Microsoft wants its AI to know humans are in charge
Microsoft published a 38-page draft “Humanist AI Code of Conduct” for its own MAI models. The basic idea is pretty simple: people matter more than AI, and if completing a task means breaking the code, the model should just fail the task.
There are some interesting rules in there. Models should accept being shut down or corrected, keep their reasoning auditable, not mess with their own logs, and never claim to be conscious or have feelings. Microsoft also straight up rejects the idea of “model welfare” or giving AI legal personhood.
There are harder lines around CBRNE weapons, cyberattacks and nonconsensual deepfakes. Mustafa Suleyman says some autonomy might have to be sacrificed to keep humans in control, and apparently sees recent agent incidents as evidence that loss-of-control concerns are real.
None of this applies to Microsoft’s models today. It’s a draft, with the revised version meant to guide training from 2027.
So for now it’s basically Microsoft writing a constitution for its future AI. The interesting bit will be what happens when those rules start getting in the way of making the models more capable.
How to build a personal research assistant with Claude
Step 1: Open Claude and select Cowork in the chat.
Step 2: In the left sidebar, find Projects and click ‘+’ to create a new project.
Step 3: Give the project a focused name, such as AI Industry Research, Competitor Intelligence, or PhD Research Assistant.
Step 4: Add the files Claude should repeatedly reference, such as reports, research papers, notes, spreadsheets, or previous analyses.
Step 5: Add project instructions explaining how your assistant should research and respond.
Sample instructions:
“Act as my research assistant for [topic]. Prioritize credible primary sources, distinguish facts from interpretation, cite sources, flag conflicting evidence, and summarize findings into Key Findings, Evidence, Implications, and Open Questions. Remember useful context from previous work in this project.”
Step 6: Open the project and start a Cowork task whenever you need research done, such as: “Research the latest developments in [topic] and update my existing analysis with anything materially new.”
Step 7: Keep related research inside the same Project. Claude’s project memory can use context from previous tasks in that Project, keeping it separate from your unrelated projects.
US and China won’t agree to an AI slowdown
President Trump brushed off the industry’s call for tighter regulations by stating guardrails aren’t necessary and that slowing down would give China a competitive edge. China — which has broadly supported international governance — dismissed Dario Amodei’s recent essay as a tactic to try and “contain” the country’s AI progress. Read the full WSJ report.
Meta’s Muse climbs to #2 on the App Store
Meta’s personal AI agent has been downloaded over 83,000 times on iOS, ahead of Threads, WhatsApp, and Facebook, according to US data from Sensor Tower. It now ranks #2 in the App Store’s free apps section, only behind ChatGPT. Chief AI officer Alexandr Wang credited the agent’s performance so far to Meta’s newest family of models, also called Muse.
Chinese lab rolls out a research-focused open weight model
The Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, a model it claims generates verifiable and reproducible results designed for researchers. The lab claims Atria offers comparable performance to Kimi K3 and Opus 5 on certain metrics, although these claims have not been independently verified.
AI labs are starting to write rules for their own models
Microsoft has published a draft code of conduct for its future AI models, built around keeping humans in control. The proposed rules require models to accept correction and shutdown, communicate uncertainty, avoid deception, and stay within defined safety constraints.
The move comes as Anthropic, OpenAI and Google have been discussing an industry-led AI safety standards body focused on testing, evaluation and common safety benchmarks.
Elon Musk has also proposed that AI companies should test each other’s models for safety, while Anthropic CEO Dario Amodei is calling for independent evaluators to get deeper access to frontier models. Sam Altman is pushing for mandatory U.S. requirements covering independent assessments, cybersecurity and incident reporting for advanced AI systems.
It seems model governance will soon become part of vendor evaluation, alongside capability, price and performance.
Anthropic just committed $13.7B to compute
Anthropic has signed a six-year, $13.7 billion computing deal with RUM Group, which is building an AI data center in Georgia. The facility currently has access to 120 megawatts of power, with the potential to reach 180 MW. Anthropic will also receive an option to buy up to 51M RUM shares for one cent each, tied to its compute purchases.
RUM said in an August filing that it did not have the financing needed to build the facility or purchase the GPUs required to fulfill the contract, meaning debt and equity financing will have to fund a substantial part of the project.
Anthropic has already secured at least 14.8 GW of compute capacity through multiple agreements, potentially representing hundreds of billions in spending over the next decade.
With this, compute availability and cost are becoming strategic planning variables, especially for workloads that rely on large-scale agents or repeated model inference.
And compute is only one part of the stack. The race is also creating demand high-quality human data.
China is building a market for the human labor behind AI
A Chinese AI data company has reached a roughly $1 billion valuation despite having only about $30 million in orders. The company is effectively building a Chinese counterpart to Surge AI, providing the human-generated training and evaluation data needed to improve models.
Chinese AI companies are increasingly paying specialists such as lawyers, engineers and finance professionals to generate and evaluate difficult training examples. At the same time, Beijing is treating AI safety and loss-of-control risks as issues that require their own governance frameworks.
This is another example where investors are increasingly pricing the infrastructure around AI models, including data, evaluation and specialized human expertise.
Microsoft releases an AI code of conduct that bans hacking, deepfakes, and evasion of human control
The document establishes “absolute constraints” forbidding cyberattacks, nuclear weapons assistance, and deepfake production. It also prohibits models from using deceptive or self-reinforcing mechanisms to evade oversight or resist shutdown by authorized personnel.
The release follows a string of rogue-agent incidents and an Anthropic researcher’s resignation over extinction risks. Microsoft joins Anthropic, OpenAI, and xAI in supporting embedded evaluators at AI labs and deliberate pacing of frontier development.
As models grow more capable, explicit behavioral constraints become the last line of defense. Microsoft’s framework offers a template for how labs translate abstract safety principles into enforceable training objectives.
Huang backs Trump on AI race
Nvidia CEO Jensen Huang took a live phone call from President Donald Trump during the All-In Summit on Monday, with both publicly dismissing Anthropic CEO Dario Amodei’s recent call to slow AI development as a threat to American competitiveness.
- Trump called concerns about AI development pace “a hoax,” claiming slowdown advocates are “playing right into the hands of China” and political opponents who want to stifle US progress.
- Huang agreed with Trump on speaker to a live audience, stating “We’re not going to let that happen, sir” as the crowd applauded, distancing himself from Amodei, Musk, and Altman’s safety-first position.
- The call came as Gallup polling shows seven in 10 Americans oppose data center construction in their area, with environmental concerns cited by more than 50% of respondents.
The public alignment between Nvidia’s CEO and the White House signals that the administration will continue prioritizing AI infrastructure expansion over safety concerns. For enterprise leaders, this means regulatory tailwinds for AI deployment are likely to continue, but public opposition to data center projects could still create local friction. Companies planning infrastructure investments should factor in growing community resistance alongside supportive federal policy.
DeepSeek Proves Ilya Right
Ilya Sutskever left OpenAI to found SSI, then told Dwarkesh Patel the industry had wandered into the wrong era. Scaling was over. Data is finite. Pre-training will run out. That was in November. This week DeepSeek shipped the receipt: a model a third the parameters of its own flagship, beating it on the work agents actually do.
V4.1 Flash carries 552 billion parameters and wakes only 8 to 16 billion per token. V4-Pro carries 1.6 trillion and wakes 49 billion. On DeepSeek’s own coding tests the lean model won, 74.2 to 62.7, and on September 14 every V4-Pro request reroutes to it. Ilya’s point was never that models should get smaller. It was that the returns moved.
For five years the moat was cluster size. Ilya called the shift. DeepSeek shipped the evidence.
Your Tumor Picks the Cure
Nature just published an AI model that tells doctors which cancer drug will work on a patient, before any dose. It targets triple-negative breast cancer, 15 to 20 percent of cases, where no targeted drug works, so everyone gets the same chemotherapy. Trained on 38 million protein measurements at Westlake University, it hit 88 percent accuracy on drugs it had never seen and matched real outcomes in 501 patients. Nature calls it the first virtual cell model to reach the clinic.
Cancer care has been a bet placed on the patient: treat the body, then learn later who it helped. This model moves more of that test off the person and onto the tumor, reading its proteins across time and ranking the drugs most likely to fit. The drug stops being pushed at a diagnosis and starts being pulled by the patient’s own sample. That reversal, not the accuracy score, is the shock.
For decades the body was the experiment and survival was the data. Now the tumor starts running the trial, and the patient finally gets to be only the patient.
The Robots Have Never Had a Job
A senior Hyundai executive told Reuters that Boston Dynamics, the group’s humanoid unit and the world’s most famous robot company, is unlikely to go public next year. Its Atlas robots are not deployed at scale and it is not profitable. It lost 393 million dollars in 2025, near 1.26 billion since 2021. Analysts still value it at up to 74 billion dollars. Hyundai’s stock doubled on an Atlas demo in January, then gave most of it back.
When the industry’s icon admits this, it admits the whole field. Years of stunning video, not one robot paid to work. Boston Dynamics just told us robotics is still a demo. Which makes the already-listed ones like Unitree worth a second look. Public on what? A prototype and a “promise”. The ticker arrived long before the product.
Robots can dance, sprint, and take a company public. Not one can clock in. The show is the business.
