Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

2 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

Overview

AmaniGPT is an experimental AI engine inspired by the Attention Is All You Need paper. That paper describes what attention needs to accomplish. Our job was finding a way to make that work fast enough on ordinary CPUs.

While it is still early days, we are building an AI engine that:

  1. Generates answers instantly so you never have to wait.
  2. Runs offline in situations where internet connections are not available or allowed.
  3. Supports personalization by allowing you to train prebuilt models on your own code and data.
  4. Protects your PRIVACY and VALUABLE IP by keeping all code, data, and messages on your devices or network.
  5. Gives you COMPLETE CONTROL by allowing you to build your own models using your own rules, code, and data.
  6. Maximizes your productivity by REMOVING TOKEN LIMITS since everything runs on your own hardware.
  7. Dramatically REDUCES COSTS by doing all of the above on ordinary CPUs that you already use today.

This is the community repo where you can:

  1. Read the official guide on training your own AmaniGPT model from scratch on Linux.
  2. Read the official guide on licensing AmaniGPT when it becomes generally available.
  3. Provide feedback and ask questions using GitHub Issues.

Benchmarks

While AmaniGPT is still experimental, the goal is to build a product that is worth using and paying for. To that end, the only benchmarks we will discuss and optimize for are the following:

  1. Does it work on your existing hardware?
  2. Is it fast enough for your use case?
  3. Does it produce accurate outputs without hallucinating?
  4. Does it meet your compliance, privacy and security requirements?

Limitations

AmaniGPT is a tool. It is not alive, sentient, or conscious. It does not feel, have an imagination, or want anything.

The hype around AI and the subsequent public backlash is based on treating AI tools as if they are human. They are not.

A calculator is not a mathematician. It is a tool that mathematicians can use for some tasks. AmaniGPT is the same.

We encourage people to use AmaniGPT where it makes sense and to do so while understanding that it has limitations.

Specifically, AmaniGPT models are only as good as their training data and will reflect the same inaccuracies, biases, omissions, and inconsistencies present in that data. Always review model responses before using them for critical work.

Releases

We plan to release one major version of AmaniGPT every year in November, with minor releases happening monthly until the next major release. If you ask us for an ETA on a specific request, the answer will always be one of the following:

  • Minor requests: "at the end of this or next month".
  • Major requests: "at the end of this or next year".

This approach allows us to build high-quality software and deliver it on a predictable schedule for everyone. If you have a specific deadline, please get in touch with our team to share your requirements and get a quote (this option is only available for paid license holders).

Licensing

AmaniGPT is currently in preview and can be used for non-production evaluation purposes only. When it becomes generally available (expected by November 2026), it will be offered under the following editions:

  1. Community Edition: free for both commercial and non-commercial use on up to 32 CPU cores.
  2. Enterprise Edition: paid for use on more than 32 CPU cores. Includes priority support and maintenance.
  3. Hyperscale Edition: paid for use on unlimited CPU cores. Includes priority support and maintenance.
  4. Sovereign Edition: paid for use on unlimited CPU cores. Includes source code for maximum independence.

For more information and up-to-date details, please review our licensing guide.

Pre-orders

If your organization is feeling the strain of AI-related capital and operational expenditures, AmaniGPT can help. By pre-ordering a paid license, you gain access to fast, private, unlimited AI on hardware you already own, along with the following benefits:

  1. Lock in preferential pricing before general availability in November 2026.
  2. Secure a perpetual license to use AmaniGPT technology, regardless of any future exclusivity arrangements.
  3. Influence the roadmap by working directly with our team to prioritize features relevant to your core needs.
  4. Ensure successful adoption by evaluating AmaniGPT with the help of our team before you make a full commitment.

For more information, please review our licensing guide.

Partnerships

AmaniGPT allows you to build and use AI models which solve problems that matter to you. We personally care about the following areas:

  1. Developer tools & IDEs.
  2. Mobile & desktop operating systems.
  3. Browsers & productivity apps.

We would especially love to partner with your organization if:

  1. You offer a product in any of the areas listed above.
  2. You have pre-ordered a Sovereign Edition license for AmaniGPT.
  3. You are enhancing products we already use ourselves.

As a Sovereign Partner, you will receive our full support and close collaboration, since your success directly benefits our own work and the broader community.

