Edition, Sat, 31 Oct 2026.

NES emulator lives inside a Qwen LLM's weights

Drew Breunig embedded a full NES 6502 CPU emulator into the weights of Qwen3.5-4B, enabling the model to play ROMs like Super Mario Bros by running the emulator at inference time.

Drew Breunig embedded a full NES 6502 CPU emulator into the weights of Qwen3.5-4B, enabling the model to play ROMs like Super Mario Bros by running the emulator at inference time.

The project, called NESformer, draws inspiration from earlier work by Divyansh Agrawal on qwenputer. By encoding the emulator's logic into the model's parameters, the system can process NES game ROMs and produce playable output through token prediction. The approach is notable as an unusual intersection of retro gaming and modern LLM architecture, though the system runs very slowly compared to traditional emulators. Simon Willison documented the project with working demos and benchmarks.

Google's AI satellite beams back its first response from orbit

Google successfully booted Tensor Processing Units on a chip-equipped satellite, running a Gemma 4 model that generated a haiku about space, marking a milestone in orbital AI computing.

Google successfully booted Tensor Processing Units on a chip-equipped satellite, running a Gemma 4 model that generated a haiku about space, marking a milestone in orbital AI computing.

The satellite represents Google's exploration of deploying AI inference in space, where computational resources could be used for on-orbit processing rather than transmitting raw data back to Earth. Google announced the successful boot of its TPUs and the model's ability to process prompts via a blog post on Friday. The demonstration opens possibilities for future AI-assisted satellite operations and space-based data processing, though the practical applications beyond this proof-of-concept remain undefined.

Anthropic's IPO path gets bumpier as junk rating looms

Blackstone forecasts Anthropic will receive a high junk credit rating from major agencies following its IPO, potentially complicating the company's ability to raise affordable capital for its massive AI infrastructure spending.

Blackstone forecasts Anthropic will receive a high junk credit rating from major agencies following its IPO, potentially complicating the company's ability to raise affordable capital for its massive AI infrastructure spending.

The junk rating concern, first reported by the Financial Times, comes as Anthropic files for its Nasdaq listing amid soaring AI compute costs. The company has been funding aggressive expansion through venture capital, but a low investment-grade rating would make future borrowing more expensive and difficult. CNBC reported that broader concerns from hyperscalers about AI spending projections have already punished tech stocks, pushing the Nasdaq into its first three-week losing streak since March. Meanwhile, Anthropic's S-1 filing emphasizes life sciences and drug discovery as central pillars of its business strategy.

Apple Intelligence launches in China with iOS 27.3

Apple announced it will bring Apple Intelligence features to China starting with the iOS 27.3 beta on November 4, fulfilling a long-pending commitment to the world's largest smartphone market.

Apple announced it will bring Apple Intelligence features to China starting with the iOS 27.3 beta on November 4, fulfilling a long-pending commitment to the world's largest smartphone market.

The rollout comes after years of regulatory and geopolitical complications that kept Apple's AI features unavailable in mainland China. The announcement was made on Apple's Chinese newsroom website, as first reported by MacRumors. The full iOS 27.3 release will follow the beta, marking Apple's first major push into Chinese AI features. The move positions Apple to compete with domestic AI capabilities as Chinese regulators increasingly scrutinize foreign technology companies' data practices.

OpenAI, Google, Microsoft and Anthropic back new scraping opt-out standard

The four leading AI companies agreed to adopt a new standard that lets publishers opt out of having their content scraped for AI training, marking a significant shift in the industry's approach to web scraping.

The four leading AI companies agreed to adopt a new standard that lets publishers opt out of having their content scraped for AI training, marking a significant shift in the industry's approach to web scraping.

The new standard, first reported by Search Engine Roundtable, gives publishers a unified mechanism to control whether their content can be used across multiple AI companies' training pipelines. OpenAI, Google, Microsoft, and Anthropic have all committed to adopting the framework, which addresses growing concerns from publishers and media companies about unauthorized scraping of their copyrighted material. The move represents one of the most concrete steps yet toward a structured opt-out regime in AI training data, though implementation details and enforcement mechanisms remain to be clarified.

ServiceNow posts its worst day ever after a Q3 beat-and-raise

A revenue beat, raised AI target, and lifted guidance couldn't stop a record 27% slide as a slower cRPO outlook spooked enterprise software.

A revenue beat, raised AI target, and lifted guidance couldn't stop a record 27% slide as a slower cRPO outlook spooked enterprise software.

ServiceNow reported third-quarter 2026 revenue of $4.11 billion, up 20.5% and slightly ahead of consensus, with non-GAAP EPS of $1.00 just under the $1.02 estimate. The company nudged its full-year subscription revenue outlook to $15.78–$15.79 billion, lifted its free-cash-flow margin target to 36%, and raised its 2026 AI annual contract value target from $1.5 billion to $1.7 billion, with agentic AI consumption in production up more than 13x year over year. Fourth-quarter subscription revenue is guided to $4.245–$4.255 billion.

The problem was bookings: ServiceNow guided fourth-quarter cRPO growth of 18.5% in constant currency, below the roughly 19.8% Wall Street expected and down from 21% in Q3. CFO Gina Mastantuono attributed the gap to early renewals pulling revenue forward, public-sector mix, and about 50 basis points of bundled AI delivery work now booked as professional services. Shares fell 27% Thursday for their worst day on record, dragging the sector down with them (Salesforce -6%, Workday -5%, Adobe -3%), even as Citi's Tyler Radke called the lack of upside "a little bit uncharacteristic for ServiceNow standards."

ServiceNow CEO calls the shrinking workforce "a gorgeous situation"

McDermott says the company's own AI tools have saved more than $1 billion a year, letting it hire fewer people.

McDermott says the company's own AI tools have saved more than $1 billion a year, letting it hire fewer people.

On the Q3 earnings call, chairman and CEO Bill McDermott said ServiceNow's "ultimate headcount has actually slightly declined, which is a gorgeous situation," describing the outcome as "less people doing extraordinary things, enjoying their work much more." CFO Gina Mastantuono said the company's internal use of its AI products generated more than $1 billion in annualized efficiency gains, "and that means we're able to not hire as many people."

ServiceNow said it is still hiring where it wants to grow: quota-carrying salespeople are up by double digits while sales operations headcount is "down significantly." Few large software companies have tied AI so directly to a smaller payroll on the record, and ServiceNow sells the same technology to its customers, making its internal results part of the pitch.

ServiceNow deepens OpenAI tie and launches the Control Tower Alliance

A ChatGPT plugin, Codex integration, and an 11-member AI governance alliance landed the same day as the stock's worst day ever.

A ChatGPT plugin, Codex integration, and an 11-member AI governance alliance landed the same day as the stock's worst day ever.

ServiceNow announced an expanded partnership with OpenAI, which the company frames as moving "from AI reasoning to governed action": a new ServiceNow plugin for ChatGPT (enabled by Action Fabric) that lets customers use the platform directly from within ChatGPT, Security and Risk solutions powered by OpenAI cyber models, and a Codex integration in the AI Workflow Factory. No exclusivity, duration, or revenue sharing was disclosed.

The company also launched the Control Tower Alliance, an ecosystem for adopting trusted enterprise AI at scale, with founding members Anthropic, AWS, Cisco, HiddenLayer, Kong, NVIDIA, Okta, OpenAI, Stripe, WitnessAI, and Zenity, alongside new AI governance offerings such as AI Passport, which verifies AI agent scopes in real time, and the personal agent ServiceNow Arc. Bulls read the deal as proof ServiceNow remains the execution layer for AI-driven work; bears note nothing is exclusive and the alliance even includes rival Anthropic.

Serval raises $150 million at a $3 billion valuation to take on ServiceNow

Sequoia leads a third round in 12 months for the AI-native ESM startup, which says one automaker moved off ServiceNow onto it in 45 days.

Sequoia leads a third round in 12 months for the AI-native ESM startup, which says one automaker moved off ServiceNow onto it in 45 days.

Serval, the San Francisco-based AI-native enterprise service management platform, raised a $150 million Series C led by Sequoia at a $3 billion valuation, bringing total funding to $277 million. Conviction and Headline joined existing investors including Elad Gil, Evantic, First Round, General Catalyst, and Redpoint. The company says it has doubled ARR each quarter since its $75 million Series B in December 2025, and has onboarded two Fortune 20 customers for organization-wide deployment, with other clients including Fox Corporation, WPP, Live Nation, Ramp, Notion, and Lucid Motors.

Serval positions itself as a rip-and-replace option for legacy ESM platforms: one auto manufacturer migrated from ServiceNow to Serval in 45 days, covering IT service management, CMDB, hardware and software asset management, and cross-department workflow automation. The company is launching ProServal, an arm that will embed 100+ "Automation Engineers" inside large organizations, and plans to add 500+ people and open offices in New York, London, and Austin.