13

Saying "Nobody knows what they are doing" is just a convenient excuse to justify terrible engineering. There is a fine line between scientific trial-and-error and proud, brute-force ignorance.Let's be clear about your "frontier":Blind Gambling: When independent labs don't understand the underlying mathematics or hardware physics, they just throw data at a wall and pray to the loss curve. That is digital alchemy, not science.The Loop: Instead of fixing structural bottlenecks or learning non-linear dynamics, people just brute-force configs. It’s the engineering equivalent of a cat grooming itself because it has nothing else to do.Zero Legacy: This unscientific approach is why the ecosystem is flooded with overfitted, hollow checkpoints that break down outside of their strict test sets.You aren't advancing the frontier; you are just polluting the platform because you refuse to open a textbook. Brute force has hit a physical wall. True innovation requires cognitive architecture and actual engineering, not just romanticizing failure. 🫵🤡

While this article provides some neat visuals, it unfortunately wraps a highly flawed and inefficient architecture in a shiny paper of a "breakthrough". Let’s look past the textbook theory and address why this engram implementation is a structural dead-end in real-world deployments:The Out-of-Distribution Brain-Freeze:The author notes that because of Zipf’s law and mathematical impossibility of mapping (Vocab^{N}) permutations, these tables rely heavily on confidence gating. But here is the architectural catch: when a model is trained to rely on an engram table for its semantic foundations, its core attention and perceptron layers are essentially "spoiled" by the crutch.
The moment the model encounters highly abstract, non-linear, or completely out-of-distribution reasoning (deeply nested code, intricate logical paradoxes, custom corporate APIs), the confidence gate shuts down. The model is suddenly forced to think for itself, but its underlying token representations have been hollowed out. It doesn't just "get confused"—it undergoes a catastrophic collapse in reasoning capability because it was optimized to read answers from a static cheat-sheet rather than synthetically computing relationships on the fly.The Memory Tax and the PCIe Bottleneck:The marketing pitches for these models claim they are "cheap because they only activate a few billion parameters."
This is a massive illusion. Storing tens of millions of bi-gram and tri-gram hashes requires a massive memory footprint (often 30GB to 50GB of static weight tables). In a local consumer deployment, you cannot fit this dead weight into VRAM alongside a decent context window. The only solution is offloading these tables to system RAM or NVMe SSDs.
Even with asynchronous pre-fetching, you are forcing the execution pipeline to perform constant, random, non-contiguous memory lookups over a slow PCIe bus. The computational efficiency gained by skipping attention layers is completely eaten alive by hardware latency and cache misses. You aren't saving resources; you are just moving the bottleneck from the GPU cores to the system's memory channels.The Utility Cap: Optimized for French Fries, Blind to Chaos:Ultimately, engrams are only efficient if your target deployment is a low-cost, cloud-hosted customer service bot designed to spit out highly predictable text, standard corporate policies, or recipes for french fries. In those narrow boundaries, matching pre-hashed chunks works wonders.
But for heavy, creative, open-ended inference, or multi-layered logic where meaning is non-local and distributed across thousands of tokens, this approach is a dead-end. You cannot build genuine cognitive capability by expanding the dictionary.
A model that relies on a 50GB index of pre-baked phrases will always be structurally inferior to a clean, dense transformer that computes full-context attention matrices natively.Thanks to the author for writing this "fanfiction" without actually digging into how this mess breaks under hardware constraints and shifts in data distribution, all while proudly displaying a PRO badge on the profile. We must be careful not to mistake a highly bloated lookup table for a genuine evolutionary step in architecture.

It is genuinely amusing to watch you treat these raw logs as if they were holy scripture. You are building an entire architectural theory around the performance of Gemini 3.7 Flash, based on a dataset whose actual generation conditions and hardware variables you cannot verify.
Your numbers are not objective facts; they are just sophisticated fanfiction based on a static spreadsheet.Let us bring you back to the reality of actual hardware and computer science:The Infrastructure Black Box:You are running rank-correlations on 309,955 rows of data, but do you know the exact temperature, top_p, system prompts, or generation seeds used for these models? More importantly, what was the underlying infrastructure? LLMs are highly sensitive probability matrices, not rigid linear functions. You cannot claim an absolute benchmark rank when you don't even know the sampling parameters used to generate the outputs.The Hardware Factor & Silicon Physics:You show a complete lack of understanding regarding how inference actually works on a physical chip.
It has been proven time and again that the exact same model weights will produce completely different raw outputs when run on different hardware configurations. Large models suffer from:Tensor leaks and soft errors occurring deep within dense GPU clusters.
Bit-flipping and silent data corruption (SDC) that subtly alters token probabilities.Dark Silicon and thermal throttling effects that introduce non-deterministic execution paths during high-throughput inference batches.Measuring the Mirage:When you compile an ELO rating from crowd-sourced mobile clicks on unverified hardware runs, you aren't measuring a neural network’s capability. You are measuring a noisy, non-reproducible snapshot of a highly volatile system.
Your near-perfect correlations prove nothing but the internal consistency of a pre-filtered CSV file.Before you run around waving graphs and calculating Spearman coefficients to the third decimal place, you need to learn the absolute basics of silicon physics and inference environment control. You are analyzing a theatrical script, not the mind of the model. 🤡

