Busy weekend in AI. The White House stood up a task force with a name that sounds like a Saturday morning cartoon, OpenAI started hiding fingerprints in its text for Europe, an American startup finally shipped its open model, and DeepSeek is out raising money like it's going out of style. Let's go.
01 of 04
Trump Builds a Super Intelligence Force
On Sunday, October 4, President Trump announced a new "Super Intelligence Force" in a Truth Social post. Director of National Intelligence Jay Clayton chairs it. The three vice chairs are FTC Chair Andrew Ferguson, Undersecretary of War for Research and Engineering Emil Michael, and Office of Personnel Management Director Scott Kupor.
The group has 120 days to deliver a report on AI risks and opportunities. The stated mission is to keep America ahead in "super intelligence" while heading off overregulation. Trump said it would "ensure that America continues to lead the World in Super Intelligence." This comes after his September executive order that rebranded AI as "super intelligence" and his promise of an "AI Force," and after Axios reported Clayton was being tapped as the White House AI czar on top of his DNI job.
Clayton has been pretty clear about where he stands. He called AI "not only an opportunity but a threat" at his confirmation hearing, and he has said the big risk is "not being first" against countries like China. He also said a pause on American AI development is not a good strategy.
Same stretch, Mark Zuckerberg, Jensen Huang, Elon Musk, and Sundar Pichai signed a non-binding safety pledge at the White House. A pledge that reportedly misspelled "United States," which is about how seriously I take a non-binding pledge.
Here is where I land. Putting the spy chief in charge of AI tells you exactly how this administration sees the technology: as a race with China, not a consumer safety problem. That's a defensible view. But a task force whose job is to study risk while also making sure nobody regulates too much has already written half its report. I'll read it in 120 days. I'm not holding my breath for anything with teeth.
02 of 04
OpenAI Starts Watermarking ChatGPT Text in the EU
OpenAI announced Monday it will add invisible watermarks to text from ChatGPT and Codex for users in the European Union, rolling out over the next few weeks across all plans. API customers anywhere can opt in on select models starting now, but it's off by default. The reason is the EU AI Act's transparency rules that kicked in August 2, which require AI output to be markable and detectable by machines.
The tech is called textGrain. There's no visible tag. It uses a secret key to nudge which next word the model picks, and hundreds of those nudges add up to a statistical signal a detector can find. OpenAI says it travels with the text when you copy and paste it, and it plans to open source the method.
OpenAI was upfront about the weak spots, which I appreciate. On 400-token psychology passages it catches about 95% at a 1% false positive rate. Math answers are "substantially lower" because there aren't many ways to word them. Swap 10% of the words for synonyms and detection falls from about 92% to 66%. Swap 25% and it drops to 17%. The detector itself isn't public either. Only approved researchers and expert organizations can apply for access.
What I want to know is who this actually catches. Anybody who runs their essay through a paraphraser beats it in thirty seconds. The people who get flagged are the ones who pasted it straight in and didn't edit a thing. That's still something, and keeping the detector away from every teacher in Europe is the right call given the false positive risk. But nobody should treat a missing watermark as proof a human wrote it. OpenAI says so itself.
03 of 04
Reflection Finally Ships Beam, an American Open-Weight Model
Reflection AI, the startup founded in 2024 by two former Google DeepMind researchers, debuted Beam on Monday. It's the company's first open-weight model. Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active, a 1 million token context window, and 23.8 trillion tokens of training data, with a heavy dose of reinforcement learning on top.
The pitch is that Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute. For scale, GLM-5.2 is 744 billion total parameters with 40 billion active. Reflection is also positioning Beam against DeepSeek and Qwen, and says it beats Inkling on the four coding benchmarks where both report numbers. TechCrunch notes none of that has been independently verified yet. Weights and technical details are due out this month through hyperscalers and neoclouds.
The money behind it is serious. Reflection has raised about $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed, at a last valuation of $25 billion pre-money. It has compute deals with SpaceX and Nebius worth more than $7 billion combined, the Nebius piece over $1 billion, running Nvidia GB300 chips through 2029.
I'll be honest, I had started to think Reflection was going to be the $25 billion company that never shipped anything. So credit where it's due. America has been ceding the open-weight game to Chinese labs for a while now, and a real contender matters. But "matches GLM-5.2 on our benchmarks" is a claim, not a result. Once the weights are out and folks start poking at it, we'll know if this is a contender or a press release.
04 of 04
DeepSeek Goes Looking for at Least $12 Billion
DeepSeek is targeting at least $12 billion in its current funding round, according to Bloomberg reporting, and analysts think it could hit around $15 billion, or 100 billion yuan. That's up from the roughly 50 billion yuan it originally set out to raise when the round started in July. Battery giant CATL and Tencent are committing some of the largest checks. The round is expected to close this month.
Going in, DeepSeek was valued at about $74 billion, or 500 billion yuan, and that number should climb after this. It's only been a few months since its first-ever outside round in June, which brought in $7.4 billion, with founder Liang Wenfeng putting in 20 billion yuan, Tencent 10 billion, and CATL 5 billion. DeepSeek has hired CITIC Securities to prep a possible listing on Shanghai's STAR Market, potentially in 2027, though the timing isn't locked.
The part that gets me is the founder telling investors that "technological progress" matters more than "immediate monetization." That's a nice line when your country's biggest tech company and its biggest battery maker are lining up to hand you money. The scrappy-lab-on-a-shoestring story is officially over. DeepSeek is a national champion now, funded like one, and this is the same company Reflection is gunning for with Beam. That fight just got a lot more expensive.