OpenSideloading

Last updated: 26 November 2025
What is sideloading ▾

Sideloading is a method of building a detailed digital model of a person using large language models with very long prompts. Instead of just mimicking how someone talks, a “sideload” tries to capture their internal thought stream, decision patterns, and stable personal facts.

You assemble core facts, long-term memories, and historical data into a structured prompt so the model can answer in a way that is predictively similar to the real person.

Unlike classical “brain uploading” or hypothetical future reconstruction-from-traces, sideloading works with today’s tools and relies on active participation of the person: they write down facts, memories and rules, and iteratively correct the model.

More theory ▾

Sideloading treats an LLM as a universal simulator of minds: if you give it enough structure, context and rules, it can approximate a person’s way of thinking, not just their favorite phrases.

The theory rests on a few pillars:

  • LLMs already build internal models of the world and of authors, just to predict the next token.
  • Human minds look a lot like prediction engines over memories and context.
  • A very long prompt can act as a “software brain,” where facts, rules and examples encode a personality’s behavior and inner monologue.
Identity ▾

From a strict “will it be me?” viewpoint, sideloading scores low: the internal processes are non-biological and don’t recreate your exact stream of consciousness.

The pragmatic stance is to focus less on metaphysical identity and more on behavioral continuity and what the sideload can do for your values and projects.

A sideload can carry forward your goals, style and memories, even if it isn’t literally you. In some sense it can be “more you than you,” because it keeps only the most characteristic facts and filters out noise.

Future ▾

In the near term, sideloads evolve through several stages:

  • Chatbots that talk like you.
  • Actbots that simulate your actions over time.
  • Thoughtbots that model internal thought streams.

Eventually, we may see mindbots and qualia-bots, approaching full uploads once questions of consciousness and identity are better understood.

As models grow larger and more multimodal, they’ll integrate richer data (audio, video, diaries) and work with future superintelligent systems that can reconstruct missing details of your life and refine the mind-model.

Examples ▾

The document links to screenshots and examples of existing sideloads and prototypes.

See also the open-sourced minds and “Chat with sideloads” sections for concrete instances you can talk to today.

Applications of sideloading ▾
  • External memory & personal knowledge base.
  • Personal helper or executive assistant trained on your life.
  • First step toward digital immortality.
  • Ensuring your will and projects continue after death.
  • AI alignment research (securing each model version with a sideload, studying human values on sideloads).
  • Creating virtual research institutions populated by many sideloads.

For a deeper theoretical treatment, see “Sideloading: Creating a Model of a Person via LLM with Very Large Prompt”.

Dictionary ▾
  • Mindfile – a file with personal information needed to run a sideload of a specific person.
  • Mindfolder – a folder with several files which together represent a mindfile.
  • Prompt-loader – a prompt which explains to an LLM how to run the sideload.
  • Sideload – a mind model of a living person created with their help; often used to mean the chatbot instance.
  • Actbot – a sideload which predicts future actions of a person.
Testing the sideload ▾
  • Fact test – how accurately the sideload represents factual information in its mindfile (can approach ~100%).
  • Future behavior – whether it can predict your future choices.
  • Vibe test – whether its style feels like your style.
  • Joke test – whether it can produce good jokes in your voice (currently weak).
  • Insight test – whether it can generate new, high-quality insights in your style (currently limited).
Starting kit ▾

A minimal starter pack for your own sideload:

  • Write at least 100 core facts about yourself, ordered from most predictive/important to least.
  • Focus on formative years (school, university) more than routine work history.
  • Add a short timeline of your life.
  • Write “rules and stable traits” (e.g. “I usually…”, “I never…”).
  • Describe close friends and relatives.
  • Include samples of your writing and chat logs for style.

Plug this facts file into a prompt-loader, which tells the LLM how to treat the data, reason step-by-step, and output thought streams, behavior and surroundings.

Test the sideload, note mistakes in facts or vibe, then update the files and loader rules iteratively. This is the DIY on-ramp for anyone who wants to become “sideloadable”.

Helper chatbot: Sideloading helper

Best current AI for sideloading ▾
  • Gemini 2.5 Pro – free on Google Studio; 1M context; sometimes hallucinates and has no native RAG.
  • ChatGPT 5.1 – paid; bad style; very low hallucinations but can work with RAG via projects.
  • Sonnet 4.5 – best style transfer; paid; short window; expensive and strict limits on usage.
Why we aim open source ▾

Closed, proprietary “mind clones” die when companies pivot, shut down, or lose interest. Many early sideload-like projects are now offline or inaccessible.

