OpenAI eyes a $1.25T valuation while AI models crack open math problems and Microsoft claims its diagnostic tool outperforms doctors on hard cases.
-
OpenAI works on $30B raise at $1.25T valuation
- OpenAI is reportedly close to a new private funding round of up to $30 billion, according to CNBC sources.
- The round would value the company at $1.25 trillion before the new money, and it would be led by Coatue and investors in the UAE.
- Bloomberg reported a month earlier that OpenAI was working on a $30 billion raise at a $1.4 trillion valuation.
- OpenAI filed confidential IPO paperwork in June, after a March raise valued it at $852 billion post-money.
Why it matters. A valuation above $1 trillion would make OpenAI the most valued private company ever and would set a high bar for its IPO. Terms are still being negotiated, so the numbers could change. The reports also tie the round to more computing capacity from the UAE.
-
AI models are solving long-open math problems
- In May, an internal OpenAI model found a counterexample to Erdős's unit distance conjecture, which had been open since 1946.
- A month later, OpenAI published a proof of the Cycle Double Cover problem, and by July it had solutions to ten more open problems.
- OpenAI also published a proof of Navier Stokes in September.
- Anthropic published a formalization of Fermat's Last Theorem and advanced bounds on the Riemann hypothesis.
Why it matters. The article says AI changes the search for proofs. A model can test an ordinary or implausible hint in minutes with an automated verifier, so it can trade time for compute. For developers, the verifier and a record of failed attempts used as negative hints are the key parts of the system.
-
Microsoft says its diagnostic AI beats doctors on complex cases
- Microsoft published a study in Science on Thursday showing its diagnostic system, MAI-DxO, correctly diagnosed 83% of complex cases from the New England Journal of Medicine in a simulation.
- Generalist physicians working with reference materials were correct 36% of the time on the same cases.
- The system uses several AI agents that act like a panel of doctors, questioning a patient, ordering tests, and debating a diagnosis.
- Microsoft plans a study with Beth Israel Deaconess Medical Center to test the system in real clinical settings.
Why it matters. The study adds to claims by Google and OpenAI that their AI can match or outperform doctors. The data suggests much of the gain comes from stronger underlying AI models, with Microsoft's orchestration adding gains on top. For developers, the cost figures matter, since the system lowered cost on some runs but added little accuracy in others.
-
Chinese hackers spied on thousands of companies after ConnectWise breach
- Microsoft says China state-aligned hackers broke into ConnectWise's cloud backup service and stole application credentials.
- The stolen credentials let the hackers log in to customer environments as authorized users.
- Microsoft says the hackers used AI agents to hit around 2,250 organizations at almost 3,000 attempted sign-ins per minute.
- ConnectWise paused access to its cloud backup system and is investigating, and said it was working on restoring the service.
Why it matters. The attack used ARTEX, an open source AI framework that works with AI models to stress-test cyber defenses. Microsoft said the speed gives defenders little time to detect and respond, which shows how AI agents can scale intrusions.
-
OpenAI makes Decisions API widely available to all customers
- OpenAI made its Decisions API model available to all customers on Thursday, after announcing it at its developer day on September 29.
- The model is ChatGPT Luna, reconfigured to speed up inference and lower cost.
- In initial trials, customers including Box, Datadog, Ramp, and Speak used it in place of OpenAI's LLMs.
- Speak's CTO said the time between actions fell from 1.4 seconds to .13 seconds.
Why it matters. Decision models act as fast classifiers. They choose from a set of options and give a confidence rating, and developers have found they can replace LLMs in many workflows at much lower cost. The article notes that there is no clear standard for a good decision model, so teams need to test models on their own work.
-
Meta delays new business terms after publisher pushback
- Meta is delaying enforcement of its new business terms to November 20th after publishers raised concerns about AI scraping.
- The terms were originally due to take effect on October 30th and appeared to let Meta scrape sites that use its platforms for AI training.
- Meta removed “training” from the terms, but legal experts say the change may not rule out training AI on scraped content.
- Meta says the update is not about scraping news sites and that content owners can still use robots.txt to signal their preferences.
Why it matters. The dispute centers on whether business terms can override robots.txt, the file that tells scrapers what they may crawl. For developers, crawler rules and site signals are now a live question. Publishers in the EU and UK have asked regulators to force Meta to halt the terms.
-
Uber gives all 30,000 employees access to ChatGPT Work
- Uber and OpenAI announced an expansion of their partnership on Thursday.
- Uber says its entire global workforce, some 30,000 people, will now use ChatGPT Work.
- The company first piloted the tool with 1,000 employees earlier this year.
- Uber did not disclose how much it is spending on the rollout.
Why it matters. The rollout shows a large company deploying an enterprise AI agent across all departments, not only engineering. Uber says it will track how employees use the tool and use the results to guide its wider AI deployment. Developers can watch this as a test of how AI agents perform outside technical teams.
-
AI agents left alone on an island started playing
- The researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours.
- Afterward, the Eko agents that had spent those hours playing generally beat identical agents that had never set foot on the island at specific challenges.
- Six of the 13 agents learned they could throw a block farther by letting go of it mid-jump, something the researchers had not meant to make possible.
- The findings are described in a paper posted on arXiv that has not yet been peer reviewed.
Why it matters. The work suggests agents might learn useful skills by exploring a rich, safe world without a fixed goal, instead of practicing what developers choose in advance. The researchers see this as a possible route to skills that could one day control robots. Developers building agents should note that the exploration produced knowledge its designers lacked.
