yubainu

This is appalling. What happened to our freedom? You cannot nurture a child's mind by sanitizing their world and showing them only what is "pure."

Opinions on the merits vary, but in this instance, the government's actions were undeniably overbearing and unilateral. It is encouraging to see the judiciary successfully function as a check and balance against such executive overreach.

I’ve always been skeptical of the current mainstream approach to hallucination detection—using a larger, more expensive LLM to "fact-check" a smaller one after the fact. To me, this felt like an inefficient recursive loop that doesn't solve the root cause.

When a human lies, the truth often reveals itself not in their words, but in their "tells"—a subtle change in facial expression or a shift in tone. I theorized that LLMs might exhibit similar "neural tells." When a model starts to hallucinate, there should be a detectable anomaly in the Hidden State Dynamics before the token is even sampled.

This led me to develop the Sibainu Engine.

My goal was to build a pre-emptive auditing layer that runs on consumer-grade hardware (RTX 3050 4GB). By monitoring the geometric stability (which I call "Layer Dissonance") between transformer layers in real-time, the engine identifies the "collapse of latent trajectory" with a core latency of less than 1ms.

Key Technical Highlights:

Efficiency: ROC-AUC > 0.90 across Gemma, Llama-3.2, and Mistral, without any additional training or fine-tuning.

Low Overhead: While the Python API adds some serialization delay, the vectorized NumPy core is fast enough to be integrated directly into any inference pipeline without bottlenecking generation.

Autonomous Recovery: I've included a demo where the engine aborts a "corrupted" session and triggers a deterministic re-generation the moment a physical neural anomaly is detected.

I believe that for LLM safety to be truly scalable, it needs to be lightweight and deterministic. I’m curious to hear your thoughts on this geometric approach and its potential generalizability to larger architectures.

I also suffer from severe tension headaches and go to a chiropractic clinic. I don't fully understand the theory, but they apparently move my spine back into the correct position by gently rocking it. I've been going for about two months now, and I'm gradually getting better. At the clinic, I'm given several types of exercises, mainly focusing on the muscles around the shoulder blades, and I perform these exercises each time until my next appointment.

Having AI review code is practically standard practice now, isn't it? In that sense, I think you could say my boss is essentially reviewing me using AI.

Ultimately, I believe the most important thing is how we effectively utilize AI. We can't and shouldn't entrust everything to AI in any field, not even after AGI is perfected. Sometimes it's important to mass-produce low-quality code, and other times it's important to create beautifully crafted code.

1M is truly amazing. However, what is the incidence of hallucination? I haven't found a benchmark, but I feel that maintaining context at 1M would likely increase hallucination. Is there some kind of mechanism to suppress hallucination?

So, the value of code writers is decreasing these days, to the point where they can be replaced by AI. Perhaps the future will be one where one architect uses numerous AIs to produce results.

In the near future, a "good programmer" might not be defined by someone who can write bug-free, clear code, but rather by someone who can prompt for code that works consistently within the context of AI. If that happens, I'll have to find a different job.

This is amazing! I have one question: how did you determine the importance of each layer? I'm currently trying to use a geometric approach to predict the behavior inside the LLM black box, but I'm having trouble analyzing the connections between multiple layers.

That's interesting news. I wonder how much kinetic energy it had. This accumulation of information might be useful if an asteroid were to hit the Earth someday. At the very least, it's more realistic than sending oil drilling experts to an asteroid.

Seeing this kind of news makes me very uneasy. There's no doubt that AI is more efficient, but what about diversity? If many companies prioritize AI over humans, won't society as a whole become very uniform?

When I see news like this, I feel very uneasy. It's not about a decrease in employment, but about the decrease in diversity in society. Compared to human diversity, I feel that the diversity of AI is very limited. I worry that everything in society will become uniform.

It's certainly important to prevent falls. Especially as adults, we tend to lack the energy to get up. In that respect, children are amazing. I recently started studying for an LLM as a hobby, but I keep falling over and spending less time getting up. I often think it would be easier to just give up and go to sleep.

Realistically, it seems impossible for a third party to verify whether or not the relevant data is included. Content derived from original content, such as movies, is likely to have learned a large amount of data. If it were to be prohibited to use such content, it would be an overly broad interpretation of copyright.

Rather than worrying about accuracy, please do something about the pain that will make you cry if you step on it!

That's amazing. I'm developing sub-tools for LLM as a hobby on an RTX3050 (4GB), but I can only run lightweight models like 1B and 2B. Is it possible to use your tool to make the CPU take over some of the VRAM movement?

I always thought it was amazing how Lego pieces fit together so perfectly that they wouldn't come off even if you lifted them, but if you wanted to remove them, they came off so easily, and I had no idea they were that precise.

This is an incredibly useful tool for managing transaction history.

