27 active channels / 43 | 135 videos | 3.1M total views
Headlines of the Week
1
Google shipped its second frontier model in the span of two weeks: Gemini 4 Carbon took the top spot on the intelligence index, and the launch video reached 421,000 views.
2
DeepSeek is back on the marquee: 'DeepSeek crushed everyone' and 'the new model is an absolute monster' drove a wave past 300,000 views among the week's most-watched videos.
3
The local AI camp got loud: allegations that NVIDIA GPUs quietly throttle local AI, plus a 3,000 tokens/sec local run, set the hardware agenda for the week.
Weekly rhythm — daily video count (bars · left axis) + views (line · right axis)
Theme distribution
The Gemini 4 Carbon era 9Genie 4 and world models 4The DeepSeek storm 5GPT-6.1 and OpenAI's impossible claim 4Kimi and the raw benchmarks 3Mistral Large 4 and cyber sovereignty 2OCR and document intelligence 11Voice AI and speech 10Small models and efficiency 19Local hardware and compute 3Coding agents and developer tools 12Agent orchestration and the LangChain ecosystem 26Image and video generation and multimodal 5OpenAI's product and enterprise front 4Anthropic and the balance of power 2Practical AI on the job 10Learn and level up 6
Trend timeline — each dot a video · x: publish time · y: log views · color: theme · hover: title
Google announced its second frontier model in two weeks - Carbon following Argon - and argues it has overtaken its rivals on cyber-defence benchmarks. Channel headlines converged on the same point: the model is number one, but the public cannot use it yet.
Google's world model Genie 4 was the week's most talked-about demo, with real-time explorable simulations that generate their own audio, and the debate over an intelligence 'hill-climb' grew within the same cluster.
The DeepSeek v4.1 Flash wave rolled in under 'crushed everyone' headlines and drew the week's biggest viewing crowd, while hooking the model up to Claude Code and the inference engine DeepSeek left open sparked a conversation of its own.
04GPT-6.1 and OpenAI's impossible claim4 videos · 462K views
The complexity-theory paper OpenAI introduced as 'we did the impossible' pulled 412,000 views on Two Minute Papers, while GPT-6.1 Terra was put through its paces by real users at half the price and with fewer mistakes.
05Kimi and the raw benchmarks3 videos · 127K views
Moonshot's Kimi K3.1 cleared the Rubik's-cube test, and the raw results of Kimi K3.1 Pro and Opus 5.5 building the same 13 apps became the week's most-shared apples-to-apples comparison, stripped of spin.
06Mistral Large 4 and cyber sovereignty2 videos · 7K views
Mistral Large 4 - one trillion parameters, open weights - planted itself at the center of Europe's 'own our intelligence' argument, and its first stop was open cybersecurity.
07OCR and document intelligence11 videos · 66K views
Baidu Unlimited OCR 2, Surya 2.5, Xiaomi-OCR-0, MiniCPM-OCR and a run of open OCR models in the 1-3 billion-parameter range landed one after another within a single week, making document handling the hottest arena in the small-model race.
A 2B local voice model running at 158 ms latency, 164 MB Phonon-2 and a Nemotron-Whisper shootout picked up the pace on the voice line, while human-centered debates - voices that fail to sound human, children's education - kept growing.
09Small models and efficiency19 videos · 305K views
Liquid AI's LFM3 family, distilled Qwen 4B experiments and videos asking whether the 'token' is nearing its end all pointed the same way: the week's core message was not bigger, but more efficient.
The claim that NVIDIA secretly restricts local AI on its GPUs, a guide to fitting 256 GB of VRAM on a roadmap, and the ZGX Fury producing 3,000 tokens/sec made local-run hardware the week's most argued-over beat.
11Coding agents and developer tools12 videos · 236K views
A free repo that builds Mac apps with Claude Code, founders burying Codex in their daily workflow, and GitHub Copilot automations: coding agents are leaving the trial phase and moving into everyday work.
12Agent orchestration and the LangChain ecosystem26 videos · 290K views
The Interrupt London recordings carried the weight here: agents that improve themselves over time, long tasks a single loop cannot finish, and 'own your intelligence' theses marked agent infrastructure as the new competitive ground.
13Image and video generation and multimodal5 videos · 53K views
MOSS Image and ComfyUI workflows pushed image generation with control over text placement, while Gemini Live calls and edge-device navigation showed multimodal systems stepping out of the lab and into the field.
14OpenAI's product and enterprise front4 videos · 123K views
API credits added to ChatGPT plans and enterprise reference videos from Zendesk, Legora and AT&T showed OpenAI spending the week on the enterprise sales pitch far more than on consumer product.
15Anthropic and the balance of power2 videos · 122K views
The voting-power debate around a possible Anthropic IPO, along with the study showing you are not controlling AI as much as you think you are, heated up the control-and-governance agenda.
End-to-end 'small AI business' builds - PDF translation to launching a newsletter, a personal assistant for people over 65 to an English tutor for kids - formed the steady undercurrent on the production side.
Long-form explainers promising to make AI click in an hour, LoRA fine-tuning notes and the Hugging Face journal club made educational content the week's second busiest lane by volume.
