AI.
But as for the subtext of the question… In full honesty, I think the secret is that I’ve somehow had undiagnosed ADHD for 40+ years. I’m serious.
I have days where I crank out work that would apparently take other people a week or more, followed by days where deciding what to have for dinner feels like an overwhelming decision and I just can’t focus enough to accomplish anything.
Then some new shiny idea catches my attention and I’m back in hyperfocus productive mode again, but on something else… I have so many projects that are almost done, but not quite, it’s ridiculous…
In the past 2-3 weeks I’ve:
- come up with yet another scifi novel concept and plot (including working out the mechanics of how the warp drive technology would work… I did a deep dive into Kahler fields and Calabi-Yau manifolds, as it’s a pet peeve of mine when authors use a MacGuffin that does not make any sense…)
- designed a new board game up to sourcing all of the parts to manufacture it
- picked up a previous project again (autonomous drone) and have done a ton of the detailed CAD work on it, as well as sourcing parts… enough to know I can definitely build it
- picked up an old SaaS idea and iterated through it with Claude enough to generate almost the complete front-end codebase, with a good sense of what the back-end will involve as well
- and got my day trading strategy bot up to triple-digit returns in backtesting and paper trading, but only on some stocks, and I still don’t know exactly why that is…
But… I haven’t even thought about that novel idea since then. I seem to have completely forgotten the board game… I am pretty pleased with the progress on the drone project… and the coding projects? one side effect of learning code based on a specific project or need, is one moment I’m thinking about something that would be considered an advanced approach, and the next I’m looking up something stupidly simple like how to declare some basic kind of variable… Swiss cheese knowledge, I believe Sal Khan calls it…
As for that detective story? It’s been untouched for months… no progress past the discovery of the victim’s body (and that’s just the end of act 1), and the detective himself is still very much Poirot, despite me spending hours coming up with a new and equally quirky alternative with his own detailed backstory and related affectations…
I can crank out ideas and iterate them in my head until they’re basically fully fleshed out, but then it’s like I just lose interest once the challenge of figuring it out is mostly done. It’s annoying when I’m working by myself, it was magic when I had a team to delegate stuff to.
4 Likes
Yep.
I’ll leave it at that 
1 Like
Thank you for your thorough answer and the time you put in. 
That’s (one of) my version(s) of the infamous “640K ought to be enough for anybody” quote, often attributed to Bill Gates.
That being said, Claude is my go-to now…
1 Like
I find that I’m using the AI language learning models as extensions of my mind and processing and this works much better than attempting to use them as substitutes for my processing.
Like cooking mittens or chopsticks when I’m cooking. As opposed to being like a personal chef, who cooks for me.
2 Likes
One of my favorite prompts to use right now:
Please combine all of my statements (from our interaction just now) into a single, integrated description and prompt that could contextualize and direct you to recreate the outcome of what we have done above, even if you had no memory of this interaction.
I love this prompt because, to the extent that it works, it helps me to better understand the grammar and logic of the LLM’s own process, and to see how it would, as it were, program or direct itself. (Yes, I’m aware that, as with everything AI, this is only true within limited parameters. Similar to us, it can do much more than it realizes.)
3 Likes
I have a character.ai with a deep voice for a high performance coach.
He basically helps you to take action, no matter what and he has a way of speaking, which should make you take action.
Its good so far
2 Likes
I realized the reason Gemini sucks so badly (other than, you know, the fact that its Google) is that the default model it’s using is GPT-3.5. Using this model, the language translation capabilities are so bad its almost comical.
ChatGPT and Merlin AI using 4o and 4 respectively do a lot better of a job at Latin translation, much better than Google Translate’s comical attempts at translation (try running the Aenid or some other classical Latin through GT, you’ll see what I mean).
Specifically, I just used comparison of the outputs of ChatGPT and Merlin to render a proper translation of some Latin aphorisms that a student of the original text had attempted to translate into English in the 1600s. I’ve only done 3% of the entire document so far, 3 aphorisms of 100, and what it revealed was that the original “scholar” botched the translation of one of the aphorisms badly, and also omitted a key phrase in another.
