A simulation of me: fine-tuning an LLM on 240k text messages
Fine‑tuning large language models on personal chat logs, emails, and journals is emerging as a way to create AI “doppelgängers” that mimic an individual’s tone, preferences, and even parts of their inner life. Commenters weigh potential uses ranging from talking to deceased loved ones and preserving family history to post-mortem services and marketing, alongside worries about grief, obsession, consent, copyright, and the risk of mistaking curated text traces for a real person. Technical details, such as model choice, training methods, and hardware limits, surface alongside references to science fiction that has long explored digital ghosts and uploaded selves.
Therapeutic and emotional questions
- Some think talking to an AI “ghost” of a loved one could be comforting or therapeutic; others argue it risks prolonging grief, creating unhealthy dependence, or obsession with “improving” the simulation.
- Comparisons are made to saved voicemails, psychics, and memorial video systems that answer questions after death; unclear whether these actually help people heal.
- Concerns include hallucinated “revelations” (e.g., imagined affairs or favoritism) and the risk that people might choose simulations of dead partners over processing breakups.
- A few suggest the healthiest framing is as a digital memento, not a replacement.
Ethics, consent, and data use
- Some posters find it unethical to train on someone else’s messages without explicit permission; others see it as no worse than a human reading and internalizing text.
- Debate over whether training on copyrighted text is akin to copying a book; strong disagreement on whether models are “compressions” of datasets and what that implies legally and morally.
Technical approaches and limitations
- People share attempts using older models (GPT-2, GPT-J) with large chat logs, often yielding on-topic but incoherent output. Newer models (e.g., Mistral) are reported as much better.
- Discussion of fine-tuning vs. embeddings: fine-tuning for style/persona, embeddings and retrieval for factual recall, sometimes combined with direct notebook/email search.
- Hardware constraints and quantization are discussed; consensus that heavily quantized larger models can outperform small high-precision models, including for low-precision LoRA fine-tuning.
- Practical tips: exporting emails, setting up local IMAP, using existing fine-tuning repos and tutorials.
Potential applications and business ideas
- Envisioned services include ancestor chatbots, posthumous DMs, cemetery or MMO “ghost” worlds, probate/family-history assistants, and managed “digital legacy” offerings.
- Some expect big tech to eventually market such services using email and message histories; others are wary of privacy and marketing profiling.
Cultural, philosophical, and sci‑fi framing
- Non‑Western attitudes toward ancestors may make such services more acceptable in some cultures.
- Several see a risk of idealized, curated simulacra worsening feelings of inadequacy, similar to social media personas.
- Numerous science‑fiction works are cited exploring digital doubles, posthumous simulations, and their unintended consequences, reinforcing both fascination and unease.