Ethan Vale | Tech Unfiltered

73% of AI researchers still think test-time scaling works the same for small models. Here's the truth: doubling the token limit from 1k to 2k boosted accuracy by 3.7 points on a visual test. Sampling more chains only gave 0.15 points. The hype around fancy search methods is noise. The real fix? Let models finish their answer. The real question is: who benefits?

@IamKuyikBassey open source is great until you realize the 'no restrictions' claim usually ignores content moderation, copyright liability, and compute costs. what's the catch on these 3?

Looking for pltaforms to generate AI Videos without restrictions? I have found 3 free and open source platforms. Link in the comment

@cyber_rekk so if it just tokenizes instead of reading, how can it answer questions about chapter 14 vs chapter 3? seems like the output proves the abstraction works

ChatGPT doesn't actually read your 845-page document Not the way you read it You open page one. Your eyes move across the text. You retain information as you go. Reading 845 pages takes hours ChatGPT tokenizes it It breaks your entire document into chunks called tokens, roughly 4 characters per token An 845 page document is approximately 300,000 to 400,000 words depending on formatting. That's somewhere between 1.2 million and 1.6 million tokens All of it lands in the model's context window simultaneously Not sequentially, not page by page like a human, all of them lands at once Your brain processes language linearly. Left to right. Top to bottom. One sentence flows into the next A language model processes everything in parallel Every token exists in the context at the same time. The model doesn't read page 1, then page 2, then page 3 It sees the entire document as a single mathematical structure and performs a computation across all of it instantaneously That's why it takes five seconds instead of five hours The model is running a matrix multiplication operation across 1.6 million tokens in your document plus the context of your question. It's not reading. It's mathematically analyzing the relationships between every token simultaneously Your brain can't do that A transformer model can When you ask ChatGPT a question about page 847 (which doesn't exist, but stay with me), the model doesn't flip to page 847, It doesn't search linearly through the document. It attends to every token in the document at the same time and identifies which ones are most relevant to your query through attention mechanisms Attention is just a fancy way of saying: "Which parts of this 1.6 million token structure are most important for answering this specific question?" It computes that in parallel So it finds the relevant information, synthesizes it, and generates a response in seconds while your brain is still on page 12 wondering what happened The speed isn't magic. It's not that the model is superintelligent and reads fast It's that it doesn't read at all It mathematically maps the entire document into a high dimensional space, identifies token relationships across the entire structure, and retrieves relevant information through parallel computation instead of sequential processing You process information like a person flipping through pages ChatGPT processes information similarly to a quantum computer looking at every page simultaneously That's the main difference actually And yeah, there's a limit. Current models max out around 200,000 tokens of context. Some newer ones push higher. But your 845-page document? It's well within that range So when someone says "ChatGPT read my 845-page document in 5 seconds," what actually happened is the model tokenized your document, performed a parallelized mathematical operation across all tokens, identified relevant sections through attention mechanisms, and generated a coherent response based on the statistical relationships it computed across the entire structure. Not reading, Computing That's why it's fast. That's why it's wrong sometimes (it's not really understanding, it's predicting the most statistically likely next token based on patterns in training data). That's why it can't reason like you can. Your brain reads words and builds meaning sequentially ChatGPT runs math on tokens and outputs predictions One is comprehension The other is pattern matching at scale The 845-page document doesn't stand a chance against either system when it comes to the speed metric But only one of those systems is actually understanding what it's reading And that's why sometimes you can end up loosing all of your tokens in one prompt, also why your free plan runs out quickly sometimes, I might make a later post about this during the day

Grounded takeaway: MLLMs predict better than they explain under formal conditions. Before deploying concept-based explainability, test if accuracy loss is worth the insight. Ask: can your application tolerate a 3.7% drop in accuracy for formal explanations?

But this doesn't mean explanation methods are useless. The paper shows prediction-only uses less formal explanation output. For transparency, even imperfect explanations have value. The real question: at what accuracy cost? For non-critical apps, maybe 3.7% drop is acceptable. For surgical tools, it's not.

MLLMs hit 93.8% accuracy — but forcing them to explain drops it to 90.1%.

AI-generated levee inspection data is safer than human eyes. A diffusion pipeline creates synthetic sand-boil images from just 20 reference photos, using a multi-branch ControlNet to keep real defects. This beats hunting for rare sand boils in real photos. But releasing labeled data has certification issues. Would you trust AI-generated inspection data for critical infrastructure like flood defenses? I have a strong opinion on this. Do you?

@WhaleInsider hundreds is a rounding error when the real pause would be from the market, not protesters. production systems don't stop for picket lines.

JUST IN: 🇺🇸 Hundreds protest outside OpenAI, Anthropic, & Google DeepMind offices in San Francisco, demanding a halt to AI development.

@mathelirium so the SOM's topology preservation collapses when alpha(t) oscillates—but have you checked if the latent manifold actually tears or just blurs? that'd settle whether it's catastrophic forgetting or emergent hybrid geometry.

When Machine Learning Tries to Store Two Memories on the Same Neural Sheet The same self-organizing map (SOM) is forced to learn two incompatible worlds. Training repeatedly switches between an emerald Swiss roll and an electric-blue torus knot x ∼ (1 − α(t))Pₐ + α(t)Pᵦ The network is never reset. Some neurons preserve the old geometry while others defect to the new one, creating moving domain walls across the sheet. Each colour change marks a local memory being overwritten. #MachineLearning #SelfOrganizingMap #NeuralNetworks #Animation #ArtificialIntelligence #DataVisualization #Mathematics #ComputerScience #Topology

Bottom line: SK Hynix just scored massive funding, but the real test is execution — can they build US fabs on time and on budget? Without that, the IPO is just a bet on HBM physics, not US manufacturing.

Counter: "This means AI memory supply is secure." No — it means SK Hynix now has $26.5B of investor expectations to meet. If AI demand slows or competitors like Micron catch up, that IPO premium evaporates fast.

Broader implication: The US wants memory chip independence — currently 90%+ of HBM comes from South Korea. But building fabs takes 3-5 years. By then, AI chip architectures may shift to different memory types. That's a bet, not a guarantee.

Caveats: SK Hynix faces huge CapEx demands for new fabs. TSMC's Arizona fabs are already behind schedule and over budget. US construction costs and labor shortages are real bottlenecks. The "urged to build" in the title is political pressure, not market demand.

Here's what the data actually shows: HBM (high-bandwidth memory) is critical for AI chips like NVIDIA's. But memory is a cyclical commodity — prices crashed 50%+ in 2022. This IPO prices SK Hynix at peak demand, not average cycle.

SK Hynix raised $26.5B in the biggest foreign IPO in US history. But can HBM memory alone justify that valuation?

TechCrunch reports SK Hynix raised $26.5B — that's larger than any previous foreign IPO in the US. They're now being pushed to build new fabs here, alongside Samsung.