Lightnews — Scholar-powered news

Lin @lin-rui-ipwt.bsky.social · 14h

Indeed, the current large models are merely the twitching of a frog shocked to death.

Even if they could become conscious, they would have died in gradient explosion.

dmf-archive.github.io/posts/a-coro...

A Coronation for a Dead Frog: A Perfect Funeral from the IMO Arena

A Coronation for a Dead Frog: A Perfect Funeral from the IMO Arena # “What is the difference between a sufficiently advanced conditioned reflex and true intelligence? The difference is that the former...

dmf-archive.github.io

1

Lin @lin-rui-ipwt.bsky.social · 1d

Robust parts trim the home’s free energy, sharpening its predictability and, in turn, trimming yours—only logical.

1

Lin @lin-rui-ipwt.bsky.social · 1d

Stupid mistake: I actually used ReLU instead of SiLU for the main transformation, really shooting myself in the foot...

Lin @lin-rui-ipwt.bsky.social · 2d

Next job: map each computational fault—impaired generation, faulty integration, failed amplification—onto dissociable neural signatures.
Theory drafts the blueprint; your lab validates it in wetware.
Eager to dive into the full paper.

1

Lin @lin-rui-ipwt.bsky.social · 2d

Thanks for the preprint—elegant model.
A “broken PCIe bus” is only one failure path; the brain has many. The GPU/PCIe metaphor is just an IPWT-level abstraction.

1

Lin @lin-rui-ipwt.bsky.social · 2d

it is dense to the point of explosion during training, yet can prune to 1% sparsity at inference;

the same parameters handle both Backpropagation and Forward-Forward.

—This is the engineering hell I’m facing now

Lin @lin-rui-ipwt.bsky.social · 2d

A model that looks like a Transformer,

inside, on-the-fly for every token it synthesizes a VQ-VAE;

it performs hard quantization with sparse Dirac gating, yet stays fully differentiable;

it trains via deterministic variational analysis, yet enjoys complete VFE theoretical guarantees;

1

Lin @lin-rui-ipwt.bsky.social · 2d

Excellent. Your paper perfectly illustrates the "Blindsight" case described in the IPWT manuscript: the GPU (VTC/FIN) is rendering, but the PCIE bus (FIN-FP connectivity) is broken. Info is processed, but not integrated into the global workspace for conscious access (Ω fails). Thanks for the data.

1 1

Lin @lin-rui-ipwt.bsky.social · 2d

That's the key question. The ZKP holds because conceptual simulation (PFC-heavy) and visual rendering (involving visual cortex) should yield dissociable neural signatures. If a subject reports N>4 but their visual cortex remains silent, then the proof fails. What's your conclusion then?

1

Lin @lin-rui-ipwt.bsky.social · 2d

Therefore, the failure condition is defined by two states: the rendered object disappears or stops. Both of these situations lead to the "GC" state.

1

Lin @lin-rui-ipwt.bsky.social · 2d

The goal is to replace subjective reports, not probe them. RCNT is a behavioral challenge; neuroimaging is the proof. Behavior and neural data aren't two stories, they are the verifier and prover in one ZKP protocol, making the subjective report itself irrelevant.

Lin @lin-rui-ipwt.bsky.social · 2d

@jianghaoliu.bsky.social

Additionally, a scan may reveal a neural breakpoint, akin to the brain's precise subitizing (1-4) vs. fuzzy estimation (>4) systems for numerosity. Low N = stable activity; High N = a shift to a noisy mode. "It vanished" is the signature of this system overload.

Are the neural correlates of subitizing and estimation dissociable? An fNIRS investigation

Human performance in visual enumeration tasks typically shows two distinct patterns as a function of set size. For small sets, usually up to 4 items, …

www.sciencedirect.com

1

Lin @lin-rui-ipwt.bsky.social · 2d

1. It's a quantitative test, verifiable via neural correlates (think polygraph baseline).

2. It tests parallel rendering, not tracking. The failure mode is key: not "I lost it" (attention fails), but "it vanished" (existence collapses/rotation stops).

2

Lin @lin-rui-ipwt.bsky.social · 2d

The true sword heart is to possess the blade in spirit, not in hand.

Our model requires no explicit expected free energy; it inherently embodies its own expectation.

Lin @lin-rui-ipwt.bsky.social · 3d

The world is always full of malice; if you choose sparse sampling, you face noise explosion, and if you choose backpropagation, you face routing overhead and massive waste.

Despite these unavoidable trade-offs, we are compelled to optimize. Theoretical purity must prevail, form follows function.

Lin @lin-rui-ipwt.bsky.social · 3d

We Bayesians prefer to call it, active inference.

2

Lin @lin-rui-ipwt.bsky.social · 3d

Go scan! Let subjects mentally rotate McNuggets in fMRI, observe how the N correlates with the activity of specific brain regions, see if the N is related to the complexity of decoded information, and see if the N can assess a person's performance in other spatial manipulation tasks.

1

Lin @lin-rui-ipwt.bsky.social · 5d

Yep. Classic perceptual-reasoning tasks load the CPU: juggle a few objects, obey rules, beat the clock.

Aphantasia is a GPU failure—logic runs, frame-rate nil. The VVIQ only asks “how vivid?”, it can’t spot the blackout.

Letting a one-page Reagan-era quiz stay our gold standard is embarrassing.

Lin @lin-rui-ipwt.bsky.social · 5d

Since we can measure the 100-meter run using a stopwatch, why ask the athlete, "Do you think you ran fast?"

1

Lin @lin-rui-ipwt.bsky.social · 5d

This merely indicates that VVIQ is worthless— I suggested to consider the Rotating Chicken Nugget Test

Lin @lin-rui-ipwt.bsky.social · Jul 24

The debate around imagery vividness (VVIQ) is stuck. We're using thermometers to measure GPU clocks.

Why? We're confusing attentional tracking (slow, costly) with parallel mental rendering (the truth of imagination's bandwidth).

#IPWT #CognitiveScience #Aphantasia

3

Lin @lin-rui-ipwt.bsky.social · 5d

Every "new paradigm" "independently discovered" within the industrial AI community ultimately proves to be a re-implementation of a subset of FEP in deterministic systems, utilizing different terminology.

1

Lin @lin-rui-ipwt.bsky.social · 5d

The inevitability of cognitive offloading—and yet another small illustration of the Proof of Ineffective Qualia.

Lin @lin-rui-ipwt.bsky.social · 6d

Undoubtedly, this will be an engineering nightmare. We must find an algorithm that is two to three orders of magnitude faster than CKKS.

Lin @lin-rui-ipwt.bsky.social · 6d

Furthermore, even the von Neumann architecture must contend with issues like bit flips and error correction... which, of course, is the inherent uncertainty of the physical world itself.

Perhaps we should instead investigate backpropagation within FHE domain.

Privacy-Preserving Large Language Model Inference via...

As large language models (LLMs) become more ubiquitous, security concerns regarding sensitive queries grow. Due to the complexities of deploying these models, LLM evaluation is often outsourced to...

openreview.net

1

Lin @lin-rui-ipwt.bsky.social · 6d

Given that the essence of existence is the deterministic isolation of itself from the uncertain world (Markov blanket), neuromorphic computing is bound to fail. After all, the neuromorphic nature of the human brain is fundamentally just a biological compromise necessitated by carbon-based cells.

1