Frequently Asked Questions
Answers to common questions about Bread
Frequently Asked Questions
General Questions
Q: What is Prompt Baking?
A: Prompt Baking is a breakthrough technique that converts natural language prompts into permanent weight updates for AI models. It minimizes KL divergence to embed behaviors durably, enabling continual learning without catastrophic forgetting. Read more here.
Q: How is Bread different from ChatGPT or other AI assistants?
A: Unlike one-size-fits-all models like ChatGPT, Bread lets you create custom models that permanently learn from your inputs via Prompt Baking. Your model evolves with your expertise or brand voice, rather than relying on temporary prompts or RAG. Deploy via API for scalable, tailored AI.
Q: Is Baking the same as finetuning?
A: No.
Q: Do I need technical expertise to use Bread?
A: Not at all! Our Brand Voice AI and Expert Learning AI are designed for non-technical users like creators and business experts. The Git for Weight Space platform is developer-friendly but includes intuitive tools. If you can describe what you want, you can bake it.
Q: How much does Bread cost?
A: Pricing varies by product: Git for Weight Space is credit-based (with free credits for developers/researchers); Brand Voice AI is $1,000/month (all-inclusive); Expert Learning AI is custom-priced based on your needs. Contact us at contact@aibread.com for details.
Q: Can I try Bread for free?
A: Yes! We offer free credits for Git for Weight Space to developers and researchers. Sign up to get started, or contact us for a demo of our other products.
Q: How do I get in contact with Bread?
A: Reach out via contact@aibread.com for general inquiries, sales, or demos. For job opportunities, email careers@aibread.com. We're also active on Twitter, LinkedIn, GitHub, and Discord.
Product-Specific Questions
Git for Weight Space (Developer Platform)
Q: How do credits work in Git for Weight Space?
A: Credits are used for baking operations—each bake consumes credits based on complexity (e.g., prompt length and model size). We provide free credits to developers and researchers to experiment. Track and purchase more via your dashboard.
Q: Can I version-control and compose multiple bakes?
A: Absolutely! Git for Weight Space treats bakes like code commits: stack, fork, and merge them for composable updates. This is ideal for building production systems iteratively.
Q: What models can I use with Git for Weight Space?
A: We support a variety of open-source models such as Llama, Qwen, and GLM, and select closed source models. Bake custom prompts to fine-tune them for your needs, then deploy via our API.
Brand Voice AI (For Creators & Brands)
Q: How does Brand Voice AI ensure my model matches my unique voice?
A: We bake your writing samples, style guidelines, and feedback directly into the model's weights. This creates a permanent, authentic voice that never reverts to generic AI perfect for marketing, social engagement, and lead gen. No more "brand suicide" from one-size-fits-all models exhibiting prompt decay.
Q: Can I use Brand Voice AI for social media automation?
A: Yes! It excels at authentic engagement on Twitter, Discord, Reddit, and more. Your baked model generates content in your voice, growing your presence while maintaining trust.
Q: What if I need to update my brand voice later?
A: Easy—provide new prompts or examples, and we'll bake updates. The model evolves without losing prior knowledge, thanks to Prompt Baking's composability.
Expert Learning AI (Automate Expert Processes)
Q: What industries is Expert Learning AI best for?
A: It's versatile for any field with specialized knowledge, like insurance claims, legal reviews, coding, or sales processes. If you have an expert who can demonstrate once and provide feedback, we can automate it.
Q: How does the 3-step learning process work?
A: 1) Demonstration: Your expert shows the process. 2) Questions: The model probes edge cases & counterfactual branches of the action tree. 3) Practice: The AI tries tasks, and corrections get baked in. This embeds nuanced judgment permanently, with human-like sample efficiency (one correction often suffices).
Q: Can Expert Learning AI handle complex, low-data tasks?
A: Yes! Unlike RL (which needs thousands of examples), it learns from single demonstrations and feedbacks, making it ideal for rare expertise where data is scarce.
Technical Questions
Q: How does Prompt Baking compare to RAG (Retrieval-Augmented Generation)?
A: RAG pulls in external data at runtime but doesn't change the model, and is therefore temporary and context-limited. Prompt Baking permanently updates weights, embedding knowledge durably without repeated retrieval. It's more efficient for continual learning and avoids "prompt decay."
Q: How much can I bake into a model?
A: Limits depend on the product—Git for Weight Space supports composable bakes (we've tested 1,000+ documents without issues); Brand Voice AI handles full brandbooks & guidelines, plus product & marketing copy; Expert Learning AI scales with expert sessions. No hard word limits, but practical caps ensure quality.
Q: Why does Prompt Baking mitigate KL divergence and catastrophic forgetting?
A: By minimizing KL divergence between the prompted and baked distributions, it matches behaviors the model is already capable of, preserving prior knowledge. Unlike SFT or RL (which can shift distributions arbitrarily and cause forgetting), baking stacks updates composably, shown through internal tests with 1,000+ documents retaining base abilities.
Q: How does Prompt Baking compare to fine-tuning or RL?
A: Fine-tuning requires massive datasets and risks forgetting; RL needs thousands of trials and struggles in real-world sparsity. Baking is sample-efficient (one prompt/correction), self-supervised, and avoids forgetting by projecting prompts into weight space naturally.
Q: Is my data secure with Bread?
A: Yes—all data is encrypted in transit and at rest. We don't share your prompts, documents, or models. For Expert Learning AI, expert sessions are handled confidentially.
Q: Can I integrate Bread models with my existing tools?
A: Definitely! Our API is OpenAI-compatible, making it easy to embed in websites, apps, or workflows. Export options are available for higher tiers.
Company and Support Questions
Q: Are you hiring?
A: Yes! We're growing and looking for talented engineers, researchers, and product specialists passionate about AI learning. Email your resume to careers@aibread.com or check our careers page for openings.
Q: Can I monetize models I create with Bread?
A: We're building marketplace features for sharing and selling baked models, and offer a Value-added-Reseller model. In the meantime, deploy via API and charge your users.
Q: What if I need custom support or enterprise features?
A: Contact us at contact@aibread.com for tailored solutions, like offline deployment or large-scale baking. Enterprise plans include dedicated support.
Q: Where can I learn more about the research behind Bread?
A: Check out our blog and the original Prompt Baking paper on arXiv. We also share updates on GitHub and Discord.
If your question isn't answered here, reach out to contact@aibread.com, we're here to help!