Resourceful Startups Prove AI Model Building is Faster and Cheaper Than You Think

If you follow the media coverage of companies like OpenAI and Anthropic, it’s easy to assume that building or fine-tuning AI models requires a budget in the billions and the resources of a major corporation. However, at Y Combinator (YC), we’re seeing a different story. Over 25 YC companies have successfully trained their own foundation models or fine-tuned existing ones, often within just three months during the YC batch. These startups are leveraging a combination of founder resourcefulness and YC’s support ($500k funding, $1m+ in cloud credits, and dedicated GPUs) to build AI models that tackle complex problems in various industries.

Technical Tricks and Creative Architectures

These companies have employed several smart technical tricks to reduce computational costs and improve efficiency:

• Optimized Model Architectures: By designing more efficient neural network architectures, these startups can achieve similar performance with less computational power.
• Industry-Specific Insights: Applying domain-specific knowledge to fine-tune models on smaller, more relevant datasets reduces the need for vast amounts of generic data.
• Transfer Learning and Pre-training: Leveraging pre-trained models and transferring learned representations to specific tasks can significantly cut down training time and resources.

Notable YC Companies

Here are a few standout examples of what these companies have achieved:

Atmo (atmo.ai)

What They Do: AI-powered meteorology for countries, militaries, and enterprises.

• Technical Highlights:
  • Developed models that provide more accurate weather predictions than the current state-of-the-art.
  • Achieved this with lower computational costs by optimizing model architectures and using industry-specific data.

Can of Soup (canofsoup.com)

What They Do: An app for creating AI-generated photos of you and your friends in imaginary situations.

• Technical Highlights:
  • Built and launched their first generative model during the YC batch.
  • Utilized transfer learning to fine-tune a pre-trained model on a smaller, curated dataset.

Deepgram (deepgram.com)

What They Do: APIs for ultra-fast speech-to-text transcription and natural-sounding text-to-speech.

• Technical Highlights:
  • Developed highly optimized neural networks for real-time processing.
  • Achieved low latency and high accuracy by leveraging efficient model architectures and cloud resources.

Diffuse Bio (diffuse.bio)

What They Do: Building foundation models in biology to design new proteins for vaccines and therapeutics.

• Technical Highlights:
  • Employed advanced deep learning techniques to predict protein structures accurately.
  • Utilized large-scale simulations and industry-specific data to fine-tune their models.

Draftaid (draftaid.io)

What They Do: AI to help engineers and designers create CAD drawings from 3D models.

• Technical Highlights:
  • Developed a model that converts 3D designs into detailed fabrication drawings.
  • Optimized the model for real-time performance using efficient neural network architectures.

Edgetrace (edgetrace.ai)

What They Do: Takes huge video datasets and generates high-quality, low-latency summaries and insights.

• Technical Highlights:
  • Developed models that can process large video datasets efficiently.
  • Utilized advanced compression techniques to reduce storage and computational costs.

Inspiration for Founders

These examples demonstrate that building AI models is more accessible than ever. With the right combination of technical know-how, resourcefulness, and support, startups can achieve impressive results in a short time frame. We hope this list inspires more founders to explore the potential of AI and contribute to advancing the field.

For more insights into how these companies are achieving their goals, check out the latest episode of the Lightcone podcast.