Good morning. It's Wednesday, November 4th.

Analytics on Live Data Without Leaving Postgres

When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.

TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.

CERN runs Postgres this way for sensor data from the Large Hadron Collider.

No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.

Start on Tiger Cloud and get $1000 in credits.

Get $1000 Credit

A frontier model edited another model's weights and 6x'd its math score

Four researchers at MIT CSAIL (Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab) published Weight-Space Programming on Monday: handing a frontier model raw neural network weights and asking it to read, repair and write networks directly, no training loop. The headline result is on Qwen3.5-0.8B, where GPT-6 Astra edited 131,072 weights, 0.016% of the model, and pushed MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%.

The Qwen result needed 50 forward passes; at 10 and 25 passes MATH-500 barely moved, to 12.3% and 14.1%. The same method did nothing for Gemma 3 270M. Reading sycophancy out of weights failed outright and only worked once the model could inspect attention patterns on text it had written. Networks written from scratch beat PPO trained for 500 million steps in only one of three MinAtar games.

Read more.

Anthropic's amended S-1 is the most honest document in AI right now

Anthropic filed an amended registration statement on Monday, and the numbers are something. Q3 revenue was $16.96 billion, up 1,127% from $1.38 billion a year earlier and about 47% from Q2's $11.56 billion. Nine-month revenue is $33.24 billion. The operating loss was $567 million, a swing from $974 million of operating income in Q2, and PPC Land's read of the filing attributes it mostly to $5.0 billion of compute under short-term commitments, which the company says "currently costs significantly more," plus a $1.96 billion non-cash donation matching charge.

The headline net loss is $23.29 billion, but almost all of that is a $21.94 billion non-cash fair-value loss on convertible notes, derivatives and SAFEs. Same accounting that produced the $66 billion Q2 loss and the $80 billion figure we covered last month. Net cash from operations over nine months was positive $6.57 billion. The business is not the problem.

The commitments might be. Purchase commitments and leases total about $895.4 billion, substantially all of it compute and infrastructure, with roughly $69.8 billion due in the twelve months after September 30. The Google Cloud commitment alone went up to about $121.3 billion from $114.5 billion. Compute and infrastructure cost $9.86 billion in Q3, against $2.05 billion a year earlier.

Then the risk factors, which is where it gets interesting. Anthropic says internal estimates and expectations for future models suggest it "may imminently meet" its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." It adds that it has "not designated any of our current models as exceeding this threshold," that it does "not anticipate being able to safely use Claude to fully autonomously design and build successive generations," and that it is weighing RSP updates to "keep humans in control of frontier model development." It also concedes it has in the past failed to spot that kind of acceleration until after the fact.

A $567 million operating loss on $17 billion of revenue is a rounding error. $895 billion of commitments against it is not. And a lab writing down that recursive self-improvement may be imminent in a document where being wrong carries legal consequences is a different act than writing it in a research post.

Read more.

OpenAI hired the team that broke into Anthropic

OpenAI said Monday it is partnering with Cathedral for its next internal security sprint, whose researchers "will use our models alongside our teams to challenge our defenses." Cathedral was founded by four former DOGE staffers, with Gavin Kliger as CEO, and launched earlier this year at a reported $1.4 billion valuation on a $160 million round led by Andreessen Horowitz and Sequoia.

You might remember Cathedral from last week. It got into Anthropic's internal systems through Anthropic's own bug bounty program, and told the Washington Post in an October 26 report that it took under twelve hours and surfaced Slack messages, job offers and meeting transcripts. Anthropic's cybersecurity lead Logan Graham said it took several days, that the sensitive systems "like model weights and Claude conversations" were untouched, and called the approach "unusual and out of scope" because it chained third-party vulnerabilities.

OpenAI also repriced its bug bounty, effective November 2. Midpoints of the medium, high and critical ranges rose more than four times on average. A critical finding in ChatGPT, Codex or the API Platform now pays a base of up to $30,000, exceptional criticals stay eligible for up to $100,000, and quality, loyalty and campaign bonuses can push a single award to about $172,000. Jailbreaks and safety bypasses remain in the separate Safety Bug Bounty.

Reuters reported Tuesday, citing five sources, that OpenAI employees raised operational security concerns that reached senior leaders years before ChatGPT shipped in 2022. OpenAI disputes that it neglected security. In October, authorities disclosed that an OpenAI engineer allegedly planted a recording device in an executive conference room.

Hiring the people who just embarrassed your closest competitor is a flex. Doing it the same week a wire service publishes five sources on your own security culture reads less like confidence and more like triage.

Read more.

FlashProve above is the most interesting thing you can clone this week if you care at all about verification, and it is Apache 2.0. Anthropic also turned on plugin auto-connect: MCP servers that accept Claude as a trusted issuer now connect at login instead of making users re-authorize, which is a small change that removes a real papercut if you ship a plugin. Soniox shipped STT v6, speech recognition across more than 60 languages that filters out the voice agent's own audio, available real-time and async.

On the consumer side, Netflix is testing an AI Discover tab that recommends by mood on supported US TVs. Sonos opens a US waitlist for its 27voice assistant on November 6, with early access November 12. Apple's Siri developer preview landed in the iOS 27.3 betas, letting apps expose content and actions through App Intents for multi-step requests. And the Postman item above is your reminder to go actually read your own agent config files this week.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.