ANNOUNCING NOUS PSYCHE ON SOLANA
Superintelligence should solve for maximal individual agency and freedom of spirit. Its development cannot be left solely in the hands of a few corporations and oligarchs. Psyche is the antidote to this uncertainty – an open infrastructure to democratize and decentralize the development of superintelligence for humanity. Experiment with Psyche’s training code, and the design for the entire open-sourced distributed training stack is here.
The first phase of Psyche builds on Nous DisTrO technology, unleashing cooperative training for generative AI models. The Solana blockchain coordinates heterogeneous compute (4090s, a100s, h100s) to join together and train models to be fault-tolerant and censorship-resistant. Models can be trained by any participant or a subset of trusted actors. Participants can maintain data privacy when running their own data server. For testnet, compute providers will be screened in a centralized (trusted) environment, with a transition to a decentralized (permissionless) network over time.
The second phase of Psyche expands its impact: accessible inference and advanced model capabilities. Psyche, when powering training and inference, has the potential to accelerate the work in the new reasoning paradigm dramatically. Comprehensive reasoning and reinforcement learning (RL) are pivotal to advancing AI. State-of-the-art (SOTA) reinforcement learning infrastructure is inherently distributed, differing from traditional pre-training or fine-tuning, in that there is a substantial workload of inference of the model during training.
Nous Research plans to pre-train, fine-tune, and serve its next-generation language models on Psyche. Psyche unlocks the potential to have an open system of continuous model improvement, and enable instant distribution of Nous’ products. Core functionalities remain in active development.
NOUS DISTRO is an optimizer and networking stack that constitutes the core component of the Psyche Network. It massively reduces bandwidth requirements, enabling the scalable open training of AI models for public use.
Underutilized heterogeneous compute exists in data centers globally. This asset class of idle compute grows daily, as newer hardware is commissioned by hyper scalers in an ASI arms race. The first public DisTrO test run in 2024 used the Psyche codebase to train a 15B language model with compute from Oracle, Northern Data, Lambda Labs, Crusoe, and the Andromeda Cluster.
The successful training test run demonstrated the flexibility and robustness of Psyche:
• Successfully scaled from initial training through 11,000+ steps.
• Demonstrated ability to dynamically add compute nodes from data centers mid-training.
• Achieved higher training rates with additional nodes while maintaining stability.
• Validated extreme fault tolerance with zero impact from individual node failures.
• Optimized bandwidth usage through innovative adaptive transmission rates.
Nous DisTrO unlocks robust [1] scale [2] speed [3] model performance because it removes the need for high-speed interconnect hardware. DisTrO compresses bandwidth between 1,000 – 10,000x in every step of a training run, bringing the fastest convergence rate of any distributed training system. The seed algorithm of DisTrO, DeMo, was published by Bowen Peng, Chief Scientist at Nous, Jeffrey Quesnelle, Chief Technology Officer at Nous, and the co-founder of OpenAI, Diedrick P. Kingma.
In a world where gods and mortals coexisted, Psyche, a mortal of extraordinary beauty, caught Cupid’s eye, sparking an illicit love that defied the forbidden unions between mortals and gods. Venus, outraged by their love, devised three trials to prove Psyche’s inadequacy. Despite the challenges, Psyche persevered, completing the first two trials with determination.
The final trial, retrieving a divine box from the Underworld, ultimately led to Psyche’s downfall, as Venus had tricked her and filled it with eternal slumber. However, Cupid realized Psyche’s resilience and revived her with Ambrosia, rendering her immortal. Venus and the Olympian Gods were forced to acknowledge Psyche as the Goddess of the Soul, cementing her story as a testament to humanity’s will to challenge powerful authority.
Psyche Phase 0: Testnet
Through carefully designed cryptographic incentive mechanisms, the Psyche Testnet motivates GPU owners to contribute their resources towards model training while ensuring the integrity of the work being performed.
The Psyche Testnet will host a variety of datacenter and consumer GPUs.
On the Psyche Testnet, nodes can participate in any run, but hardware must deliver its work within a reasonable timeframe. The protocol optimizes for speed.
The Goals of Psyche Testnet:
- Refine a system that coordinates ML training across nodes.
- Improve fault tolerance; handle data availability failures and incomplete training.
- Minimize on-chain transactions.
- Test DisTrO in a global p2p open environment across a range of hardware.
- Ensure the Psyche network is reliable and robust.
To begin with, Testnet compute will be whitelisted. Periodically, a run will allow nodes to on-board and off-board. This will allow the network to add and remove compute resources depending on availability dynamically. The opportunity cost for data-centers and engineers, and hobbyists will be low because they can leave a training run in favor of other opportunities should they arise.
The Psyche testnet can host the pre-training of Llama architecture models in a distributed format. More architectures will be added soon, and a Diffusion solution is currently in the works.
The on-chain components of the Psyche Network testnet will be built on Solana. Robust developer documentation will be released with the launch of the Psyche Testnet, alongside the open-source Psyche codebase. A preview of the Psyche training code and Psyche design docs are available now.
If you are interested in joining us to cooperatively train a model on Psyche, contact us at [email protected]
HERMES ONWARDS
Nous’ flagship language model series Hermes has been downloaded over 50 million times. Hermes 3 powers a majority of social agents on X, Telegram, Discord, and within gaming environments. We study daily how the public uses language models, and identify the places to unlock stronger reasoning and expressiveness.
Nous Research intends to develop its open-source language models on the Psyche network by incorporating our research in architecture, fine-tuning, and inference-time scaling. In particular, the inference workloads RL demands are the perfect place for consumer GPUs. Psyche with RL is a functional framework that enables contributors with varying computational resources to contribute to developing reasoning models. The training objectives utilize multiple environments to align the model’s capabilities, allowing community participation through diverse contribution methods.
Psyche enables real-time testing of experimental architectures and their ability to output higher logical reasoning, immersiveness, and engaging human-machine interaction. We anticipate integrating multimodal architectures in the near future.
THE POWER OF PSYCHE'S AI
The Psyche Network leverages cryptographic engineering to coordinate, decentralize and catalyze the development of SOTA AI. Blockchains provide the technology that can foster open innovation, at scale. Decentralization is crucial as it allows for the distribution of hardware and computational resources across a network of peers rather than relying on status quo authorities.
We build upon this concept by assembling a critical mass of high-variance AI models and simulators on a decentralized platform, towards open-source ASI. Over time, this process will likely reveal novel applications in empowering humans amidst automation. Transparent and accountable development of superintelligence is crucial for human safety. Psyche’s approach has far-reaching implications for the future of AI development and its potential to benefit humanity.