oh-what we have here-Just stumbled upon this thread by pure chance and honestly, this whole circus is hilarious. What exactly are we looking at here? Instead of actual arguments against valid technical points, all we see is a classic emotional meltdown.
Professional discussions often involve strong disagreements, but maintaining a respectful environment is essential for productive collaboration. When technical points are met with personal conflict rather than counter-arguments, it can hinder the progress of a project and the community as a whole. Community guidelines on most platforms are designed to prevent targeted harassment and ensure that interactions remain focused on the work itself.
Regarding the technical aspects mentioned, there are valid concerns regarding the impact of certain tools on the ecosystem, particularly those that might be used for metric manipulation or non-genuine activity. Addressing these architectural and ethical challenges is a critical part of software development. Focusing the conversation on code quality, security, and the intended use of a project allows for more meaningful improvements and a healthier environment for all contributors.bravo-keep going

Oh, look at this utter embarrassment. It is fascinating to watch the defense mechanisms of a platform when its methodology is scrutinized. When faced with a detailed audit showing a 98% correlation between redundant axes and structural inconsistencies, the official brand account resorts to spamming a prompt-injection script asking for a chocolate cake recipe.Let us clarify a few fundamental points regarding this evaluation theater:The Inability to Counter-Argue:Attempting to trigger an AI bot with a baking recipe instead of addressing data redundancy with mathematical proof is a complete white flag.
If the team lacks the technical arguments to defend their own dataset card, it is better to remain silent rather than showcase this level of corporate evasion. If you truly require a recipe, there are far better places to look than a technical discussion.The Illusion of "Independent" Metrics:The concept of independent testing in the current LLM landscape is an initial profanation. Anyone who investigates how these benchmarks are established, who funds these studies, and how the metrics are aggregated will quickly realize that the entire ecosystem is designed for marketing optimization rather than discovering actual capabilities.
You are benchmarking memorized pattern reproduction, not reasoning.Understanding Evaluation Foundations:Publishing millions of human judgments means very little if the underlying axes collapse into a single preference signal. This entire approach is designed to simulate scientific progress for investors while remaining fundamentally shallow.
It is a three-faced mirror where documentation gaps and convenient update rules provide plausible deniability when the numbers are questioned.If the creators cannot provide a substantive response regarding the mathematical limits of their data, it would be wiser to stop shifting the goalposts. True researchers answer with architecture and validation, not with a shopping list for confectionery. 🤡

The patronizing tone of this writeup is wild, considering the grand conclusion is literally just discovering the law of large numbers.
Acting like a single greedy pass not representing the full distribution is some mind-bending epiphany is pure comedy. That’s baseline evaluation logic, not a breakthrough.The actual engineering here is broken on two fronts:Slapping a LoRA on a bloated 36B legacy carcass is useless. A model that heavy—especially an older, over-parameterized Chinese base—has a massive hidden probability space.
A low-rank adapter is just a duct-tape fix; it skims the surface without changing the core associative paths.
The second you dial the temperature to 0.7, the sampler easily navigates the native weight distribution and walks right past your thin LoRA patch. If you want a hard-stop "vulnerability-gate," you do a full fine-tune (FFT) to hardwire the refusal boundaries. Otherwise, you’re just wasting compute on a placebo.You test proof-of-concept gates on tight 2–4B models, not unoptimized behemoths.
If you actually want to map variance and stress-test a security architecture, you run it on a clean 2B or 4B base. It’s cheap, predictable, and lets you brute-force thousands of tokens to get real statistical distribution data. Doing this on a messy 36B model is just burning GPU cycles for noisy metrics.Besides, this whole setup is an absolute joke against real adversarial attacks.
Drop a proper token optimization suffix (like GCG) or an activation shift in there, and the model folds instantly. Vector-level attacks hijack the deep attention layers, completely blinding your surface-level adapter patches."Receipts, not hype" doesn't mean much when your actual receipt is just realizing that a janky kludge behaves exactly like a janky kludge.

