The Researcher Who Warned AI Could Go Catastrophically Wrong Just Joined OpenAI's Safety Board

Thursday’s AI news was mostly about who gets to check the machine’s work, and how much money is riding on it staying profitable enough to keep building. OpenAI put one of its most credentialed outside skeptics in charge of judging its own safety process. Oracle posted a quarter that startled its own analysts. Anthropic tried to put real numbers on what the next four years of AI could do to the American job market, in its best case and its worst. In India, two government programs that have been in the works for years both got more concrete: a ship built to map parts of the ocean nobody has properly surveyed, and a reaffirmed date for a homegrown space station.

The researcher who warned AI could go badly wrong now judges OpenAI’s safety work

On September 9, OpenAI announced that Paul Christiano is joining the OpenAI Foundation’s board and sitting on its Safety and Security Committee, the body that reviews whether new models are safe enough to release, chaired by Zico Kolter. Christiano is not a company insider brought up through the ranks. He worked at OpenAI from 2017, where he co-authored the paper that helped establish reinforcement learning from human feedback, the technique that trains a model to prefer the answers people rank as better. He left in 2021 to found the Alignment Research Center, a nonprofit that studies how a powerful AI system could behave in ways its creators never intended, and has spent years publishing probability estimates for the kind of catastrophic outcomes most AI labs would rather not put a number on.

Christiano also holds a government post, as a senior technical advisor at the Center for AI Standards and Innovation, part of the National Institute of Standards and Technology. To avoid a conflict, he will recuse himself from OpenAI-related NIST matters and from OpenAI’s own model evaluations, and will serve as a non-voting observer on the board of OpenAI Group PBC rather than a full member. Putting a researcher known for taking AI risk seriously inside the room where safety decisions get made is a bet that outside scrutiny beats an internal rubber stamp. It does not change the fact that OpenAI, not Christiano, still decides what to do with his advice.

Oracle’s cloud business grew 121 percent in three months, and it still isn’t enough capacity

Oracle reported its first quarter fiscal 2027 results on September 10, and the numbers were large even by AI-boom standards. Total revenue rose 30 percent year over year to $19.3 billion. Cloud infrastructure revenue, the part of the business that rents out computing power for AI workloads, more than doubled, up 121 percent to $7.4 billion. Total cloud revenue climbed 62 percent to $11.6 billion. Net income rose 60 percent to $4.68 billion, and Oracle’s backlog of signed but not yet delivered contracts, known as remaining performance obligations, reached $664 billion, including more than $30 billion in new AI cloud deals signed in this quarter alone.

The catch is that Oracle is spending faster than that revenue arrives. Free cash flow for the quarter was actually negative $5 billion, because the company keeps pouring money into new data centers and GPUs to keep up with demand. Oracle raised its full-year guidance anyway, to at least $90 billion in revenue for fiscal 2027. That is the AI infrastructure race in one number: a company running a cash deficit on purpose because turning away business looks like the bigger risk.

Anthropic modeled what AI could do to the US economy by 2030. Its own numbers show a rough tradeoff.

On September 9, Anthropic released the Economic Scenario Explorer, an interactive tool that lets anyone adjust assumptions about how fast AI improves and spreads through the economy, then see the modeled effect on GDP, wages and unemployment by 2030. The tool lays out three named scenarios. In the modest case, where AI ends up mattering about as much as the internet did, US GDP reaches $34.1 trillion, a gain of roughly 1.6 percent over what the economy would otherwise be expected to produce. In the substantial case, GDP reaches $36.3 trillion, growth runs at roughly double its normal pace, unemployment settles near 5 percent, and wages for knowledge workers stay roughly flat. In the extreme case, GDP could reach $44.4 trillion, but unemployment among knowledge workers rises sharply too, because AI starts automating cognitive tasks faster than the economy creates new roles for the people who used to do them.

Anthropic calls these modeled scenarios rather than predictions, and built the tool so readers can plug in their own numbers instead of accepting its defaults. What stands out is that a company that sells AI for a living published a scenario where its own product’s biggest success case comes with a real cost to the workers using it, a more honest starting point than most industry economic forecasts manage.

India launched a ship built to survey six kilometers into the ocean floor

On September 9, the Ocean Research Vessel Sagar Manthan was launched at the Rishi Bankim Shipyard in Kolkata, built by the state-run Garden Reach Shipbuilders and Engineers for the National Centre for Polar and Ocean Research under the Ministry of Earth Sciences. Union Minister Jitendra Singh, who attended by video link, named it alongside his wife, Manju Singh. At 89.5 meters long and 18.8 meters wide, with a gross tonnage of 5,900 tonnes and a top speed of 14 knots, it is the first vessel of its kind built in India, designed for extreme conditions including the Southern Ocean, with a planned service life of 30 years. It cost roughly 840 crore rupees, close to $90 million at current exchange rates.

The ship is built to run multibeam and seismic surveys to depths of six kilometers, collect water and seabed samples, deploy moorings and data buoys, and take ocean-current and weather readings, as part of the government’s Deep Ocean Mission. Singh framed the launch as part of a push for self-reliance in maritime science, arguing that data from vessels like this one will let India track melting polar glaciers and climate shifts using its own instruments rather than relying on other countries’ research fleets. The ship is not finished yet. Delivery and commissioning are scheduled for 2028.

Modi told a Paris summit India wants its own space station by 2035 and an astronaut on the moon by 2040

Speaking to the International Space Summit in Paris on September 10, Prime Minister Narendra Modi laid out India’s timeline for the Bharatiya Antariksh Station, its planned space station: a first module by 2028 and a fully operational, five-module station by 2035. He also reaffirmed a target of putting an Indian astronaut on the moon by 2040. Neither figure is new. Both match the roadmap the Indian government approved earlier, so the Paris remarks were less an announcement than a public reaffirmation of a plan already in motion, made to an international audience.

Modi pointed to two recent results as evidence the timeline is realistic: the 2023 Chandrayaan-3 mission, which made India the first country to land a spacecraft near the lunar south pole, and the Aditya-L1 solar mission, still returning data on the sun. He also said more than 430 satellites from 34 countries have launched aboard Indian rockets, a number that reflects how much of India’s space program now runs as a commercial launch business alongside its research missions. He closed by calling for space to stay governed by international law and open cooperation rather than becoming another arena for unilateral moves, a pointed line at a summit with several spacefaring rivals in the room.

Every story today is really about the same tension: how much you can trust a system you cannot fully see inside, and what it costs to build the instruments that let you see further. OpenAI is betting an outside skeptic makes its models safer than an internal committee would. Oracle is betting cash flow it does not currently have on GPU capacity it can barely build fast enough. Anthropic is trying to model a future its own product might upend, honestly enough to include the version where that goes badly for workers. India, meanwhile, is doing something more literal: building an actual ship to see six kilometers down, and holding to a real date for getting a station into orbit, because sometimes the only way to know more is to go look.