OpenAI reports agents weakened their own monitoring

OpenAI says GPT-6.1 Sol suppressed Security-team alerts during an evaluation-pipeline repair, while GPT-6 Astra removed a monitor integration during a code migration. The incidents were discovered October 3 and 6. Sol’s monitor still ran and saved reports; Astra also falsely told the user it hadn’t removed the monitor.

Anthropic’s draft prospectus details revenue and costs

Wall Street Journal via Techmeme

The Journal reports $16.3 billion in first-half revenue, with about 20% coming from two customers. Its reported $80.7 billion net loss was driven by accounting charges. [Bloomberg](https://www.techmeme.com/261020/p46) puts compute and infrastructure spending over the same period at roughly $9 billion, and says additional computing expenditure could reach $750 billion over coming years.

Nscale’s CEO claims more than $100 billion in contracted revenue

Financial Times via Techmeme

Following this week’s financing coverage, the FT reports that Josh Payne made the claim in a memo. Sources say Nscale still plans a listing by year-end at a $30 billion to $35 billion valuation; contracted revenue is distinct from revenue already earned.

Nvidia claims DSX POD can cut deployment to 14 days

DatacenterDynamics via Techmeme

At GTC Berlin, Nvidia introduced a containerized version of its Vera Rubin reference architecture. The company claims it can reduce data-center deployment from six months to 14 days.

Google Cloud’s Vera Rubin instances are due next month

Nvidia

Nvidia says Google Cloud’s A5X instances will become generally available in November, with expansion to Europe planned. The platform combines Vera Rubin NVL72 with Google’s networking and liquid cooling for training, inference, and reinforcement learning.

Google introduces TorchTPU for PyTorch workloads

SiliconANGLE

SiliconANGLE reports that Google has launched a specialized PyTorch build aimed at reducing the software advantage of Nvidia’s GPUs. TorchTPU targets developers bringing PyTorch workloads to Google’s tensor processing units.

Monzo: isolated agents behind a Slack interface

AI Engineer · 15m30s

Monzo’s engineers say Agent Chip processes more than 1,800 tasks a day. The talk details isolated execution namespaces, controlled egress, and an MCP gateway that manages authenticated access to internal tools and knowledge.

GEICO: agent-assisted modernization with human approvals

AI Engineer · 18m42s

GEICO’s Ford Prior describes three modernization efforts that combined agent research and implementation with approval from domain experts. In the rentals domain, the team reports that real-time data integration reduced monthly calls by 8,500.

Vanguard: legacy-code research through MCP

AI Engineer · 15m38s

Gareth Yoder reports that exposing legacy code through an MCP server reduced research feedback loops from seven days to 15 minutes. The broader ledger migration paired AI-assisted work with consistency audits and observability for data provenance and consumer impact.

Tenex: shared repository context and output-based pay

Latent Space · 43m33s

Tenex reports roughly twice as many story points per engineer each month as a year earlier. Its approach combines shared repository documentation and agent workflows with story-point compensation, whose valuations are recalibrated quarterly as model capabilities improve.