Everyone's bragging about 200k, 1M, even 10M token context windows in 2026. Nobody's talking about what actually happens inside them. I just fell down this rabbit hole and I can't stop thinking about it: A new paper published January 2026 tested hundreds of thousands of data
LLMS
-

Open Source Data: Essential Missing Component for Large-Scale Training
By
–
More Open Source Data. The main missing ingredient for large scale training.
-
40 years of code: Generative AI, time-saving but errors
By
–
Dear Rafik, I've been coding for 40 years, I'm 52 years old and I started at 12. I use all generative AI coding tools daily, I admit it's fantastic, I save an incredible amount of time, it's undeniable, but they all code with errors, I have to go over everything,
-

Overview of Loss Functions in Machine Learning
By
–
Machine Learning loss functions #infographic Loss Functions in ML Explained: https://
datacamp.com/tutorial/loss-
function-in-machine-learning
… Comprehensive Review of Loss Functions in Deep Learning: https://
arxiv.org/pdf/2504.04242
v1
…
————————
#AI #DataScience #DataScientist #Mathematics -
Claude Code Now Writes 4% of All Code Globally
By
–
Double pod with @fabknowledge of @SemiAnalysis_
! Claude Code for Finance + The Global Memory Shortage https://
latent.space/p/valuemule One year ago today, @bcherny released Claude Code to the world as part of the Sonnet 3.7 launch. One year later, it is now writing 4% of all code -
Multi-agent system builds from YouTube tutorials autonomously
By
–
This guy literally built a multi-agent orchestration system with Opus 4.6 to watch YouTube tutorials and execute them autonomously 😳 pic.twitter.com/hZYy1QkruD
— God of Prompt (@godofprompt) 24 février 2026This guy literally built a multi-agent orchestration system with Opus 4.6 to watch YouTube tutorials and execute them autonomously
-
LLM Performance Degradation Over Extended Context Windows
By
–
Performance degrades over the course of 100k tokens even, let alone the whole currently supported window… After a few turns of coding Python, it just can't reliably use its tools anymore. Requires constant jumping back and/or offloading.
-

GPT-5.3-Codex Live: OpenAI’s Advanced Agentic Coding Model
By
–
GPT-5.3-Codex is live on Poe! OpenAI's most capable agentic coding model combines advanced reasoning with autonomous coding workflows. 25% faster than its predecessor, built for tasks you can start and walk away from. Try it alongside every top model in the Poe app and API at
-
Inception’s Mercury 2: The First Reasoning Diffusion LLM
By
–
BREAKING 🚨: Inception has launched Mercury 2, the first reasoning diffusion LLM with 5x the performance of top-speed-optimised models.
— 🚨 AI News | TestingCatalog (@testingcatalog) 24 février 2026
Early access OPT-IN is open 👀
Free public testing available too 👀 https://t.co/mObq0kpiMK pic.twitter.com/wD4B1NhfVtBREAKING : Inception has launched Mercury 2, the first reasoning diffusion LLM with 5x the performance of top-speed-optimised models. Early access OPT-IN is open Free public testing available too
-
Perplexity AI Shifts to Subscription-Only Model to Prioritize Trust
By
–
Perplexity's model is subscription-heavy ($20/mo Pro, enterprise seats $40+), now fully ditching ads to protect AI answer trust—no sponsored slop. Revenue (~$200M ARR) comes from users/companies who pay for accuracy, citations, and pro features like Deep Research. Aligns