It is an intriguing and aesthetically profound approach.

Thanks for the precise critique. You are right: Recall 54% is the "danger zone." In a regulated or production environment, missing half of the structural collapses is functionally equivalent to zero protection. The 88% precision proves the signal exists, but the threshold for "collapse" in latent space is currently too rigid. The "Geometric approach (SIB) + Rule-based output validation" hybrid you suggested is the most logical path forward. • Geometric Drift (Layer-Internal): Catches the "process" of losing logical coherence (structural entropy). • Rule-based (Output-Layer): Catches the "result" of semantic violations (pre-defined constraints). My next focus is analyzing the "Silent Failures" — the 46% we missed. If the latent space doesn't show geometric collapse but the output is still a hallucination, it suggests the model is confidently drifting into a "parallel" but structurally stable manifold. That's a different failure mode that geometry alone can't catch. Reproducibility is the only way to move this out of "voodoo AI" territory. Glad the raw_logs.csv helped.

I’ve been exploring why LLMs "break" during inference. Most current hallucination detection methods look at the final text (semantic analysis) or use another LLM to double-check (self-consistency). These are effective but extremely slow and expensive.

SIB-ENGINE is my attempt to solve this at the geometric layer. By monitoring the "Anchor Drift" (how hidden states deviate from the prompt’s latent trajectory), I found that hallucinations often manifest as a structural instability before the token is even sampled.

The Numbers:

Recall: 53.89% (It catches about half, but it's consistent)

Precision: 88.52% (Low false-alarm rate is my priority)

Overhead: <1% (Running on an RTX 3050 with 4GB VRAM)

AUC: 0.8995

I've released a Lite version (1-axis) on GitHub so you can see the fundamental logic and run it on your own machine. I’ve also included the raw_logs.csv from my N=1000 test run on Gemma-2B for full transparency.

I’m particularly curious if anyone here has experimented with similar geometric approaches or has thoughts on how this might scale to 70B+ models where the latent space is significantly denser.

Happy to dive into the technical details!

I’ve been exploring why LLMs "break" during inference. Most current hallucination detection methods look at the final text (semantic analysis) or use another LLM to double-check (self-consistency). These are effective but extremely slow and expensive.

SIB-ENGINE is my attempt to solve this at the geometric layer. By monitoring the "Anchor Drift" (how hidden states deviate from the prompt’s latent trajectory), I found that hallucinations often manifest as a structural instability before the token is even sampled.

The Numbers:

Recall: 53.89% (It catches about half, but it's consistent)

Precision: 88.52% (Low false-alarm rate is my priority)

Overhead: <1% (Running on an RTX 3050 with 4GB VRAM)

AUC: 0.8995

I've released a Lite version (1-axis) on GitHub so you can see the fundamental logic and run it on your own machine. I’ve also included the raw_logs.csv from my N=1000 test run on Gemma-2B for full transparency.

I’m particularly curious if anyone here has experimented with similar geometric approaches or has thoughts on how this might scale to 70B+ models where the latent space is significantly denser.

Happy to dive into the technical details!

I’ve been exploring why LLMs "break" during inference. Most current hallucination detection methods look at the final text (semantic analysis) or use another LLM to double-check (self-consistency). These are effective but extremely slow and expensive.

SIB-ENGINE is my attempt to solve this at the geometric layer. By monitoring the "Anchor Drift" (how hidden states deviate from the prompt’s latent trajectory), I found that hallucinations often manifest as a structural instability before the token is even sampled.

The Numbers:

Recall: 53.89% (It catches about half, but it's consistent)

Precision: 88.52% (Low false-alarm rate is my priority)

Overhead: <1% (Running on an RTX 3050 with 4GB VRAM)

AUC: 0.8995

I've released a Lite version (1-axis) on GitHub so you can see the fundamental logic and run it on your own machine. I’ve also included the raw_logs.csv from my N=1000 test run on Gemma-2B for full transparency.

I’m particularly curious if anyone here has experimented with similar geometric approaches or has thoughts on how this might scale to 70B+ models where the latent space is significantly denser.

Happy to dive into the technical details!

One thing I didn’t emphasize in the post: this work started partly from thinking about how black-box generative models might be audited under emerging regulations like the EU AI Act, where access to model internals or weights can’t be assumed.

Instead of aiming for human-readable explainability, ONTOS looks at whether it’s possible to leave behind reproducible, quantitative traces of structural stability during generation — something closer to audit evidence than a narrative justification.

I don’t claim this says anything about factual correctness or ethics. The narrower question is: was this generation process structurally stable, predictable, or already collapsing internally, even if the output still looks fluent on the surface.

I’m curious whether people see structural monitoring like this as complementary to existing safety / compliance approaches, or fundamentally limited in ways I might be missing.