AXI framework pre-registered test: per-step cost reduced 4.5 to 9.3 times
Understand this story
Shalendra Gaur releases a pre-registered test report for the AXI framework. The framework keeps the precise state of AI tasks outside the conversation, validates each step's reply with a deterministic gate before applying it, and sends only short new prompts to the model. The report claims that on Gemini 3.8 Flash, open-source models, and NVIDIA T4, L4, and Google TPU v6e, all 1,500 steps maintained precise state, prompts shrank by an average of 169 times, per-step cost fell 4.5 to 9.3 times, and the number of sessions each chip can host increased by up to 185 times (measured by memory headroom). The report says predicted values were hash-locked before each run, and all numbers can be recomputed from saved files.
AI overview based on source reports · Updated 2 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIAXI Framework: The Cure for the Cost Crisis in AI
AXI is a host-side layer that holds an AI job's exact state outside the chat, checks every reply with a deterministic gate before applying it, and sends the model a short fresh prompt at each step. Pre-registered tests on Gemini 3.8 Flash, open models, an NVIDIA T4 and L4 and a Google TPU v6e report exact state on 1,500 of 1,500 steps, prompts 169 times smaller on average, one exact step 4.5 to 9.3 times cheaper, and up to 185 times more sessions per chip. Predictions were hash-locked before each run and every figure can be recomputed from the saved files; the architecture is by Rajnish Choubey and the tests were built and run by Shalendra Gaur, with code at https://github.com/gaurshalendra2412/axi-runtime.
Activity for this story
Not enough continuous observations to show a trend yet.