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

71–80 of 145 posts

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

#71

Post by jacquesm »

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…

"Do me a favor, boy. This scam of yours, when it's over, you erase this god-damned thing.".

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

#72

Post by jacquesm »

ilrwbwrkhv wrote: post #61

We need to learn to let go. Chatting with a deceased loved one is basically equivalent to the ressurection stone in Harry Potter. A faint reflection which will drive people to insanity. This is not healthy at all.

Harry Potter was very late to that particular party.

John Crowley, 'Snow' (1986) comes to mind as well as Gibson in 1984 with Neuromancer. And likely the concept is much older than that.

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

#73

Post by jacquesm »

fractallyte wrote: post #47

My comment on something similar, 5 years ago ( https://news.ycombinator.com/item?id=18546242 ): Better be careful with that... From "Bicycle Repairman", by Bruce Sterling (1996) (This is a spoiler, BTW!): "The mook speaks just like the Senator did, or the way the Senator used to speak, when he was in private and off the record. The way he spoke in his diaries. As far as we can tell, the mook was his diary.... It used…

But way after Gibson.

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

#74

Post by valray »

szmerdi wrote: post #53
orenlindsey wrote: post #5

I wanted to do this myself but I don't have enough messages yet. Also, I wish the author shared some more technical information on how to do it.

Similar project with enough technical details: https://asmirnov.xyz/doppelganger

Alexander Smirnov's article was discussed on HN: https://news.ycombinator.com/item?id=38434914

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

#75

Post by Sohcahtoa82 »

eightysixfour wrote: post #18
Sohcahtoa82 wrote: post #16

Earlier quoted context omitted.

> I recommend fine tuning mistral What kind of hardware does that take? I've got an RTX 3080, i9-9900K, and 32 GB of RAM right now. I eventually plan on upgrading to an RTX 5090 when those come out, as well as i9-15900K (Or whatever they're gonna call it) and 128 GB of RAM once the next Intel CPU comes out. My understanding is that even with an RTX 5090 (Assuming 32 GB of VRAM, nothing has been announced yet), if I u…

Pretty sure an fp32 7b model uses just shy of 27gb of vram. The performance losses at int8 are minuscule, even down to int5 is still very usable. Why would you try and run it fp32 or fp16?

My naive intuition is that you'd want to train on the highest precision you have the horsepower for, and then after training, you might be able to get away with quantizing to a lower precision for inference.

I would also expect that fp32 might be overkill, but that anything less than fp16 would start to show a noticeable loss in quality, especially in training. I can't imagine int8 being good at all.

And I would think training with int5 would produce garbage. I didn't even know int5 was an option.

Do you have any published benchmarks comparing the quality of results between the different precisions, going all the way down to int5?

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

#76

Post by chrisco255 »

wyldfire wrote: post #2

> I immediately recognized this would be in extremely poor taste and removed the data. Enough on that topic! Poor taste for a blog post. But I wonder if LLMs could advance to a point where it might be therapeutic to have a conversation with a deceased loved one.

I think it would be ok from the right frame of mind. To think of the LLM as a digital memento and not as some kind of replacement for the deceased.

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

#77

Post by citizenpaul »

alchemist1e9 wrote: post #13
mewpmewp2 wrote: post #8

Earlier quoted context omitted.

> it might be therapeutic to have a conversation with a deceased loved one. Would it? I'm not a specialist, but it would seem a bit off to me to approach it like that. I'd think best is to accept that your deceased one has gone, and it seems like having a fake one might just prolong the grievance process and possibly also develop into an obsession of trying to replace your deceased one with the simulation, upgrading…

Have you ever watched the original Superman movies and Clark talks with the AI hologram of his biological father from his home planet? I believe in the future our descendants will likely find it normal to talk with AI versions of their ancestors, perhaps not us parents right now, but I suspect my children will willingly train and generate AI versions of themselves for future generations to interact with.

Yeah but clark never met his father it was meant to for information transfer. He was not using it as a crutch to avoid dealing with the mental health aspects of his fathers death.

Also as others have pointed out it suffers from the hollywood positive-ification of things and never attmempts to take on any serious aspects of the tech. Its just a plot device.

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

#78

Post by justinlloyd »

Exuma wrote: post #62
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…

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.

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

#80

Post by a_bonobo »

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's this recent cool SF novel by the Polish (and undertranslated!!!) SF author Jacek Dukaj called 'The Old Axolotl'. In it, people used gaming systems to copy their 'beings' into machines before a planetwide wipeout event of biologic life. The machines, with the human simulacra within them, survive for hundreds of years. What's interesting about the books is that the machines feel human and have all the human emotions, but somewhat foreshadowing LLMs, they cannot learn like humans learn. The robot society is essentially stale, forever, nobody develops. (at least that's my interpretation as to why the robots fail in the book)
Post reply on HN