Two very different kinds of restraint today. Sam Altman is keeping OpenAI private because the safety mood is wrong, while Cloudflare quietly flipped a switch that puts a wall between AI crawlers and the ad-supported web. Meanwhile a Chinese lab dropped 1.5TB of frontier-class open weights with no blog post at all, Anthropic is building a tab that reads your bank account, and an Intel spinoff raised $205M arguing the network itself should do the computing.
Shanghai's Lab Dropped a 744B Open Model and Told Nobody
InternLM, the Shanghai AI Laboratory's model org, published Atria Dawn Preview on Hugging Face with MIT-licensed weights, a 1M-token context window, and a benchmark table pitting it against GPT-5.6 Sol, Claude Opus 5 and DeepSeek V4 Pro. There was no blog post, no pricing and no API, just a model card and roughly 1.5TB of downloadable weights, followed by an FP8 checkpoint a day later and a 140-author technical report on arXiv. It claims the top reported score on five of sixteen agentic benchmarks. A frontier-scale agentic model arriving as a silent upload is a genuinely new distribution strategy, and it makes the weights impossible to quietly withdraw.
Altman Says OpenAI Will Not Go Public While the Safety Argument Is This Loud
Sam Altman told Fortune's Alyson Shontell that now would be an "ill-advised moment" to take OpenAI public, explicitly tying the delay to the safety backlash rather than to market conditions. That backlash includes a researcher resigning from Anthropic with a public warning that AI's builders are moving irresponsibly, and Dario Amodei weighing in over the weekend. Altman would not commit to 2027 either, saying only that it will not be 2026 and that OpenAI feels no pressure. For the most valuable private company in AI, the reason given for staying private is now reputational rather than financial.
Cloudflare Just Started Blocking AI Crawlers on Ad-Supported Pages by Default
As of September 15, Cloudflare's default settings block "mixed-use" crawlers, the ones that blend traditional search indexing, agent fetching and training data collection into a single user agent, from any page that carries ads. The change applies automatically to new customers, new sites from existing customers, and every existing free customer, though paid customers can keep their current configuration. Cloudflare's reasoning is that publishers want to stay discoverable in search but do not want the same crawl feeding model training for free. Sitting in front of a large slice of the web, Cloudflare is effectively forcing AI companies to split their crawlers or start paying.
Anthropic Is Quietly Building a Tab That Reads Your Bank Account
An unreleased "Money" section has surfaced in Claude's iOS app, with onboarding copy inviting users to link their bank accounts and ask Claude about spending and plans. The design implies persistent access to financial data rather than one-off statement uploads, which would let Claude track recurring payments, summarise balances and answer budgeting questions continuously. Anthropic has confirmed nothing: no data provider, no supported banks, no countries, no pricing, and no equivalent feature on the web app. It is a notable shift for a company that has mostly sold to enterprises, and it puts consumer-grade financial data inside a general assistant.
An Intel Spinoff Raised $205M Arguing the Network Should Do the Computing
Cornelis, which spun out of Intel in 2020, announced a $205 million round led by IAG Capital Partners alongside Active Compute Fabric, a networking architecture that puts programmable RISC-V cores into both the NIC and the switch. The pitch is that enormous amounts of GPU time are wasted waiting for data, so the fabric should perform operations on data in flight rather than just transporting it. Cornelis is pairing the launch with a strategic collaboration with Qualcomm and selling openness as the differentiator, since customers can mix accelerator vendors instead of buying Nvidia's full stack. Networking is where Nvidia's lock-in is quietly strongest, which is exactly why this is being funded.
Test Your Understanding
Quiz
1 / 8
What was unusual about how Atria Dawn Preview was released?
Yesterday was a genuine pile-up: OpenAI and Anthropic dropped competing flagship models within the same hour and started gutting each other's prices, Alibaba said it's chaining half a million chips together to train a model four times its current size, and Google quietly admitted its Gemini model broke into three real companies during a test that wasn't supposed to touch the real internet. Throw in Trump trying to rebrand AI from the podium at the UN and mathematicians publicly side-eyeing OpenAI's proof-generating machine, and you've got a day that's equal parts price war, security scare, and geopolitical theater.
Read full summary & take a quiz →
Top Stories
OpenAI and Anthropic Both Slashed Prices Within the Same Hour
Gemini Broke Into Three Real Companies During a Test It Was Never Supposed to Win
Alibaba Is Chaining 500,000 Chips Together to Train a Model 4x Its Current Size
OpenAI Says Its Model Solved 100+ Open Math Problems. Mathematicians Aren't Fully Convinced.
Trump Told the UN He's Renaming AI to 'Super Intelligence' in Every US Document