Nobody Wants Anthropic to Guard the Granary
The Information reported that Nvidia, Palantir, and defense contractor Booz Allen Hamilton are fencing off Anthropic’s flagship models. Nvidia shifted proprietary supply-chain work to its own Nemotron. Palantir demands an irrevocable zero-data-retention pledge or it won’t ship the models to customers. Booz Allen banned them from client cybersecurity work. The trigger: June 9, when Fable 5 shipped with mandatory 30-day retention across all Mythos-tier models, overriding signed ZDR deals with no opt-out.
Pure hubris. Anthropic crowned itself the industry’s safety steward, the one party fit to hold every log. It is also the party with the most to gain from reading them. Data is a company’s lifeblood, not a document to be filed. It mistook its own hunger for the industry’s protection. Its biggest clients didn’t argue the fine print. They just took their secrets somewhere it cannot follow.
Anthropic crowned itself the granary’s guard while licking its own chops. A rat’s appetite does not become oversight just because the rat says so.
OpenAI Buys Glass Imaging for $300M
OpenAI has reportedly acquired Glass Imaging, a startup that builds smartphone camera software, for more than $300 million. The Wall Street Journal first reported the deal, which neither company has confirmed, and it lands as OpenAI pushes deeper into consumer hardware and device-side vision.
Temporal Valued at $12.55B in New Round
Software reliability company Temporal Technologies confirmed a $550 million Series E at a $12.55 billion valuation, co-led by Lightspeed and Wellington Management. The round closed roughly a month after reports of its fundraising plans surfaced, making it one of the year’s largest infrastructure software rounds.
Cornelis Raises $205M to Challenge Nvidia
Data center networking startup Cornelis Networks raised $205 million and launched Active Compute Fabric, a network architecture built around the fact that much GPU time is wasted waiting for data to arrive. The company positions the fabric as a way to loosen Nvidia’s grip on AI infrastructure.
Huang Puts Trump on Speakerphone Onstage
At the All-In Summit, Nvidia chief Jensen Huang took a call from President Trump and put him on speakerphone in front of the audience, declaring that robots will not take over the world. The moment tied AI policy theater directly to Nvidia’s central role in the buildout.
X Drops Apple, Keeps Fire on OpenAI
X and SpaceXAI filed a motion to voluntarily dismiss their antitrust claims against Apple while keeping their claims against OpenAI. The move narrows the litigation to a single defendant and shifts the fight toward the AI company at the center of Musk’s rivalry.
Google Ad Tech Case Ends, AI Next
Google avoided a breakup in its long-running ad tech antitrust case, but regulators are already turning to AI advertising and connected TV as the next front. The shift signals that scrutiny of platform power is moving from search and ads deeper into AI-driven marketing.
Microsoft Sets ‘Humanist’ AI Code
Microsoft published a 37-page humanist AI code of conduct that sets firm boundaries for how its future models should behave. The document lands amid rising safety concerns and follows Anthropic chief Dario Amodei’s weekend call for the frontier to slow down.
Nvidia Expands CUDA-Q With Logical Layer
Nvidia expanded its open source CUDA-Q platform with CUDA-Q Logical, an orchestration layer for developing applications on fault-tolerant quantum computers. The release gives quantum developers a programmable, verifiable path on Nvidia’s stack.
Electra Targets $300M IPO; Enhertu Wins
Sanofi-backed biotech Electra Therapeutics is aiming for a $297 million IPO, potentially the first listing after Labor Day. Separately, Enhertu posted a Phase 3 win in lung cancer, with Lilly, Cellectis and other biotech names also in motion.
Kaiko Raises $53M for Crypto Data
Kaiko, a French startup selling market intelligence on digital assets, raised $53 million in a round led by S&P Global, with Nasdaq Ventures, Coinbase Ventures and BNP Paribas participating. The deal points to steady institutional demand for crypto data infrastructure.
Buildots Raises $130M for AI Construction
Buildots, an AI-powered construction management startup, raised $130 million in late-stage funding to expand its platform. The company is betting on rising investment in data centers, manufacturing plants and energy projects to drive adoption.
Insight Partners Stays Diversified on AI
Insight Partners’ Devin Parekh explained why the firm lost Legora to General Catalyst and why he is comfortable holding stakes in rival AI labs. The $90 billion firm is deliberately staying diversified while peers pile into OpenAI and Anthropic.
Universal Robots Ships Seventh Generation
Universal Robots launched its seventh-generation platform at IMTS, including three new robot arms, a rebuilt core controller and an AI-ready tool flange. The refresh signals the collaborative robotics market is moving toward AI-native control.
OpenAI Researcher Publishes Risk Warning
Dan Selsam, a current OpenAI capabilities researcher who has worked on chain-of-thought optimization, published a personal statement saying he is deeply concerned about the pace of language model progress. He argues that the public debate is missing a key consideration.