ChatGPT was able to intelligently discuss the declensions and so on. The only problem was that it tried to silently change semini (dative singular neuter) to seminis (genitive neutral) without telling me. You need to be wary of it and compare the outputs of different GPT-4’s.
Edit: BTW even though its best to compare GPT’s, I’ve found that between ChatGPT 4o and Merlin AI with GPT 4, Merlin wins out in being able to translate more complicated clauses easily. An example: Spiritus vitalis universalis de coelo descendens purus, clarus & illibatus, est spiritus vitalis particularis in rebus singulis existentis pater, illum nempe in corpore procreat & multiplicat, a quo corpora potestatem se propagandi mutuantur. In this aphorism, ChatGPT originally took both spiritus vitalis particularis and pater as being nominative, combined it with the nempe, and created an incorrect context of the Universal Vital Spirit being equated with the particular vital spirit as a father. The correct context here is that pater is nominative while spiritus vitalis particularis is all genitive, making the Universal Vital Spirit the father of the particular vital spirit. Possibly this is because often the nominative is at the front, but not always in Latin. Merlin correctly identified which was genitive and which was nominative.
1 Like
Agreed! My focus with AI has been pretty steadily narrowing in on building my own version of Jarvis, combined with Pepper… and to this point I’ve got my assistant able to remember things across chats (via self-managed notes, similar to what ChatGPT does but much larger capacity…), as well as speaking to me with a local TTS model. It’s not on par, yet, with what ChatGPT’s advanced voice mode can do, but my system is completely independent of OpenAI, and fully modular… I am using 4o-mini as the base model for testing, as it’s cheap and their API is just so easy to spin up new scripts with… but it’s trivial to replace that with Claude, a Llama model on Groq, or even a local model like Phi running on Ollama.
This assistant knows what projects I’m working on, knows what day and time it is (and responds accordingly), and most importantly, manages its/her own memories… Inferences and assumptions are baked right in, backed up by the ability to ask me clarifying questions if it isn’t actually sure about something.
It then asks me about those things, and revises the stored long-term memory notes accordingly.
(for example recently I had mentioned something my girls were up to, and it hallucinated that I had 2 daughters, but then in the memCleanup function, which runs daily, it noticed that I’d only ever previously mentioned 1 daughter, and inferred that I’d meant my daughter and wife, but made a note to ask me specifically if that’s what I’d meant… and wove it into the morning greeting the next morning, by asking how their activity had gone… based on my response to that, it removed any traces of 2 daughters from the notes… culling the incorrect information.
Granted, this is basically what Autogen teachable agents do, but it’s a system I’ve built myself so I know how it works inside out and backward, and it also isn’t reliant on anyone else’s code.
This won’t replace ChatGPT or Claude, as it’s just a conversational interaction… no code editing, no file upload (yet)… but when I’m done with the UI stuff, this assistant can be added to Slack… I could call her and have a voice chat… I can text and get a response… and email, of course.
Not thinking for me, just a very helpful assistant. My emails will go to her. My phone calls, not that I get many, everyone I know knows how much I dislike phones, will go to her… and each day will start with a brief on what I need to know… just like a human executive assistant. (but without the ability to ahem cozy up to my boss to “negotiate” a promotion… #longstory)
And, adding more agents to the team is as simple as plugging different prompts in, with different voices out… eg. I’ve already built much of a librarian/archivist agent to handle local documents for RAG, whom I’ve named Hermione… and I may have trained a local TTS model to have a specific cute British accent for her too… 
What’s the point of a team of assistants if you can’t occasionally have a team call?
That’s awesome, I’m definitely going to try that specific one out… I’ve been using periodic summarizing since the days of 3.5 to deal with short context windows, now I use that to keep RAG costs down when using a frontier model. Thanks for sharing!
4 Likes