OP
OP
- Joined
- Nov 23, 2008
- Messages
- 2,054
- Reaction score
- 260
Last few weeks has been about strictly proving function by function by triggering both the original binary aswell as the reconstructed versions functions side by side in runtime after initially having completed the entire static analysis of dwarf information.
Here's an image showing the workflow. Most additions are tests for each actual function where each test is adapted to each function separately. This is just from todays prompt and work.
The goal is not to recreate the original but to create a behaviourally accurate version of it, meaning it should behave as the original in all regards for every function. Once we reach 1579 strictly proven functions we will move on to cleaning the code, asan and then stress testing.
Current estimated cost is around 1K usd in LLM licenses alone. I decided that AI with good proof and a strict work structure would yield better results as AI has improved quiet alot.
Here's an image showing the workflow. Most additions are tests for each actual function where each test is adapted to each function separately. This is just from todays prompt and work.
The goal is not to recreate the original but to create a behaviourally accurate version of it, meaning it should behave as the original in all regards for every function. Once we reach 1579 strictly proven functions we will move on to cleaning the code, asan and then stress testing.
Current estimated cost is around 1K usd in LLM licenses alone. I decided that AI with good proof and a strict work structure would yield better results as AI has improved quiet alot.
Last edited:
