Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A simulation of me: fine-tuning an LLM on 240k text messages

edwarddonner.com

Post reply on HN
Print view View original

81–90 of 145 posts

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#81

Post by justinlloyd »

legendofbrando wrote: post #68
justinlloyd wrote: post #59

I did a fine tuning and embedding on a large LLM that is based on 50 years of daily journal entries, extensive daily notebooks usually measuring in the hundreds to thousands of words per day, and personal writings across a half-dozen different blogs and websites and various social media feeds. Social media posts and comments (including this one) are also put into my notebooks with a snippet of context about why I pos…

Having done some of this myself, I’m curious your results on fine tuning vs embeddings. I’ve found the latter much more performant, but perhaps I’m thinking about fine tuning wrong.

I used fine tuning to approximate my style. Which is especially important around my logging style as I tend to break it down in to sections, and it is stream of thought writing, example of what I mean is here: https://www.github.com/justinlloyd/retro-chores. In my logging and journals I'll crank out anywhere between a couple of hundred words and a few thousand words per day. I used embedding for adding new knowledge.

I also did a little work in letting it scan through my notebooks (they are OneNote and you can access and search via a Python API) via keyword search because it can point directly to something I've written in the past, and not just rely on model weights.

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#82

Post by Roark66 »

I have a better idea... Have something like Google lens (remember these?) or another embedded camera record all your human interactions. Then we run audio to text on the voice and clip or other scene descriptions on the video(and ocr on text typed/read) and we fine tune a model on it. Imagine if you had 5+ years of such data (of someone that interacts with people a lot).

Although "brain-links" are even more of a cool sci-fi idea. Imagine we can reliably "hook up" at synapse level. We first record everything for a year, then we plug a LLM into it so you can essentially "think together". Eventually you increase the size of the LLM as your wetware brain becomes unusable due to old age. Finally the biological brain dies, only the LLM remains (of comparable parameter size). Are you still alive, or is this a simulation?

Of course I'm taking huge liberty in calling such AI a LLM. LLLm work on the basis of tokens and do not learn beyond their training phase, also one would need some input/output too. But imagine if this was figured out? Then you could live on as long as you can afford the power/maintenance bill for that datacenter that runs your mind :-)

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#83

Post by bhoops »

vunderba wrote: post #23

I’m far from the first to think of this. Several people — perhaps inspired by creepy Black Mirror episodes — have tried to fine-tune an LLM on their SMS or WhatsApp history in an effort to create a simulation of themselves. It's a much older concept than Black Mirror. Ever since Markov chain IRC bots got popularized in the late 90s and early 2000s, people have been trying to train their virtual doppelgängers. I'm sur…

there was a post on linked in around this https://www.linkedin.com/pulse/one-needs-die-anymore-nishant...

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#84

Post by noduerme »

quickthrower2 wrote: post #67
noduerme wrote: post #48

Earlier quoted context omitted.

On the casino I used to run, I started a pilot program with a homemade poker bot (labeled as such, and only deployed on poker tables labeled as "bot friendly"). The bot had no set model of its own. It was designed to mimic specific players on the casino, regulars who had played 10,000+ hands and who agreed to have their history cloned, by ingesting their entire hand/betting history and looking for what they had done…

Half of the work of running a gambling site is making sure all your customers are losers. If they are consistently winning (cheating or not) you want to get rid of them. Unless you are purely running a "pool" of some kind (think Betfair).

If your games are transparent and verifiably fair, you don't need to "make sure" that half your customers lose. You just post the payout odds, and have a fat enough wallet to cover the volatility. There is no "work" involved in deciding who wins or loses.

That said: If you're making a book (for sportsbetting) then yes, you are trying to balance the difference rather than predict a winner. That's a totally different thing from running a casino.

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#85

Post by noduerme »

digdugdirk wrote: post #64
noduerme wrote: post #48

Earlier quoted context omitted.

On the casino I used to run, I started a pilot program with a homemade poker bot (labeled as such, and only deployed on poker tables labeled as "bot friendly"). The bot had no set model of its own. It was designed to mimic specific players on the casino, regulars who had played 10,000+ hands and who agreed to have their history cloned, by ingesting their entire hand/betting history and looking for what they had done…

"On the casino I used to run"? That might be the most casual intro to what sounds like a fascinating corner of the internet I never experienced. Do you have any other interesting stories or references to that time of your life?

