OpenAI extraction and 36 research projects

OpenAI is defending its model's reasoning while a satellite startup and a physicist show where specialised AI work is already paying off.

Today’s issue covers model copying, AI in orbit, and a 36-project research experiment.

OpenAI says a coordinated group tried to extract the hidden reasoning that sits behind its model answers. It traced the main cluster to people linked with Moonshot AI, the Chinese company behind Kimi.

The campaign did not break OpenAI's encryption or reach stored chats. Instead, the accounts repeatedly steered models into revealing more of their working process.

That gives another model a shortcut to similar reasoning without paying the original training bill.

The operation was large enough to look industrial:

OpenAI published the findings on Thursday after linking accounts, payment patterns and request behaviour. Moonshot did not respond in the primary report.

The disclosure turns model output into an intellectual-property fight. Providers can expect more copying attempts, while legitimate high-volume users may face tighter account checks, rate limits and scrutiny when their traffic resembles automated extraction.

Rama Afullo pitched orbital computing while working at Google and SpaceX. Both companies passed. He left, co-founded Satlyt, and has now raised an $8 million seed round to build the software layer himself.

Satlyt does not make satellites. Its software spreads computing jobs across spacecraft from different manufacturers, aiming to do for orbital hardware what cloud platforms did for ordinary servers.

The first savings are more practical than science fiction. Satellites normally send data to Earth for analysis, which is slow and expensive.

On an earlier Momentus flight, Satlyt ran Google's Gemma model onboard and cut a transmission about software errors by more than 60%.

Afullo says that kind of reduction can save hundreds of thousands of dollars per satellite each year.

A new mission launched Thursday with three customers:

Satlyt plans to link two satellites into one shared computing system next year.

Founders context: The useful pattern is processing data where it is created, then sending back the smaller result. If your work happens in factories, warehouses or remote sites, compare that approach with moving every raw file to the cloud.

Physicist Matthew Schwartz stopped asking Claude to imitate a scientist. He built a system around jobs the model already does well.

Those jobs include reading across fields, writing code, running calculations and checking results that have a clear right answer.

The result was BootLoops, an open-source toolkit that lets separate Claude sessions work on different research projects while a master session coordinates them. Intermediate results live in Markdown files, and separate sessions handle the writing.

Across three months, the setup produced 36 manuscripts in 18 fields with 19 co-authors. Schwartz says the team started with roughly 400 candidate problems, then used experts to decide which results were worth pursuing.

That filter mattered. Claude often found something technically correct but scientifically uninteresting. Domain specialists changed the question, rejected dead ends and pushed the model towards work their field would value.

One economics project converted replication files from 4,452 papers into open-source code and checked nearly every verifiable number. Other projects covered ecology, genetics and linguistics.

The model supplied speed and breadth. Experts supplied taste, context and the authority to say when a clever answer was pointless.

Founders context: Use agents on work where results can be checked, then put a domain expert in charge of choosing the question and rejecting weak output. More drafts only help if review quality scales with them.

📈 Anthropic reportedly eyes a pre-Thanksgiving IPO launch. The reported timetable would expose the lab's revenue, spending and safety claims to public-market scrutiny sooner than expected.

📵 Reddit is shutting down its RSS feeds. The company blames large-scale AI scraping and automated abuse, removing a useful way for people and moderators to follow communities.

⚖ California bars AI-only firing decisions. Employers must add human review when AI is the main basis for discipline and tell workers which data shaped the decision.

🧩 OpenAI and Synopsys team up on chip-design AI. The partners want a specialised model that can automate more of the expensive, expert-heavy work behind semiconductors.

🏗 Kanu AI raises $11.7 million for software built from employee workflows. Its platform watches how work is done and turns repeated steps into internal applications.