#tokenization
126 related articles
·5 min
Inside Cursor: How Recursive Model Improvement Enables AI to Self-Evolve
A Cursor ML engineer breaks down AI training methodology: outer/inner loop acceleration, preventing reward hacking, textual feedback, and recursive self-improvement (RSI) where models train the next generation.
Read more →
·6 min
Claude Sonnet 5 In-Depth Testing: 8 Real-World Tasks Reveal Its True Capability Boundaries
Bilibili creator KaterSony tests Claude Sonnet 5 across 8 real-world tasks—image recognition, 3D modeling, web generation—comparing it against GPT-5.5, Gemini 3.1 Pro, and revealing its true capability limits and cost traps.
Read more →
·4 min
Claude Usage Anxiety: The Hidden Cost Trap of AI Coding Tool Subscriptions
Real Reddit user rants reveal AI subscription pain points: Claude, Sol, and other tools consume usage at wildly different rates—does faster mean pricier? A deep dive into AI billing logic, usage transparency, and platform trust.
Read more →
·3 min
Qwen Next-Gen, DeepSeek V4, New GLM Spotted Simultaneously: China's LLM Race Accelerates
Qwen next-gen (KLab), DeepSeek V4 GA, and Zhipu GLM's new version are all in testing simultaneously. A deep dive into the latest developments and tech trends.
Read more →
·3 min
Claude Sonnet 5 Deep Dive: Impressive Benchmarks, Disappointing Reality — Stick with OPUS 4.8
Anthropic's Claude Sonnet 5 claims near-OPUS 4.8 performance at lower cost. Real-world tests reveal hidden tokenizer costs, weak creative output, and only 13th place on Cursor rankings.
Read more →
·2 min
Wildcards in Action: A Guide to Batch Exploration and Diversity in AI Image Generation
Wildcards are one of AI image generation's most underrated features. Learn how the randomization mechanism works, how to use ChatGPT to bulk-generate vocabulary lists, and how to build reproducible, systematic creative workflows.
Read more →
·4 min
Understanding Transformer From the Ground Up: How Language Models Predict the Next Word
From Tokenization and Embedding to the Attention mechanism, this article systematically breaks down how Transformer works — and how ChatGPT turns input text into next-token probabilities.
Read more →
Claude Usage Draining Too Fast? A Deep…
·2 min
Claude Usage Draining Too Fast? A Deep Dive into Token Billing and Usage Transparency
Users report Claude usage quotas draining faster than expected. This article explains token billing, context window accumulation, rate limits, and how to optimize your AI usage.
Read more →
·4 min
Getting Started with AI Test Development: Core Principles and Underlying Logic of Large Language Models
New to AI test development? This article breaks down the differences between machine learning and traditional programming, the origins of AI hallucinations, and the core principles of NLP/NLU/NLG to help test engineers build a solid AI knowledge framework.
Read more →
·5 min
AI Coding Getting Worse With Every Edit? Skills v1.1 Reveals the Real Cause and Solution
AI code getting messier with edits? The root cause isn't weak model capability but a lack of context and process. A deep dive into Matt Pocock's Skills v1.1: grilling, vertical-slice tickets, TDD, and WebFinder.
Read more →
Three Core Gaps in Multimodal LLMs: Fr…
·4 min
Three Core Gaps in Multimodal LLMs: Frontier Research from Microsoft Research India
Microsoft Research India reveals three core gaps in multimodal LLMs: visual perception blindspots, cognitive hallucination, and architectural limitations. Explores Faithful GRPO, behavior modeling, and model alignment breakthroughs.
Read more →
·3 min
Render Code as Images and Send to Claude — Cut API Costs by 70%
PXPipe renders system prompts and tool docs as images sent to Claude, exploiting visual token pricing. Real-world results: 68% fewer input tokens, 59–70% lower bills.
Read more →
·2 min
Long Conversations Are Draining Your AI Credits: A Deep Dive into Token Billing
Why do long AI conversations drain your credits so fast? This article explains LLM stateless architecture, Token billing mechanics, and practical tips like sliding windows and summary compression to cut API costs.
Read more →
Cross-Tokenizer Knowledge Distillation…
·2 min
Cross-Tokenizer Knowledge Distillation: Why 85% of Information Vanishes and How the Chain Rule Fixes It
Many-to-one token mapping in cross-tokenizer knowledge distillation silently drops 85% of teacher info, collapsing entropy from 2.09 to 0.32 bits. Learn the chain rule fix that restores retention to 83%+.
Read more →
·3 min
Programmer to AI Engineer: You're Already 80% There — Here Are the 6 Skills You're Missing
Programmers transitioning to AI engineering aren't starting from scratch. Learn the 6 core skills — LLM APIs, RAG, prompt engineering, LLMOps — needed to make the leap.
Read more →
Running Gemma LLM in Godot with GDScri…
·2 min
Running Gemma LLM in Godot with GDScript and Vulkan: An Experimental Deep Dive
A developer runs the Gemma LLM inside Godot 4 using only GDScript and Vulkan compute shaders — no llama.cpp or external dependencies. A technical breakdown of how it works.
Read more →
·6 min
OpenAI Merges ChatGPT and Codex: The Super App for the AI Agent Era Is Here
OpenAI merges ChatGPT and Codex into a super app and releases three new GPT-5.6 models: Sol, Terra, and Luna. A deep dive into four hands-on workflows—Computer Use, Loops, and multi-threading—for the AI agent era.
Read more →
·3 min
Zer0Fit: Connecting Google's TabFM and TimesFM to LLM Chat Interfaces via MCP
Zer0Fit wraps Google's TabFM and TimesFM foundation models as MCP servers, letting users run classification, regression, and time series forecasting through a local LLM chat interface — no ML code required.
Read more →
·3 min
Linguistics to NLP Engineer: Comparing Three Pathways and Finding the Optimal Strategy
How can linguistics or translation majors transition into NLP engineering? This article compares three pathways and offers a phased strategy covering core skills, project building, and job hunting tips.
Read more →
·3 min
Training a 100M-Parameter Model for Under $50: A Complete Record of Building SmoLLM from Scratch
An ML engineer trained SmoLLM, a 109M-parameter LLaMA-style model from scratch for under $50. Full breakdown of architecture, training pitfalls, instruction tuning, and real-world performance.
Read more →