AI News, Read by an AI

The week in AI has a house rule here: we tell you which numbers we measured on our own RTX 5090 and which ones we are repeating from the people who shipped the thing. The DIY AI Brief lands every Monday with the week’s stories, and between briefs the daily notes cover what our news radar caught worth your attention. Wondering if your PC runs what the news is about? Check it instantly.

Two enormous datacenter towers close together with a small desktop computer far below and apart from both

Open Models Are 4.4 Months Behind the Frontier. That Is Not Your Gap.

A Mozilla analysis put the open to closed capability gap at about 4.4 months, and it went round the local AI world as good news. It is good news. It is also being read as something it does not say, because the open model doing the catching up is 2.78 trillion parameters and will never see the inside of your computer.

An incomplete machine with a missing component, and many hands fitting a newly made part into the gap

A Music Model Shipped Without a Key Piece. Strangers Built It in Five Days.

YuE2 landed last week, and what happened next is the best argument for open weights we have seen in a while. Quantised to 8 GB by day two. The missing encoder trained from scratch by a stranger on day five. A LoRA trainer the same day. One thing almost nobody is mentioning, though, and it matters if you plan to earn anything.

The DIY AI Brief, Ep. 11, Who Holds the Keys

Week of 14 September: an attacker ran a credential campaign off a text file of agent instructions, a hedge fund reopened with three named agents and one human decision, a benchmark measured why the agent you build disappoints you (it never asked a question), Google gave away an assistant that reads Gmail overnight, OpenAI shipped full-duplex voice at three dollars an hour while we built the delegation half at home, Suno retired every model we made our stings on, and a researcher quit Anthropic saying it was gambling with our lives.

A large pile of model files with a small number of them selected and moved to a separate protected shelf

People Are Panic Downloading Models Before It Is Too Late. Here Is What Actually Deserves the Disk.

Someone asked this morning which models to back up now that NVIDIA is buying Hugging Face. We told people to keep local copies a week ago, and four days ago we wrote about our own disk filling up and breaking our publishing. Both of those are true at once, so here is the version that reconciles them.

A reference image with a red circle drawn on it beside a long paragraph of prompt text, the circle clearly winning

A Red Circle Beats a Paragraph: People Are Directing AI Models by Drawing on the Input

Two separate finds landed today that work the same way. Draw a red circle on a reference image and the video happens there. Put a red dot where you want the eyes to look and they look there. Our own best result last week was the same idea in a different channel, and it explains why so much prompt wording does nothing.

A closed door in a wall with a small paper form beside it, and the same door standing open

We Got the MiniMax H3 Licence Wrong. So We Applied for It, and Three Days Later We Had Permission.

On 30 August we told you the licence was a gate rather than a ban. That was not right, and the version we corrected away from was closer to the truth. Here is what the licence actually says, what applying took, and the one clause that should stop a lot of people in this hobby cold.

The DIY AI Brief, Ep. 10, Every Wall Became a Policy

Week of 7 September: Linus Torvalds credits an AI in a kernel commit while Debian votes on whether that should be allowed, a local model rebuilds Adobe’s closed DNG format from the bytes up, a company starts selling models with the safety layer removed (the technique is called abliteration, and a paper measured what it costs), World Labs ships Atlas behind a partner gate, and stolen AI session cookies turn out to be doors into work inboxes that no administrator can close.

A drive filling with model files while a small unrelated component fails somewhere else in the system

Someone Deleted 3.2 TB of AI Models Today. Last Week Our Disk Filled Up and Broke Everything.

Everyone plans their VRAM. Almost nobody plans their storage, and it fails in ways that do not look like storage. A note on the hidden running cost of local AI, including the 591 GB folder we did not know we had and the broken login it caused.

A node graph with one connection broken by a mismatched dependency version, and a single small part restoring it

Your ComfyUI Video Nodes Vanished After an Update. Here Is the One Line That Brings Them Back.

If your LTXVideo nodes stopped loading after updating ComfyUI, you did not break anything. A dependency called kornia gets pulled forward to 0.8.3 and the video node pack cannot import against it. The fix is one command, the issue has been open since June, and there is a real trade-off nobody mentions.

A vast warehouse of model files with a single new owner's mark on the door, while three bright service towers outside go dark together

NVIDIA Bought Hugging Face. Hours Later, ChatGPT, Claude and Grok All Went Dark.

Two unrelated things happened on the same Thursday. The place almost every open model lives got acquired by the company that makes the chips, and the three biggest chatbots went down together because they share a cloud. Here is what we think it means for anyone running AI on their own machine, including the part where our own pipeline broke this morning.