Discover Amazing Content, Share Life Moments
Connect Our Wonderful World
New Discovery by Anthropic: Large Language Models Have an Internal "Thinking Workspace" — What Is J-space?
Anthropic's latest research finds that large language models like Claude have an internal structure similar to the human "global workspace" called J-space. It holds the reportable, controllable intermediate concepts that the model uses for reasoning, even when these concepts never appear in the model's final output. This article breaks down the study's key findings, experimental methods, and what this discovery means for understanding AI consciousness.
How Grok ran on 100K GPUs in 23 minutes, and an auto-tuning loop for RAG
xAI served Grok on 100K GPUs using SGLang, got it done in 23 minutes, at a cost so low it makes DeepSeek API look embarrassing. Meanwhile, a self-tuning loop lets you say goodbye to the nightmare of manual RAG tuning — all code is open sourced.
100M Parameters, Runs on CPU, 6x Real-Time Speed: Kyutai Labs Releases a Lightweight TTS That Eliminates the Need for a GPU
Kyutai Labs has quietly launched Pocket TTS, a lightweight text-to-speech model with 100 million parameters that runs on CPU, featuring 200ms latency and requiring only 5 seconds of audio for voice cloning. It is open-sourced under the MIT license, allowing free commercial use. This is arguably the most practical implementation of local AI speech synthesis to date.
Tencent Hunyuan Launches Hy3: 295B Total Parameters, 21B Active Parameters, Performance Rivaling Trillion-Parameter Models — But The Real Highlight Is Reliability
Tencent Hunyuan has released Hy3, a 295B MoE model with 295B total parameters and 21B active parameters. Its performance matches that of trillion-parameter flagship models. With three iterations in six months, it has cut hallucination rates by half and delivers stable tool calling. It is open-sourced under Apache 2.0, with free API access available for two weeks.
Ran Out of Claude Quota Again? This Open-Source Tool Automatically Switches You to Free Models
An open-source tool called 9Router lets you automatically switch between more than 40 AI providers. When your Claude quota runs out, it automatically falls back to cheaper models, and switches to free models if those are also unavailable. It supports Claude Code, Cursor, Copilot and more. Setup takes just two steps, and it's completely free.
DeepTutor v1.5: Build Tutoring Around a Data Cycle, Not a Pile of Features
A team from the University of Hong Kong has open-sourced DeepTutor v1.5, an agent-native learning workspace. Its core belief: Tutoring should be a continuous data cycle, not disconnected features. It unifies instruction, practice, behavior tracking, inspectable memory, and active IM partners, all running on the same agent loop. It is fully open-source, and has already earned 25.2k stars on GitHub.
35B MoE Runs at 79 Tokens/s Locally, Outcoding Claude Sonnet 4.5? Real-World Performance of Qwen 3.6 on DGX Spark
llama.cpp benchmark results for Qwen 3.6-35B-A3B on DGX Spark: 79 generation tokens/s with 2k context, dropping to 31 tokens/s with 256k context. MiaAI Lab released a one-click startup script. More counterintuitively, a locally 8-bit quantized 27B model outperformed Claude Sonnet 4.5 in overall coding test scores.
When It Comes to AI-Powered SaaS Automation, Even the Best Model Only Solves Half the Problem
Zapier and Artificial Analysis have released the independent evaluation AutomationBench-AA: Claude Fable 5 leads with a score of 48.6%, just 0.1 percentage points higher than Opus 4.8. All models violate business rules, and Gemini 3.5 Flash stands out for its exceptional cost-performance. Finance tasks are the hardest, with models completing only about one-third of objectives on average.
Reddit Uses LLMs to Kill the Garbage Created by LLMs – Makes Sense
Reddit's LLM-powered tool blocks 23 million spam views daily, most of which are also generated by LLMs. Fighting fire with fire, and it's working reasonably well.
2026 Tech Layoff List: Is AI an Excuse or the Truth?
In 2026, tech companies are hitting record revenues while conducting massive layoffs, with AI being the most frequently cited reason. Giants like Microsoft, Google, Meta, and Oracle are cutting headcount, but is AI really replacing humans behind these layoffs? This list documents all major layoff events this year where AI was cited as a reason—dense with information and worth a close look.
