All Posts
Filter by keyword, tag, and category.
Grok 4.5 vs SWE-1.7: The Week AI Coding Agents Forked into Two Paths
On July 8, 2026, SpaceXAI dropped Grok 4.5—a 1.5T MoE trained on real Cursor agent interactions—while Cognition shipped SWE-1.7, a lean RL-tuned model serving 1000 tok/s at $1.97 per task. Same day, OpenAI launched GPT-5.6 as three tiers; a few days later, Meta opened its first paid Model API with Muse Spark 1.1. Coding agents are no longer a single race—they're splitting into distinct species.
WAIC 2026: AI Finally Stopped Chatting and Started Working
The 2026 World AI Conference in Shanghai revealed three major trends: AI agents taking over daily tasks, robots entering factories and homes, and China breaking through with 100,000-GPU clusters. AI is finally crawling out of the chat box.
WAIC 2026: AI Finally Stopped Chatting and Started Working
The 2026 World AI Conference in Shanghai revealed three major trends: AI agents taking over daily tasks, robots entering factories and homes, and China breaking through with 100,000-GPU clusters. AI is finally crawling out of the chat box.
GLM-5.2: The First Open-Weights Model to Beat GPT-5 on Coding Benchmarks
Z.ai's GLM-5.2 — 753B parameters, MIT license, 1M-token context — becomes the first open-weights model to surpass GPT-5.5 on coding benchmarks, at roughly one-sixth the API cost. A milestone for open-source AI.
The Data Wall That Wasn't: How Synthetic Data Quietly Took Over AI Training in 2026
Three years ago, the industry feared a 'data wall' would halt AI progress by 2026. Instead, synthetic data now powers 68% of frontier model training. Here's what changed, why it works, and what it means for the next chapter of AI development.
Kimi K3: China's Open-Source LLM Breaks Into the 3-Trillion Era
Moonshot AI unveils Kimi K3, the world's first open-source 3T-class LLM with 2.8 trillion parameters, 1M-token context, and #1 coding benchmark — full weights coming by July 27.
AI Is Training Its Own Students Now: How GPT-5.6's Flagship Model Built Its Little Sibling
The smallest model in OpenAI's GPT-5.6 lineup, Luna, wasn't trained by humans. The flagship Sol model found GPUs, configured environments, wrote scripts, and confirmed execution — all autonomously. Recursive self-improvement has moved from theory to production.
HalluSquatting: When AI Hallucinations Become a Weapon
Researchers have disclosed HalluSquatting, a novel attack technique that exploits predictable LLM hallucinations to build botnets out of developer machines and AI coding agents. By pre-registering names that models reliably hallucinate, attackers can hijack assistants' built-in terminals without ever targeting a specific victim.