LLMs as MoldMakers; not 3D Printers
harsha kokel
Lightning Talks by Women at Almaden by AWING
June 5th, 2025
MoldMakers vs 3D Printers
Use cases?
- 🧱 Make a mold once
- then replicate fast, is consistent, and scalable
- 🖨 Print a custom object every time
- bespoke, unique, slow
LLMs as 3D Printers
Generate each piece from scratch
Engineer Every Response
- ⏳ Time and token costs
- 🔁 Inconsistent format, tone, and logic
- 🧩 Redundant efforts
LLMs as MoldMakers
Build Once, Use Many Times
Use the LLM to:
- Create reusable templates, frameworks, or data structures
- Generate code, logic flows, or prompts as "molds"
Benefits:
- ⚡ Faster inference
- ✅ Consistency and reliability
- 💲 Lower costs
A historical perspective
An efficient past:
- General-purpose solvers
- SAT Solver, Domain-independent Planner, ASP Solver, theorem provers, CSP solvers etc
- Built once → Solved many problems
- Modular, composable, and computationally efficient
Image generated by ChatGPT
Transformative middle:
- AI-inspired Assessment Tools
- Epic Sepsis Model, AI-QRISK, etc
- Retrieve pattern from data to build tools
- Used training once, applied many times
An inefficient present:
- Fragile outputs
- High compute cost for every minor variation
Planning with LLMs
as 3D Printers
Planning with LLMs
as MoldMakers
Thought of Search
LLMs as 3D Printers MoldMakers
Generate Once , Solve Many
Thank you