Exploring GPTs: ChatGPT in a trench coat?
OpenAI’s new “GPTs” — customizable ChatGPT-based bots that can be given instructions, tools, and private knowledge bases — are seen as a powerful but uneven step toward making AI agents accessible to non‑technical users. Commenters compare them to the Assistants API, debate how much they really offer beyond clever prompt engineering, and note practical issues such as weak retrieval performance, opaque hidden prompts, and awkward integration with external data and OAuth. Many still view the main innovation as UX and distribution: GPTs lower the barrier to building and sharing specialized agents, while raising fresh concerns about data collection, corporate control, and AI mediating more customer and personal interactions.
GPTs vs Assistants API / Technical Differences
- Many see GPTs as “ChatGPT with a pre-prompt,” i.e., custom system instructions plus optional tools and files.
- Assistants API is described as lower-level: more control over threads, context initialization, and UI, but you must build the frontend and handle polling.
- Some report only modest extra control with Assistants in practice; OpenAI still drives conversational flow.
- One noted difference: multiple Assistants can be attached to a single thread, while GPTs are single-model.
Actions, Function Calling, and Auth
- GPTs can call external APIs via “actions” backed by OpenAPI specs, functionally similar to function calling.
- OAuth is available for GPT actions, but constrained by domain rules; setup can be tricky.
- POST actions often require repeated “Allow/Deny”; a special header (
isConsequential: false) can reduce prompts.
RAG / “Knowledge” and File Handling
- GPT “knowledge” is effectively RAG: files are chunked, embedded, and stored in a vector database (likely Qdrant for some flows, though others mention Azure services).
- Behavior seems size-dependent: small files may be inlined into prompts; large collections often work better merged into a single text file.
- Several users report poor or inconsistent retrieval quality, citation control, and indexing failures, especially with many or large files.
- Raw text files tend to work better than complex formats; PDFs/Markdown are hit-or-miss.
Prompt Transparency and Control
- Strong desire for a “view source” option for GPT prompts and configuration; many are wary of hidden instructions and unknown APIs.
- Users note GPT prompts can frequently be leaked with clever queries; attempts to hide them are often defeated.
- Some argue that “view source” would mainly serve power users but could drive community-driven improvement.
Use Cases and UX Impressions
- Reported uses include: retro game consoles, domain-specific analysts, training/workshop automation, RAG over technical documents, personal assistants, and experimentation with JavaScript Code Interpreter.
- GPTs are praised as a massive UX win for non-technical users who struggle with system prompts and custom instructions.
- Others find them unnecessary “toys” compared to directly prompting the base model.
Risks, Data, and Business Concerns
- Concerns include: shovelware content, corporate customer-service deflection, dehumanizing efficiency, and OpenAI harvesting uploaded documents and user creativity as a competitive moat.
- It is unclear to some whether GPT-uploaded documents are used for model training; API uploads are said not to be.