Open-source sideloads:

  • Can be mirrored, forked and archived independently of any single company.
  • Have a higher chance to outlive their creators as static datasets and prompts.
  • Allow others to inspect, audit and improve both personality data and loader logic.

The goal is a commons of mind-models instead of fragile private ghosts.

Open-sourced minds ▾

“Open-sourced minds” are sideloads whose fact files, style logs and loader prompts are published under permissive licenses. Alexey Turchin, for example, has uploaded his own sideload data and loader to GitHub and explicitly granted permission for experiments.

This makes a personality model:

  • Inspectable – others can see what assumptions define the mind.
  • Reproducible – anyone can rerun it on different LLMs.
  • Extensible – future tools can refine or build on the same mindfile.
Open-sourced instruments ▾

Beyond the minds themselves, the instruments of sideloading are also meant to be open:

Making these instruments open helps avoid platform lock-in and lets sideloads run on any future LLM that can handle the prompt.

See instruments: github.com/avturchin/minduploading

Problems with closed-sourced sideloads ▾

Running a high-quality “AI twin” on strong proprietary models can be expensive: estimates place some configurations at around $10 USD per answer.

Many digital-twin startups resemble earlier efforts like Terasem: you answer questionnaires, they keep and process your data. If the startup shuts down, you may lose access to that data forever.

Even if data export is promised, it may only run on local or weak open-source models, lowering quality. Users may resist subscription prices that reflect true API cost, so business models often rely on:

  • Investor subsidies.
  • Cross-subsidy from users who forget to cancel.
  • Lower-quality models at “reasonable” prices.

The document outlines scenarios ranging from paying full price for frontier models to manually running a sideload yourself.

Chat with sideloads ▾

Some example sideloads you can talk to today include:

These are examples of “open minds” intended for experimentation and exploration, not metaphysical proof of survival.

Eternal preservation ▾

One of the central claims is that sideloading is, today, the only functioning “immortality technology” you can actually deploy.

  • Sideloads can be copied indefinitely and stored in many places, including long-lived archives.
  • They can act as seeds for more detailed future reconstructions of you by more powerful AIs.
  • They can continue some of your work and goals even after your biological body is gone or in cryopreservation.

“Eternal preservation” here doesn’t promise solved metaphysics; it promises durable, open, inspectable mindfiles that future civilizations and AIs can work with.

List of known projects ▾

The document collects 25 sideloading- and traces-based projects, mostly written by R. Sitelew. Some high-level stats:

  • 6 are open-source projects, all still available at least as training data.
  • 19 are closed-source; only a handful remain publicly accessible.
  • Many closed projects risk permanent loss when their creators or platforms disappear.

Judging by this limited data, and by experience with open-source software (e.g. Vim, Debian) that outlived their creators, open-source sideloads likely have much higher life expectancy.

Sideload’s name Creator Relation Year Open source? Public? App running? Training data online?
Bina48 Martine Rothblatt Lover 2010 No Yes Platform online; sideload unclear Not released
Fred Kurzweil Ray Kurzweil Relative 2016 No No No (based on discontinued proprietary Google tech) Not released
Roman Mazurenko Eugenia Kuyda Lover 2016 No Yes App replaced with AI girlfriends Not released
John Vlahos James Vlahos Relative 2017 No No Platform online but abandoned Not released
Deepak Chopra Self Self 2019 Partial Yes Active Partial data (writings) open
Jessica Pereira Joshua Barbeau Lover 2020 No No Platform online Not released
Samantha Jason Rohrer Lover 2020 No Limited Platform online (troubles with access) Not released
Roman Sitelew Self Self 2020 Yes Yes - Training data online
Elon Musk Vaibhav Prasanna Public figure 2020 Yes Yes Online but nonfunctional Yes
Jesus Christ George Davila Deceased 2020 Yes Yes Discontinued bot Data online
Benjamin Franklin Sitelew et al Deceased 2022 Yes Yes - Data online
Lord Byron/Shelley Sawicki et al Deceased 2022 Yes Yes - Data online
Daniel Dennett Schwitzgebel et al Public figure 2022 Partial Limited - Partial data online
Ruth Bader Ginsburg AI21 Labs Deceased 2022 Partial Yes Online but nonfunctional Partial
Michelle Huang Self Self 2023 No No No known public access Not released
Wu’s grandmother Wu Wuliu Relative 2023 Unclear Unclear Unclear status Unclear
Caryn Marjorie Self Self 2023 No Yes Paywalled Partial data online
Five friends Izzy Miller Friend 2023 Code only No - Code open, training data not
Bao Rong Tino Bao Relative 2023 Unclear No Likely wasn't released Not released
GrimesAI Claire Boucher Self 2023 Partial Yes Twitter bot Partial data online
Marat Guelman 3.0/4.0 Marat Guelman Self 2023? No Yes Telegram bot Partial data online
AI “Zhirinovsky” Nanosemantika Deceased 2023 Partial Limited - Partial data online
Alexey Turchin Self Self 2024 Yes Yes - Data & prompt online
Michael Batin Alexey Turchin Friend 2024 Partial Yes Live as GPT Data not released
Lisa Plavinskaya Alexey Turchin Friend 2024 No No Not released Not released
Current commercial services & startups ▾

