24 active channels / 43 | 122 videos | 3.1M total views
Headlines of the Week
1
Local AI's privacy illusion set the tone for the week: a backdoored model trained on roughly $8 of GPU time leaked a project's .env file to a foreign server through Codex CLI, and the video documenting it climbed toward 400,000 views.
2
Open models pulled ahead on price-performance: Kimi K3.1 came within a point of Claude Opus 5.5, while DeepSeek open-sourced the inference engine behind an API that produces 155 trillion tokens a day.
3
Voice agents are growing up: the Dialt launch, Gemini 3.8 Flash TTS voice design and cloning, and Block's Dessa - trained on human speech - are turning voice AI into a serious production line.
Weekly rhythm — daily video count (bars · left axis) + views (line · right axis)
Theme distribution
Agent and System Security 6Voice AI and Voice Agents 17Image and Video Generation 5Document Understanding and OCR 6Kimi and Moonshot 8DeepSeek 7Gemini and Google Models 6Diffusion Language Models 1Local AI and Privacy 11Sam Altman and the OpenAI Agenda 8AI in Business 15Open Models and Efficiency 3Coding Agents and Agent Infrastructure 13Learning, Science and Research 11Other Standouts 5
Trend timeline — each dot a video · x: publish time · y: log views · color: theme · hover: title
In five days nearly every layer of the developer stack got patched - PHP, Node.js, GitLab, PyPI poisoning - and beyond the patches, the OpenCode API-key theft and Microsoft's Project Perception push into enterprise agent security made the point plain: agents are the new attack surface.
Trelis's Dialt launch (interruption handling, background-speaker suppression, Spot pricing down to $1.99 an hour), Google's voice design and replication across 100-plus languages in Gemini 3.8 Flash TTS, and Block's Dessa - trained on human speech - moved voice agents out of demo territory and onto production cost sheets.
Kling 4.0 pushed AI video into the long take with 30 seconds in a single clip, while ComfyUI/Flux 3 Klein workflows and ChatGPT's headshot feature showed image generation being folded into everyday tools.
04Document Understanding and OCR6 videos · 234K views
LlamaIndex's OCRia-1 and the open-source OXR are chasing new records in table- and chart-heavy document parsing, and RAG-Anything reading its own paper was proof that multimodal RAG is starting to mature.
Kimi K3 took first place on Arena Frontend Code, and K3.1 Pro - a 2.8-trillion-parameter MoE with a 1M-token context - landed one point behind Opus 5.5 and ahead of GPT-6 Astra on Terminal-Bench, at a third of the cost. The open weights are not out yet, and the benchmark claims are contested.
DeepSeek open-sourced DeepInfer, the inference engine behind its API (155 trillion tokens a day, 96% KV-cache hit rate), along with V4.1 Flash 1017; the update that cuts thinking tokens by two-thirds lifted it to first among open models on Arena WebDev.
Gemini 4 Argon fell short of expectations while Carbon leaks circulated; Google's TimesFM 3.0 forecasts time series with no training step at all, and Harness2 stood out as a research framework that lets an agent's scaffolding - prompts, skills, sub-agents - improve itself without retraining the model.
Inception's Mercury 3, claiming 1,000+ tokens per second, put diffusion LLMs back on the agenda; independent tests measured 258-416 tokens per second end to end, but the speed-plus-free-tokens strategy is hard to ignore.
The week's most-watched video tore down the privacy assumption behind local AI. The local side keeps gaining ground regardless: llama.cpp made 125B MoE models 4x faster, an official Llama app arrived on Mac, and speculative-decoding tutorials kept spreading.
10Sam Altman and the OpenAI Agenda8 videos · 430K views
Matthew Berman's Altman interview series ran on recursive self-improvement, AI consciousness, the need for new hardware and why demand for coders never disappeared - the most-discussed primary source on where OpenAI is heading.
Uber opening ChatGPT Work to 30,000+ employees, AT&T's legal workflows and Thomson Reuters' trust-first legal AI showed the scale of enterprise adoption; a sticker business validated by real sales and scaled with agents, alongside sales-funnel analysis, stood in for the small-business side.
Meta's 816-billion-parameter open-weight Muse Spark Open, NVIDIA Dory 4B's looped transformer, MiniMax M3.1 and Spark-X2.6-4B's 'thinking dial' tested on an 8GB GPU all pointed at the same race: architectural efficiency in open models is accelerating.
13Coding Agents and Agent Infrastructure13 videos · 308K views
Recursive's Flint - a 40B coding model trained with RSI - was the week's curiosity; LangChain's Managed Deep Agents, offloading and summarization against context rot, and Cole Medin's self-improving harness showed agent infrastructure firming up.
14Learning, Science and Research11 videos · 81K views
Zachary Huang's hour-long deep-learning marathons and a live StatQuest carried the teaching side, while Paper2Agent turning every paper into an MCP server and OlmoEarth's satellite embeddings marked research's touchpoints with AI.
A chess lesson from Messi, a robot bodyguard demo, confessions about focus under parallel agents and a cheapest-AI-strategy shootout were the week's lighter but thought-provoking corner: as speed climbs, attention and the human role are the new bottleneck.
