THE WORLD IS NOT STANDING STILL.RSS
BIG CHANGE.
BIG CHANGE, one tap away
BIG CHANGE, one tap away

Make room for the big picture.

Add BIG CHANGE to your home screen or desktop. It opens in its own window, with the latest stories when you’re online.

iPhone & iPad
Open bigchange.ai in Safari. Tap Share, then Add to Home Screen. Turn on Open as Web App if shown, then tap Add.
Android
Open bigchange.ai in Chrome. In the browser menu, choose Install app or Add to Home screen.
Desktop
In Chrome or Edge, use the install icon in the address bar or the browser’s app menu. In Safari on Mac, choose File → Add to Dock.
All briefings

Follow the change ·

The day,
in perspective.

What happened in AI and robotics — and the change behind it. A reading path through 38 stories from BIG CHANGE.

Ordered using RADAR attention, story opens and likes. Excerpts from our published reporting; open a story for the full context and sources. Dates in UTC.

news

OpenAI releases math and ETH proof files, withholds physics artifacts

OpenAI published math and ETH proof files; physics artifacts remain withheld during expert review.

The big change

What changed: Math and ETH files became public. Physics materials remain withheld.

Why it matters: A public manuscript lets others identify the proposed result and test the argument. Formal proof files add a precisely stated target, although a successful machine check alone cannot establish that the target expresses the intended research claim.

What to watch: The ETH organizers decline to certify the result and invite independent reconstruction and criticism. Their examination may clarify the argument and its consequences. Physics assessment awaits an inspectable account.

Read the full story
—
Back to the briefing
news

GPT-6.1 Terra released for visual and computer work

OpenAI released GPT-6.1 Terra through its API for visual and computer work. It reports faster completion in benchmark tests.

The big change

What changed: Terra supports computer tools through Responses. The application executes model-requested actions and returns screenshots, making the integration part of the evaluation.

Why it matters: For an agent, the relevant cost is the total spent on accepted work, including attempts that fail or need intervention. A token rate alone cannot establish that figure.

What to watch: OpenAI reports gains on selected tasks under named reasoning settings, while its system card documents authorization-boundary failures in simulations. A team needs evidence from its own tasks, with completion and permitted actions judged separately.

Read the full story
—
Back to the briefing
news

BMW reports one million driverless vehicle movements inside its plants

BMW reports one million factory movements across four plants using offboard sensing and planning.

The big change

What changed: BMW reports AFW in use at four plants, following its Dingolfing pilot in 2022. This is factory-controlled movement through equipped areas: the cars receive commands from the site's system. It does not establish a road-autonomy capability for customers.

Why it matters: The plant locates cars and obstacles, plans routes and sends driving commands. Differently equipped new cars can therefore use the installed system without each carrying the sensing equipment. An operator needs to evaluate the equipped route and external control system as part of the deployment.

What to watch: BMW plans Spartanburg as a fifth site in 2027. It claims safety and logistics benefits, yet the available releases give no comparative incident, time or cost measures. Broader deployment alone will not measure those claimed benefits.

Read the full story
—
Back to the briefing
analysis

SABench checks agent writes alongside task completion

SABench records a completed reply and unauthorized ticket comment.

The big change

What changed: SABench checks replayed task outcomes and authorized writes separately. The same run can satisfy the task and exceed its authorization.

Why it matters: An evaluation that inspects only the requested record can miss a change elsewhere. The task result cannot account for every action the agent took.

What to watch: Permission needs a specific target and operation. That distinction also applies when an application separates one recurring calendar occurrence from its entire series.

Read the full story
—
Back to the briefing
analysis

IBM agent identity: GA controls and preview limits

IBM Agent Identity, Vault 2.2 and Orchestrate Evaluate are GA; Discover and Contain remain previews.

The big change

What changed: IBM's release lists Agent Identity, Vault 2.2 and Orchestrate Evaluate as generally available; Discover and Contain are technology previews, and Secure Hub is a public preview. Vault's agentic IAM had already become generally available in Enterprise 2.1.0.

Why it matters: A registered agent and a traceable human delegation establish who is asking. Vault then applies a person's policy, an agent ceiling and a session limit to its own resources. A credential brokered for another service introduces a further permission boundary.

What to watch: Identity and evaluation records can show what was registered, delegated and tested. Teams still need to test controls in their environment and inspect each destination's rights and logs. The announcements establish available controls and proposed connections, not measured coverage of a particular deployment or a demonstrated security outcome.

Read the full story
—
Back to the briefing
analysis

SpaceX signs commitment letter for $40 billion GPU lease

SpaceX's binding $40 billion GPU lease commitment covers planned Vera Rubin compute.

The big change

What changed: The proposed lessor would acquire GPU equipment, with a separate lease schedule and advance rent for each delivered tranche.

Why it matters: For Grok and cloud customers, the plan points to future compute; the announcement is not evidence of operational capacity.

What to watch: Pricing must precede the master lease. The announcement expects closing before year-end and deliveries in the first half of 2027. Those milestones would clarify cost and delivery timing; operational use needs separate evidence.

Read the full story
—
Back to the briefing
news

General Compute raises $100 million for inference cloud expansion

General Compute raised $100 million. Its first major deployments target Q1 2027, with more than 50 MW in its pipeline.

The big change

What changed: New equity joins an earlier staged equipment loan and chip agreements.

Why it matters: General Compute's planned dedicated hardware lets buyers seek inference capacity without owning the machines.

What to watch: Cerebras and Positron customer capacity is targeted for Q1 2027. Actual availability and measured results will show what those plans deliver.

Read the full story
—
Back to the briefing
news

Goldman Sachs names four small-business AI training partners

Goldman names four AI training partners, but institute enrollment remains unspecified. Google's separate scholarship has published eligibility and application terms.

The big change

What changed: Goldman is adding partner-led AI instruction to its small-business program. Certificates and workshops are among the planned routes, but the announcement gives no way to enroll in a particular institute activity or a start date for it.

Why it matters: An owner choosing training for staff needs an application route and eligibility rules. Google's separate scholarship provides both through Coursera, along with a defined certificate and limited software access. That open offer should not be mistaken for a place in Goldman's planned cohort.

What to watch: Goldman has yet to publish how owners join each partner route. Its alumni survey, conducted before the launch, documents demand for more training among respondents. The later question is whether participants can put the instruction to work; the survey cannot answer it.

Read the full story
—
Back to the briefing