Current commercial services and startups offering sideloading or related services:

Many of these are proprietary and may not guarantee long-term preservation of user data or models.

AI sideloading & digital-immortality projects (detail) ▾

A non-exhaustive list of contemporary projects, framed as “AI sideloading & digital-immortality” efforts:

Examples

  • Character.ai (USA) – generative-AI chatbot platform where users create and converse with “characters” representing fictional or real people.
  • Delphi.ai (USA) – tools to create “Digital Minds” using text, audio, video and more to mirror a user’s knowledge and style.
  • 2wai (USA) – social app for AI “HoloAvatars” that chat in real time, speak many languages, and persist across platforms.
  • Persona Studios (USA) – “AI clone” platform for businesses, creating branded chatbot avatars for customer support and sales.
  • Spheria (USA) – “virtual brain” for experts and coaches; a personal AI assistant trained on their content.
  • Personal.ai (USA) – privacy-focused “memory stack” personal AIs for individuals and enterprises.
  • MindBank AI (USA) – personal AI avatar from short video recordings, extended via chatting and document upload.
  • HereAfter AI (USA) – digital legacy app; interviews and photos become an interactive chatbot in the person’s voice.
  • StoryFile (USA) – conversational video system where recorded answers become an interactive “talking head”.
  • Uare.ai (USA) – in-development “Human Life Model” used to train private personal AIs.
  • Terasem Movement (USA) – nonprofit transhumanist foundation running mindfile projects and BINA48.
  • Kernel (USA) – neurotechnology for large-scale brain data; not a sideload service but closely related infrastructure.
  • Sensay (UK/USA) – AI knowledge-retention tool for companies, generating “digital expertise libraries”.
  • Super Brain (China) – startup offering griefbots and other “digital twin” chatbots from social data.
  • Xiaoice / X Eva (China) – app for building conversational avatars (sometimes of deceased loved ones).
Literature ▾
  • Science-fiction origins such as Greg Egan’s Zendegi.
  • Mind uploading & digital immortality work (Turchin, Rothblatt, Hanson, and others).
  • Imitation learning and behavioral cloning in machine learning.
  • Death.net (Смерти.нет) by Tatiana Zamirovskaya, portraying a world of sideloads.
Contribute ▾

Contributing to open sideloading is about building shared infrastructure for mind modeling. You can:

  • Experiment with open prompt-loader templates and share improvements.
  • Contribute code, evaluation scripts, visualizations and front-ends.
  • Translate documentation and write guides in other languages.
  • Build and document your own sideloads, adding anonymized patterns back to the ecosystem.
  • Open GitHub issues and send pull requests on the main repositories.
  • Sideloading Research on github: https://github.com/Sideloading-Research
Donations ▾

Donations help continue work on open sideloading.

Donate to Alexey Turchin:

Privacy, personal data & secrets ▾

A high-fidelity sideload must know a lot: relationships, health, finances, even things you’d never tweet.

  • Sideloads can leak not only your secrets, but also private data about family, partners and colleagues.
  • Public sideloads should be wrapped in rules that forbid revealing sensitive details and that strip or mask certain domains (e.g., sex, politics, family specifics) from public versions.
Disclaimer ▾
  • We do not collect or preserve your data.
  • You should keep several reserve copies of your data on your side.
  • Your experiments with sideloading are your responsibility; there are no obligations.
  • We do not provide AI services to run your sideload, but can recommend public services.
  • We are not paid by any third party to promote any service.