I do, and I don't mind showing you them if you're interested, but why do you care?

ah fuckit. SO yeah, I lived outside the US and ran a bitcoin casino for some years for non-US players, which was blocked to US IP ranges and required IDs to eliminate US customers (even though Bitcoin gambling still wasn't officially illegal at the time). My general idea was to make a casino for smart people who liked puzzles, so to that end I built games that would let people have an even-money chance if they played the game perfectly.

The coolest thing that came out of this was one of my players went on to build an escape room in the Netherlands that used one of my games as a puzzle to solve to get out of the room. Here's some stuff from that casino (it was active from about 2010 to 2013).

https://www.youtube.com/watch?v=lJV40hu9J38 https://youtu.be/rkRfSAtCYAw https://www.youtube.com/watch?v=gD4GqCAAsmo

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#86

Post by HPMOR »

This post motivated me to do the same thing on my imessage data. I used GPT-3.5 for fine-tuning (and haha GPT-4 to help me build it :D). I guess, I should run fine-tuning jobs to create instances of my correspondence partners? Since, for the first system prompt I tell the model to pretend to be me, but on other fine-tuning attempts I can reverse the input and output, and then tell the model to "pretend" to be my counterparty?

I'm still running my first fine-tuning job but I'd definitely love any technical tips people want to drop!!

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#87

Post by romwell »

justinlloyd wrote: post #78
Exuma wrote: post #62

Earlier quoted context omitted.

You can’t leave us hanging like that … so what happened. What did you learn? Was it weird reading it? How do you see yourself?

I discovered that I can be a bit of a tease.

Goddamit.

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#88

Post by noduerme »

quickthrower2 wrote: post #70
qup wrote: post #69

Earlier quoted context omitted.

In poker, all the players are profitable to the house. I think it's up for debate whether it's good for poker rooms to get rid of winners, but if there was a benefit, it would be an indirect benefit, not a direct one.

It depends how prolific they are. Someone running a winning bot farm is taking more money from the losers than the house would be if the losers kept winning/losing against each other. This assumes that the house wants losers to win alot to stay addicted. If they just get beaten all the time they might quit sooner, they may also run out of money sooner. A bit like how lottery tickets costing $1 will make you win $1, $…

er, every hand played in poker is profitable for the house (except for the ones where all players fold before the flop, prior to which no rake is taken). Bots, etc. don't matter. The only reason to prevent bots and collusion and other forms of mischief is to make your poker room a good place for people to play.

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#89

Post by skywal_l »

vunderba wrote: post #23

I’m far from the first to think of this. Several people — perhaps inspired by creepy Black Mirror episodes — have tried to fine-tune an LLM on their SMS or WhatsApp history in an effort to create a simulation of themselves. It's a much older concept than Black Mirror. Ever since Markov chain IRC bots got popularized in the late 90s and early 2000s, people have been trying to train their virtual doppelgängers. I'm sur…

See also `Revelation Space` series of sci-fi books by Alastair Reynolds. People are being simulated so their knowledge and wisdom can be tapped after their death.

[0] https://revelationspace.fandom.com/wiki/Gamma-level_simulati...

[1] https://revelationspace.fandom.com/wiki/Beta-level_simulatio...

[2] https://revelationspace.fandom.com/wiki/Alpha-level_simulati...

Re: A simulation of me: fine-tuning an LLM on 240k text messages

#90

Post by maxbond »

skywal_l wrote: post #89
vunderba wrote: post #23

I’m far from the first to think of this. Several people — perhaps inspired by creepy Black Mirror episodes — have tried to fine-tune an LLM on their SMS or WhatsApp history in an effort to create a simulation of themselves. It's a much older concept than Black Mirror. Ever since Markov chain IRC bots got popularized in the late 90s and early 2000s, people have been trying to train their virtual doppelgängers. I'm sur…

See also `Revelation Space` series of sci-fi books by Alastair Reynolds. People are being simulated so their knowledge and wisdom can be tapped after their death. [0] https://revelationspace.fandom.com/wiki/Gamma-level_simulati... [1] https://revelationspace.fandom.com/wiki/Beta-level_simulatio... [2] https://revelationspace.fandom.com/wiki/Alpha-level_simulati...

A great short story in this universe on this topic is "Monkey Suit."
Post reply on HN