Roadmap

AmaniGPT is part of our vision to make AI more sustainable and useful over the next 10 years, in multiple phases:

  • Phase 0 (Preview): we are here now. The goal is to finalize and validate CPU-only AI as a more affordable and sustainable alternative to GPU-heavy AI.

  • Phase 1 (General Availability): expected by November 2026. AmaniGPT should offer fast, offline, unlimited, private AI on ordinary hardware. The marginal cost of training and inference should be close to zero. Intelligence can be everywhere at once, simplifying how we live, move, produce, and trade. Everything can think: from houses to cities, cars to planes, farms to factories, businesses to entire markets. Everything is your agent. Everything works for you.

  • Phase 2 (Agents Reboot): expected by 2027-2028. Agents have struggled in the past because failure is expensive. However, when the cost of inference is close to zero, the cost of experimentation and failure also becomes zero. Agents will be able to work together in teams to explore hundreds of possible solutions to your most pressing problems without costs spiraling out of control.

  • Phase 3 (Symbiotes): expected by 2028-2030. With the cost of training and inference close to zero, we can build hyper-personalized agents (called symbiotes) that are paired with individuals for life, continuously learning and evolving with them. If you are a professional (e.g. a surgeon, engineer or architect), your symbiote learns from you until it becomes a professional in the same field and can assist you in your life's work.

  • Phase 4 (Embodied Symbiotes): expected by 2030-2036. To cover the full spectrum of human knowledge and skills, symbiotes will gain physical embodiments outside phones and computers, such as humanoid robots, self-driving cars, drones, submarines, etc. Your embodied symbiote allows you to be in multiple places at once and carry out physical tasks (e.g. a mechanic symbiote can repair multiple cars at the same time under the supervision of its human partner).

Some of the features required to achieve the above vision already exist in prototype form within AmaniGPT today (e.g. it is natively multimodal which should be useful when building symbiotes that can see, hear, speak, and move).

The mission over the next 10 years is to validate the architecture, optimize it for speed and reliability, and build an accessible ecosystem so everyone can contribute. We welcome you to join us on this ambitious journey and help us build AI that makes our daily lives more productive, enjoyable, and meaningful!

Thriving

The greatest existential threat facing humanity has been the same since day one: unsustainable growth.

We as a species are growing faster than our environment, infrastructure, schools, hospitals, and courts can keep up with. If 100 people are born today, they all immediately need food, healthcare, and education, but it takes 20+ years for any of them to become farmers, doctors or teachers. By the time they are ready to contribute, millions more people will have been born. The math simply doesn't work.

The only way we have survived this long is through increasingly advanced forms of automation:

  1. Tools: we started by creating tools that extended our physical abilities (e.g. hammers and axes).
  2. Machines: tools can't work without us, so we created machines that do more work for us (e.g. washing machines).
  3. Software: machines are limited to specific tasks, so we created software that is easier to update for new tasks.

Software is beginning to reach its limits because it can only do what it is programmed to do and it takes too long to adapt to new tasks or inputs that weren't anticipated by its programmers. As outlined in the roadmap, AI allows us to build software (in the form of agents and symbiotes) that can learn and adapt to new tasks "just in time" without manual reprogramming.

AI allows us to do something that has never been possible: instant replication and global distribution of knowledge and skills. This is how 100 people born today have immediate access to 100 farmers, doctors, and teachers. This is how a planet with billions of people living in cities avoids food riots, disease epidemics, and surging crime rates. AI is how humanity survives and thrives over the next 100 years.

As the AmaniGPT roadmap unfolds, we plan to engage with people at all levels of society to gather thoughts, help leaders define effective policies, and ultimately build a world that our children will be proud to inherit.

Stay tuned for more details on this topic in the coming months, especially regarding the COCO (Capability, Objectives, Constraints, Oversight) blueprint for building, evaluating, and deploying AI systems that operate with one prime directive: to protect and enhance the dignity of human life.

In the same vein, we will also share our Marshall Plan for AI: a novel approach to ensuring people at risk of job displacement are able to transition into new jobs so they can continue to thrive in an AI-powered world.

Contact

Eric Mutta (the creator of AmaniGPT) is a tech entrepreneur and software engineer with 27 years of experience. For media, licensing, and pre-order inquiries, please contact: eric@amanigpt.com

About

AmaniGPT: fast, offline, unlimited AI without GPUs πŸš€

Resources

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors