Good morning. It’s Friday, November 6, and we’re covering Anthropic’s accelerating research pace, a new loophole for NVIDIA chips, and OpenAI’s predictions for major breakthroughs in healthcare.

Plus: yesterday’s poll results on Grok Bot usage, and a Friday Fact on the humble beginnings of the deep learning revolution.

→ Make Enterprise Content More Useful at Every Step

OpenAI president Greg Brockman predicted that AI will produce major advances in biology, health and cures within 12 to 18 months. He also suggested that technology could eventually make death a choice, a claim that goes far beyond the evidence available today.

Claude Can Analyze Your Spending

Anthropic is rolling out Personal Finance in Claude to U.S. Pro, Max and Ultra subscribers. The beta connects bank and card accounts through Plaid, analyzes up to two years of read-only transaction data and can build charts, calendars and dashboards. Claude cannot move money, manage accounts or make investments.

Moonshot AI, the Chinese company behind the Kimi assistant, has explored renting European data centers with clusters of more than 8,000 NVIDIA Blackwell chips, CNBC reports. The capacity would be used mainly to run models and could also support training. Overseas cloud access would give the company a path to advanced chips restricted from direct sale to China.

The Nasdaq 100 plunged more than 6% yesterday as doubts grow over whether AI investments can deliver expected returns. The S&P 500 fell 4%, wiping out $2.5 trillion in market value, amid Democratic midterm victories and concerns over data center restrictions. Qualcomm and Arm dropped 12%, NVIDIA lost 9%, and semiconductor stocks slid 11%.

In Part 1, we introduced the Grokbot playbook. In Part 2, we’re putting it to work: collecting invoices, exploring ways to lower bills, preparing sales follow-ups, and turning tasks you demonstrate into reusable workflows. Get practical prompts and our team’s favorite use cases to try yourself.

Read Part 2 →

The model that set off the deep learning era did not run in a data center. Find out more at the bottom of the newsletter.

New from Box: Reimagining business processes with content and security at the core

→ Make enterprise content more useful at every step with Box

Anthropic appears to be developing AI models roughly twice as quickly as it was a year ago, according to an outside assessment by the Model Evaluation and Threat Research group, or METR. The researchers studied Anthropic from August through October and estimated that AI-assisted research and engineering explain most of the acceleration.

If the estimate holds, advanced models are no longer just products; they are becoming tools that help build their successors. Anthropic says it has not yet seen strong internal bottlenecks that would stop the trend over the next year. → Read the full AP report.

Aina opens preorders for a $399 AI ring: Preorders open at $399 for a ring with screen, joystick and mic; shipping January 2027, $499 standard.

OpenAI data center payments: New pledges to cover infrastructure and power costs, publish energy and water data, skip NDAs with officials.

OpenAI eyes 2027 IPO readiness: Greg Brockman says the company could be IPO-ready in 2027, citing enterprise growth and business confidence.

Apple cuts internal AI budgets: Reportedly cut daily AI spend for many engineers from $300 to $200 in October, with some teams at $100 or less.

$1 million for unsolved materials problems: Startup offers $1 million for five open problems, including superconductivity and quantum chemistry.

In 2012, Alex Krizhevsky trained AlexNet on two NVIDIA GTX 580 graphics cards in his bedroom at his parents' house. The network was too large for one 3GB card, so he split it across both. Training ran for 90 epochs over five to six days. That October, the model won the ImageNet competition with a top-5 error rate of 15.3%, more than ten percentage points ahead of the runner-up.

Until then, most image recognition relied on hand-engineered features. AlexNet showed that a deep neural network trained on consumer gaming hardware could beat them by a wide margin. Researchers across the field switched direction within a few years. In 2025, the Computer History Museum, working with Google, released the original source code. The paper describing it has been cited more than 172,000 times.

Here's how you voted: A majority (68%) have never tried Grok Bot, while 15% use it daily and 9% have tried it but stopped.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team