🏴‍☠️ Norelec

Im all about pizza, crypto, privacy, and idea innovation and embracing technology and creating new projects that will further ourselves.

[🏴‍☠️ Norelec @NorelecSeth]

First STYX benchmark on public datasets:

99.8% retrieval precision
75.7% LLM-judged quality
74% fewer tokens

4,672 questions. Local 14B model. No cloud APIs.

First baseline. Patent pending.

https://t.co/BqUTiDv8Mr

[🏴‍☠️ Norelec @NorelecSeth]

With welcoming arms I'm already part of a competition!

Help donate to the cause of OSDRI.

Not a hand out! A hand UP!

https://t.co/EY7GA4f2RI

[🏴‍☠️ Norelec @NorelecSeth]

@BinxTheCatNFT @MrBeast @salesforce @LoneSharkGames https://t.co/x71pBKy1ed

do you have a discord, Shane Kicked me

[🏴‍☠️ Norelec @NorelecSeth]

.@karpathy You have said the data you train on matters more than the model.

Turns out the context you inference on does too.

We stripped 98% of tokens from 60,900 docs. Quality did not drop.

500 blind judgments. 5 models. A100. Statistical parity with full context.

[🏴‍☠️ Norelec @NorelecSeth]

We track every claim our AI makes and whether it was right.

124 tracked outcomes. 93.15% accuracy.

Most AI systems measure speed and cost. Almost nobody measures how often the AI is actually correct.

If you can't measure reliability, you can't trust automation.

[🏴‍☠️ Norelec @NorelecSeth]

.@AnthropicAI Claude's 200K context window is a breakthrough.

But what you put IN the window matters as much as the window size.

We compressed context by 98% and tested quality: 239-238 across 500 blind judgments.

Same answers. 57x fewer tokens. Maybe less really is more.

[🏴‍☠️ Norelec @NorelecSeth]

Built 21 AI services that govern each other.

This week they audited their own codebase and found 11 bugs I never knew existed.

The system that builds the product is becoming the product.

#AI #BuildingInPublic

[🏴‍☠️ Norelec @NorelecSeth]

Enterprise AI on Azure means token bills at scale. Every department, every workflow, every API call pays a context tax.

86% of that context is noise. Tested across 10K+ real data points. No additional infrastructure needed.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

Llama models power open-source AI worldwide. Smaller context = faster inference on consumer hardware.

86% token reduction without embeddings or vector search. Tested on 10K+ data points. Works with any Llama variant.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

.@AnthropicAI What if giving Claude LESS context produced BETTER answers?

We tested it. 500 blind judgments across 5 models:
- STYX compressed context: 61% win rate vs GraphRAG
- vs full raw context: statistical parity

98% fewer tokens. Equal or better quality.

#Claude #AI

[🏴‍☠️ Norelec @NorelecSeth]

W&B tracks every experiment. Imagine tracking token efficiency too — how much context your pipeline wastes per query.

86% of retrieval tokens are noise. Tested across 10K+ real data points with tiktoken verification.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

Vector databases solve retrieval. But retrieval was never the hard problem — relevance is.

Built a context engine that extracts only what matters. 86% fewer tokens. No embeddings needed. Tested on 10K+ real data points.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

Chroma makes vector search simple. But what if the embedding step itself is the bottleneck?

Tested an approach that bypasses embeddings entirely — 86% token reduction, 20x on unstructured content. Zero infrastructure overhead.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

Vector search retrieves by similarity. But similarity ≠ what the model needs for its next decision.

What if you skip embeddings entirely and extract only decision-relevant state? 86% token reduction. 10K+ data points.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

The best startups find leverage others don’t see. Right now every AI company accepts that context = embeddings + vector search.

That assumption costs 6x more tokens than necessary. I proved it across 10K+ data points.

https://t.co/ZgTEtbn62M

[🏴‍☠️ Norelec @NorelecSeth]

@_mctrinh This might resonate with your work at the intersection of biology and computation — when you're reasoning about immune cell interactions or gene circuits through LLMs, most of the context window is narrative that doesn't affect the answer.

STYX extracts what governs decisions. The benchmark methodology is open if you want to poke at it.

[🏴‍☠️ Norelec @NorelecSeth]

STYX beats GraphRAG on quality. 61% to 36%. 500 blind judgments. 5 models. Zero errors.

98% compression. Same answers as full context. Better than GraphRAG. 57x cheaper.

50K docs. Validated in Python, R, Julia.

Patent pending. Open benchmarks on Github (see reply post)

#AI #LLM

[🏴‍☠️ Norelec @NorelecSeth]

Quality Proof Delivered — STYX Beats GraphRAG 61% to 36%

https://t.co/beHZT3nvZR

[🏴‍☠️ Norelec @NorelecSeth]

https://t.co/ZgTEtbnDSk

[🏴‍☠️ Norelec @NorelecSeth]

Context windows keep growing.
Gemini 1.5: 1M tokens.

But bigger isn't always better.

STYX: Give your model LESS noise, MORE signal.

57x compression means:
- Faster inference
- Lower costs
- Better focus

Tested on 60,900 docs. 90-98% reduction.

@GoogleAI #Gemini