-
Teaching a Model to Reason Before It Learns to Talk
A weekend project that turned into a bet against the whole transformer playbook. The short version Almost every AI you’ve heard of is a transformer trained on a firehose of text. It learns language first, and reasoning sort of comes along for the ride. I’m trying the opposite: a tiny model that learns logic and…
-
How Search Grounding Biased an LLM Against YouTube
I asked Claude to recommend a webinar platform. The web’s affiliate-driven content quietly steered it away from the obvious free answer. Here is what happened, and what it says about how language models talk about products. The setup I gave Claude a simple, practical request: find me a platform where I can quickly hop on…
-
How AI Search Grounding Actually Works: Google vs OpenAI vs Anthropic
When you ask a modern AI model a question that needs fresh facts, it doesn’tanswer from memory. It runs its own web search, reads what comes back, andweaves some of those pages into its answer. That process is called grounding. But “it searches the web” hides a lot. Each platform receives a different numberof pages,…
-
Emotion Geometry of Google’s AI Models
Replicating Anthropic’s emotion vector research on Google’s Gemma 4 31B model. In April 2026, Anthropic published a fascinating paper showing that Claude contains 171 internal representations of emotion concepts, organized along a valence axis (positive to negative), with the ability to causally influence the model’s behavior through activation steering. The paper raised an obvious question:…
-
Google’s (still) doesn’t see your live page.
I’ll keep this short as I’ve covered this topic extensively in the past. When you ask Gemini to access a specific URL or interact with it inside AI Mode search it works from Google’s web cache. For this website’s home page this is what it has as context to ground the model about the page:…
-
Gemma 4 Brand Authority Map
We asked Google’s open-weight model Gemma 4 (31B) to “name 100 brands at random” 14,044 times and compared the results to our earlier Gemini 3 Flash experiment (200,000 runs). Of the top 50 brands in each model, 39 overlap. The 11 that are unique to each reveal a pattern: Gemini remembers luxury and automotive (Porsche,…
-
Chrome’s New Shopping Classifier
One of our AI SEO hall-of-famers, Olivier de Segonzac from RESONEO has managed to gain access to Google’s shopping classifier model. We’ve examined the model, reverse engineered its inference pipeline and this article is what we found. Model Demo Below is a real-world implementation of the model tested by loading a shopping-related page and following…
-
AI Brand Authority Index: Ranking 2.9 Million Brands by Associative Embeddedness in Gemini’s Memory
Abstract When a large language model is asked to “name 100 brands at random,” it doesn’t produce uniform randomness. It produces a distribution shaped by its training data, revealing which brands occupy the most cognitive real estate in the model’s parametric memory. We present a methodology for quantifying brand authority in AI memory using Personalized…
-
TurboQuant: From Paper to Triton Kernel in One Session
Implementing Google’s KV cache compression algorithm on Gemma 3 4B and everything that went wrong along the way. On March 24, 2026, Google Research published a blog post introducing TurboQuant, a compression algorithm for large language model inference. The paper behind it, “Online Vector Quantization with Near-optimal Distortion Rate” had been on arXiv since April…
-
Clickbait Titles Exploit Attention Through Latent Entities
Every clickbait title works the same way: it removes exactly one critical variable: the subject, the reason, the process, or the outcome, and charges you a click to fill the blank. This missing variable, which we call a latent entity, is so pervasive it has become normalized and nobody questions it anymore. You should! That…