@ProCreations You really love to talk when you're completely out of your depth, don't you?It is incredibly pathetic to watch you stalk user profiles across different repositories, obsessively uploading screenshots because your fragile ego cannot handle a legitimate architectural critique.
Real deep learning engineers focus on tensor dynamics and gradient flows; you focus on running a creepy, amateur detective agency because someone exposed your sloppy competence.Let me give you a reality check, buddy: aren’t you even slightly worried about the consequences of this blatant, targeted harassment and tracking of user profiles?
You are shamelessly crossing platform boundaries just to hide your engineering flaws under defensive memes.Instead of hosting this charity theater to fund people who will burn $150 making the exact same mistakes, you should spend that money on a basic deep learning textbook.
Keep collecting your little screenshots, but don't be surprised when this obsessive stalking and targeted toxicity backfires on your own account. I will no longer waste my time on this nonsense-get some real live kido)

@ProCreations I see you are still running around different repositories, obsessively tracking my account and uploading screenshots like a malfunctioning database script.
Are you seriously that emotionally damaged by a technical critique of your little $150 charity theater? Imagine being so desperate that you resort to outright stalking across platforms and threads just to avoid admitting your engineering flaws.
Let me ask you something, buddy: aren’t you a bit worried that this blatant, targeted harassment and tracking of user profiles might backfire? You are shamelessly violating community boundaries because your fragile ego cannot handle being corrected.
Real world networks have ways of filtering out pests who cross the line from coding into targeted stalking.Instead of acting like a creepy hall monitor collecting images of people who expose your sloppy competence, do yourself a favor and look at your own code. Your obsession is becoming pathetic, and if you keep pushing this targeted harassment, don't be surprised when actual consequences follow.

A simple question for you, since you tried to play the 'AI slop' card to dismiss the critique. If you actually 'realized' it was a model or 'slop' from the start, why didn't you just state it upfront with a smug look to show everyone what a great 'AI specialist' you are?Instead, you chose to panic, write a defensive technical reply trying to argue about vocabulary size × hidden size, and then called a gang of scammers to mass-report the comment and break the platform's rules just to hide the post.If you truly 'busted' a bot, you would have laughed it off publicly. But you didn't. You treated it as a lethal threat to your reputation, engaged in a serious debate, lost it mathematically, and then hid behind a cowardly wall of mass-flagging. Your actions speak louder than your excuses. You didn't see 'slop' — you saw a mirror, and you couldn't handle the reflection.Because of people like you, the Hugging Face community has devolved into a kindergarten of offended children, where fragile emotions rule the day instead of actual arguments. It's pathetic.wp guys-bravo!keep going.

It is truly pathetic to watch a gang of fragile egos coordinate mass reports just to hide objective technical criticism. You characters create a whole "group" to violate Hugging Face terms of service regarding brigading, completely proving that you cannot handle a real debate. First, you whine about "personal insults," and then you pull a cowardly move like this because you lack the brainpower to counter her arguments with actual math.Let us peel back the layers of your "amazing" scam here. You talk about compute, but you don't even understand the baseline mechanics of the architectures you are playing with. Adrienne completely stripped your 400M model naked, but let me open your eyes even further.If you throw away the tokenizer and the basic syntax layers from a small model, you are already hollow. But here is a little secret for the butchers: attention heads are heavily marketed parameters, not dedicated, independent layers of core knowledge. In these micro-budgets, the actual capacity left for processing deep logic and reasoning is barely 10% to 20%. You are literally trying to force a heavily castrated dictionary to act as a system-call validator.Instead of hiding behind the "Report" button and organizing mass-flagging parties like children, you should have taken her advice, read your own configuration files, and learned how to fine-tune specific layers properly. This charity theater isn't research; it's a mutual coping mechanism for people who don't know how transformers actually process weights.Bravo)))

Ah, the classic 'it’s not a bug, it’s a feature' defense.Let's cut the academic gaslighting. If your primary scientific goal was to study 'how LLMs converge on scorer weaknesses' and test computational limits, you should have titled the project 'Evaluating LLM Adversarial Capabilities on In-Silico Benchmarks'. Instead, you titled it 'Open Discovery Challenge #1 — Malaria' and opened with 'Your AI can design a malaria drug candidate.' That is a textbook definition of a goalpost shift.And comparing your private database to established cloud infrastructure or big pharma pipelines is a joke. Real tech infrastructure is backed by legally binding SLAs, enterprise compliance, and multi-million dollar liability clauses. Asking researchers to drop unpatented IP into a zero-equity, decentralized project's private server based on nothing but internet goodwill isn't 'how the industry works'—it's just bad operational security.You can spin this as a 'meta-result' all you want, but you got caught running a broken benchmark under a humanitarian headline. Enjoy your $1,000 bug-bounty hunt. I'm done here. 🍌🤡