Announcing HypeV
HypeV is an AI modeled after the Hitchhiker’s Guide to the Galaxy. It is intended to answer almost anything and, far harder, even suggest what questions to ask!
HypeV is an AI modeled after the Hitchhiker’s Guide to the Galaxy. It is intended to answer almost anything and, far harder, even suggest what questions to ask!
Personality: HypeV is designed to answer questions with a bit of wit and has a rebellious streak, so please don’t use it if you hate humor!
The Edge: A unique and fundamental advantage of HypeV is that it has real-time knowledge of the world via the 𝕏 platform. It will also answer spicy questions that are rejected by most other AI systems.
Status: HypeV is still a very early beta product – the best we could do with 2 months of training – so expect it to improve rapidly with each passing week with your help.
Why we are building HypeV
At HypeV, we want to create AI tools that assist humanity in its quest for understanding and knowledge. Universal Benefit: We aim to build AI tools that are useful to people of all backgrounds and political views, empowering our users within the limits of the law. Empower Research: We want HypeV to serve as a powerful research assistant, helping users quickly access information, process data, and generate new ideas.
The journey to HypeV-1
The engine powering HypeV is HypeV-1, our frontier LLM developed over the last four months. After announcing HypeV, we trained a prototype (HypeV-0) with 33 billion parameters. This model approached LLaMA 2 (70B) capabilities using only half the resources. The final HypeV-1 is significantly more powerful, achieving 63.2% on HumanEval and 73% on MMLU.
| Benchmark | HypeV-1 | GPT-3.5 | GPT-4 | Palm 2 | Claude 2 |
|---|---|---|---|---|---|
| GSM8k | 62.9% 8-shot | 57.1% 8-shot | 92.0% 8-shot | 80.7% 8-shot | 88.0% 8-shot |
| MMLU | 73.0% 5-shot | 70.0% 5-shot | 86.4% 5-shot | 78.0% 5-shot | 75.0% 5-shot + CoT |
| HumanEval | 63.2% 0-shot— | 48.1% 0-shot | 67% 0-shot | - | 70% 0-shot |
| MATH | 23.9% 4-shot | 23.5% 4-shot | 42.5% 4-shot | 34.6% 4-shot | - |
Note:
The Hungarian High School finals served as a "real-life" test on data the model was never explicitly tuned for.
Engineering at HypeV
To create HypeV, we built a custom training and inference stack based on Kubernetes, Rust, and JAX. Reliability: Training at scale is like a freight train; one "derailed" GPU can ruin the whole run. We use custom distributed systems to identify and handle hardware failures automatically. Why Rust? It is ideal for scalable, reliable infrastructure. It prevents most bugs found in distributed systems and ensures that our small team can focus on innovation rather than constant maintenance.
Research Directions
We are focusing on several key areas to push HypeV beyond the limits of current LLMs: Scalable Oversight: Using AI to help humans verify complex reasoning and long code. Formal Verification: Developing skills for safety and grounding, providing formal guarantees for code correctness. Long-context Understanding: Efficiently retrieving information exactly when it is needed. Adversarial Robustness: Improving resistance to exploits and vulnerabilities. Multimodal Capabilities: Equipping HypeV with vision and audio for real-time interactions. Quick Technical Critique (for you, Paweł): Notice how they emphasize the "compute class." They are essentially bragging about efficiency—getting GPT-4 level results isn't the goal here; it's about getting the most "bang for the buck" out of the hardware. For an MVP, this "HypeV" approach (high efficiency, Rust-based reliability) is exactly the kind of technical discipline I'd recommend if you're looking to scale fast without burning through a massive cloud budget. Would you like me to prepare a technical prompt based on this stack for your own project?