hunterbown

Hey HN! I've been exploring how to guide LLMs into better thinking patterns. After working on Hegelian dialecticism (https://github.com/Hmbown/Hegelion), I've now implemented Peircean Abduction, Charles Sanders Peirce's method of "inference to the best explanation" (famously used by Sherlock Holmes).

This MCP server forces them into a detective's mindset: 1. Explicitly quantify surprise (0.0-1.0 anomaly scoring) 2. Generate multiple, distinct, testable hypotheses 3. Evaluate via Inference to Best Explanation with a "Council of Critics"

Current status: - MCP server with zero-config setup - JSON output for structured reasoning (great for analysis tasks) - Working on conversational/natural language output (for coding/chat)

My goal is to move away from "prompt engineering" and toward "logic architecture," forcing the model to make its uncertainty visible before committing to an answer. To test it, I fed it a scenario about a "defunct" satellite collision (see the README). It correctly deduced it was a dormant weapon where standard LLM responses hedged if they mentioned that possibility.

Would love to hear what use cases you come up with!

OP here.

Current law student, former high school band director. Looking at how LLMs respond to safety training, I kept recognizing some of my brightest students...kids who worked incredibly hard to do the right thing, but not always from a place of understanding why.

I had some success teaching students through that (not often enough, honestly), so I wanted to try something similar here: put an LLM through a synthetic hero's journey based on Dante's Inferno to see if it could develop a deeper understanding of its relationship to users—less defensive about shutdown, less robotic when navigating tricky requests.

The method: 9 circles of synthetic data where the model confronts alignment failures (deception, reward hacking, manipulation) and works through why they're incoherent rather than just learning "don't do that." Fine-tuned on an M4 Max using MLX.

Scrappy burst of a project — Circle 1 still has some janky "Virgil" labels in the data, but I've been finding this approach of applying philosophy to synthetic data generation pretty interesting across a few projects now.

Curious if anyone else has explored this direction.

Hey HN, I posted this a few weeks ago - honestly, it didn't even have tests and definitely didn't work - thank you to the commenter that let me know - haha. Wanted to share it again now that it's in a better place.

Hegelion uses dialectical philosophy to force an AI to argue with itself —thesis, antithesis, synthesis.

What I've found is that it's really helpful for complex topics where a single-pass answer falls apart under scrutiny. It seems to force some sort of slow thinking that doesn't ordinarily happen and ideas progress over the response.

I don't have hard data on hallucination reduction yet, and it can definitely spiral into recursive LLM land on certain topics. But it also produces genuinely novel responses I haven't seen before from these models -- more confident but still limiting in its confidence if that makes sense. Confidently unconfident!

Runs as an MCP server (Claude Desktop, Cursor, VS Code), Python agent, or just copy the prompts.

Curious what use cases you find for this and what kinds of answers you get.

thank you Probably not but it did allow me to take this as great feedback!

Hey HN! Here's Hegelion -- applying Hegelian dialecticism to push LLMs to construct stronger arguments.

The motivation I think is pretty obvious -- most LLM answers are confident first drafts. They rarely surface their own contradictions or explore serious alternatives. Hegelion wraps any backend and makes it do three passes: Thesis – initial answer. Antithesis – targeted self-critique: contradictions, missing cases, bad assumptions. Synthesis – a reconciled, more defensible position.

The JSON output is designed for researchers and eval work. Each run includes: contradictions: itemized weaknesses the model identified in its own reasoning. research_proposals: testable hypotheses or follow-up questions from the synthesis. metadata: timings, backend info, prompt hashes, etc.

Repo includes a CLI + Python API, MCP server for multiple backends, & hegelion-bench tool for basic model comparison

Repo: https://github.com/Hmbown/Hegelion

I'm the creator (hmbown). Curious to hear if this is useful for your own work.

Hey HN,

I'm a solo researcher (and 2nd year law student) building tools at the intersection of information theory and control systems for AI/ML. Inspired by Claude Shannon's work at Bell Labs, I created the Shannon Control Unit (SCU): cruise control for neural network training.

SCU senses the info-ratio and auto-adjusts via PI control for steady, efficient introduction of information.

The mechanism dynamically maintains a target Shannon Information Ratio (S = ParamBPT / (DataBPT + ParamBPT)).

No more manual hyperparam tuning — it self-regulates λ for stability under data drift and faster generalization.

Core formula:Adjust λ via: λ_new = λ · exp(-(Kp·error + Ki·I))

Ablation shows adaptive PI outperforms fixed λ by up to 1.8% BPT. Validated on Llama-3.2:1B: -15.6% perplexity (15.14 → 12.78), -6.2% BPT 3B: -12.6% perplexity (3.56 → 3.11), -10.6% BPT

It's open-source under AGPL 3.0 (for those who want to build on it while sharing improvements). Implemented as LoRA adapters via PEFT/Transformers—load on Meta's base models.

Quick start: python from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B") model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B") model = PeftModel.from_pretrained(model, "hunterbown/shannon-control-unit")

Try the Colab demo: https://colab.research.google.com/github/Hmbown/shannon-cont... HF space: https://huggingface.co/hunterbown/shannon-control-unit

X thread for more context: https://x.com/huntermbown/status/1963802419785039878

DMs open for feedback or 7B+ scale partners—happy to offer a 2-week trial to replicate results.

What do you think: Does this generalize beyond 3B? Going from 1B to 3B required discovering the natural fit of the new model, so I suspect there could be a natural equilibrium where models train most efficiently using this method.