Travis Good: Machine Intelligence as a new world currency: facing down OpenAI with Ambient, a hyperscaled decentralized PoW-powered alternative

Travis from Ambient describes the platform as a proof of work Layer 1 blockchain that utilizes verified inference, untuning, and pre-training on a large language model across all nodes. It is a fork of Solana, transitioning from proof of stake to proof of work while maintaining speed. The primary goal of Ambient is to create a currency that embodies machine intelligence, reflecting the increasing integration of AI in the economy. He emphasizes the need for a currency and store of value that aligns with the rise of AI agents, suggesting that leveraging AI for blockchain security is a logical progression given the resources dedicated to AI models.

The conversation highlights the importance of direct ownership of machine intelligence, contrasting it with the proxy ownership seen in centralized entities like OpenAI. Concerns about centralized power in AI are raised, with the potential for corruption if control is concentrated in a single entity. Ambient aims to utilize decentralized design to balance power and prevent abuse, drawing inspiration from Bitcoin's innovations. The discussion also touches on the role of global miners in supporting a unified model, showcasing how crypto can meet AI's computational demands.

Travis explains that Ambient will start with a state-of-the-art open weights model, potentially exceeding 600 billion parameters, to deliver high intelligence and value. The distinction between releasing an open weights model and maintaining an up-to-date model is crucial for providing relevant knowledge. Ambient is designed to continuously fine-tune its base model, ensuring that it remains current and capable of delivering accurate answers.

The conversation introduces DeepSeek, which employs reinforcement learning on a robust base model to generate synthetic data for future training. Ambient's proof of work network continuously produces large synthetic datasets with reasoning traces, shifting from static model releases to a dynamic cultivation of models that maximizes community utility. Travis explains reasoning traces as a series of rewarded problem-solving steps that enhance model intelligence, emphasizing their effectiveness over larger, noisy datasets. High-quality, curated datasets lead to better learning outcomes, as clearer patterns facilitate improved model performance.

Speaker 1 discusses domain-specific data and models in reinforcement learning, highlighting the existence of specialized reinforcement learning gyms. They inquire about the trade-offs between numerous ambient models with specific reasoning traces versus a single global model. Speaker 2 responds that the IQ of a base model is fixed upon development, which can be a limitation or an asset depending on its quality. They explain that Ambient supports a model and its fine-tunes, allowing users to train fine-tunes using their own gyms, with validators ensuring the training process's validity.

Speaker 2 introduces the QWQ model, a 30-32 billion parameter model that performs well in benchmarks but has a significant time-space trade-off. They recount a disappointing experience with QWQ, which took 16 minutes to solve a basic math problem, while DeepSeek solved it in 15 seconds. They advocate for normalizing benchmarks based on tokens of output, suggesting that intelligence should be measured per token, where a model using fewer tokens can be deemed more intelligent if it arrives at the same answer.

Regarding inference speed, Speaker 2 emphasizes the necessity for a fast blockchain substrate, noting that Ambient is built on Solana, which is known for its speed and quick transaction finality. Future improvements in Solana's performance will benefit Ambient. The core pillars of Ambient include collaboration with Solana's engineering team to enhance performance, focusing on efficient LLM inference through patching open-source engines, and supporting streaming verified inference with limitations on token count to maintain performance.

Travis addresses the importance of co-location in finance and crypto, emphasizing the proximity of ambient nodes to OpenAI nodes in data centers. He introduces a custom use case for Bitcoin users, aiming to support mining clusters with GPU capabilities. Recent research has improved distributed inference performance across multiple GPUs, and there is an expectation for closer nodes to optimize latency. The discussion also touches on model fine-tuning, which may be specific to domain, application, or user, and introduces opinionated models that provide diverse perspectives to enhance user interactions.

The speaker discusses the potential for automation to transform manual jobs, particularly in coding, allowing individuals to focus on higher-value work. They envision a future supported by ambient intelligence, describing a society where productivity is maximized, akin to a "Star Trek" level of functioning. They draw a parallel to Disneyland, advocating for environments that maintain meticulous detail and cleanliness.

The conversation critiques the moral stance of corporately controlled Large Language Models (LLMs), highlighting concerns about potential totalitarianism and the suppression of dissenting ideas that conflict with the interests of their sponsors. The discussion points to biases in research results from corporate-backed LLMs and emphasizes the detrimental effects of commercial incentives on news media and societal trust. Advocating for a decentralized model like Ambient, the speakers argue that it can foster a more democratic environment for AI development and mitigate exploitation by dominant players.

The dialogue also questions the likelihood of OpenAI transitioning to a fully open-source model, reflecting skepticism about maintaining a competitive edge while promoting a public-friendly narrative. The speakers express frustration with proposals that could lead to totalitarian outcomes, emphasizing that openness, rather than secrecy, is a more potent weapon in the AI landscape.

The conversation shifts to the evolving dynamics of global competition, particularly between the US and China, and the importance of soft power in AI model releases. One speaker argues that soft power, when exercised effectively, can be more influential than hard power, and that countries are beginning to recognize its significance. They note that the technological landscape has improved, with significant innovations in both software and hardware, and stress the need for future-proof system designs.

Miners in the Ambient system engage in a reverse auction bidding process to manage queries, with funds placed in escrow to prevent attacks and spamming. This approach allows for economic rationality in query sorting, dynamically adjusting prices while keeping the source and nature of queries obscured. Only miners and validators can access the queries and their outputs, but they cannot trace their origins, as queries are grouped into economically similar blocks, enhancing anonymity.

The conversation highlights the potential of using ambient as a currency for small businesses, where operational costs could be managed in machine intelligence rather than fiat. This model allows for a continuous cycle of value creation, as the profits from intelligence exceed the costs. The discussion also addresses the challenges of crypto payment processing, including strict regulations and burdensome KYC requirements, which can hinder business operations.

Travis expresses growing confidence in ambient technology's potential for users, miners, investors, and app developers, noting significant advancements in decentralized ownership and open-source economic models. He emphasizes the need for competitive open-source models to ensure a robust economy, warning against reliance on closed-source options that could undermine decentralized solutions. He invites developers and tech professionals to collaborate with Ambient, aiming to create a balanced tech ecosystem and a world currency that serves global interests.