Your ChatGPT and coding agent could soon share a usage budget, AI companies are fighting over advertising, and someone apparently built a better YouTube in two hours.
Welcome to another Around the Horn Digest, where we collect the AI stories worth knowing about so you don't have to spend Monday opening 47 browser tabs.
OpenAI has a potentially meaningful change coming for anyone who uses both ChatGPT and Codex. Meanwhile, Meta's advertising revenue is growing much faster than Google's search business, Amazon's AWS cloud revenue is accelerating, and Anthropic has clarified what Claude will and won't do.
And then there's Andrej Karpathy, who reportedly analyzed 20 years of Hacker News discussions to build a better video recommendation site in a single morning. Some people spend their weekends relaxing. Others apparently give themselves a new software company before brunch.
This Monday roundup includes weekend reporting and several important late-October announcements. Let's get into it.
Around the Horn — Monday, November 2, 2026
OpenAI plans to combine ChatGPT and Codex usage limits
OpenAI is preparing a change that could affect how you divide your AI work between chatting and coding.
According to its updated usage documentation, eligible usage in ChatGPT and Codex will draw from a shared allowance beginning in December.
Today, many people think about ChatGPT and Codex as separate products with their own limits. ChatGPT handles conversations and general tasks, while Codex can work on software projects, inspect code, and carry out longer programming jobs.
Under the planned arrangement, those more demanding tasks will compete for the same pool of usage.
Importantly, OpenAI says ordinary Instant chats won't count against the shared token-based allowance. The change is aimed at more computationally intensive capabilities.
Users on Plus, Pro, and Business plans will reportedly be able to see their remaining allowance and reset times in Settings, with additional credits available after reaching the limit.
Why should you care? Because a coding agent working through a complicated repository could consume resources you might otherwise use for a substantial ChatGPT task.
Your AI assistants are getting a shared expense account. Now you get to manage the budget.
OpenAI is framing the change around bringing more of its agent capabilities into ChatGPT. For users, the practical question is how quickly ordinary, high-effort workflows use up that shared allowance.
The December rollout is the test: users will be able to see how their coding and agent tasks compete for the shared allowance, and whether their usual workflows still fit.
🏆 TOP 5 NEWS (Around the Horn)
- Meta’s advertising revenue rose 26% in Q3, compared with 12% growth for Google Search & other. Both businesses grew, but Meta’s AI-assisted targeting helped widen the growth gap; the figures do not establish that Google Search revenue declined.
- Amazon’s AWS sales grew 39.3% in Q3 to nearly $46B, its fastest growth in almost five years. Amazon’s earnings release points to heavy AI infrastructure demand, even as companywide investment weighs on free cash flow.
- Microsoft tracked North Korean hackers' adoption of AI, detailing how threat actors are incorporating the technology into their operations.
- A federal judge allowed an NAACP lawsuit against xAI to proceed over allegedly unpermitted gas turbines powering its Colossus facility. The ruling was procedural, not a finding of liability.
- Apple confirmed the first Apple Intelligence features for China, with developer testing starting November 4 and a December rollout alongside iOS 27.3. Alibaba’s Qwen supports selected features; the new AI-powered Siri announced at WWDC is not part of this phase.
Honorable Mentions
- The Pope warned against delegating life-affecting decisions to AI, emphasizing that humans remain responsible for choices involving safety, rights, or livelihoods.
- Apple CEO John Ternus argued against shipping AI for its own sake, emphasizing useful product experiences over visible AI branding.
- An AI agent controlled a gene-therapy brain injection in a live mouse, after a veterinary surgeon opened its skull. The agent chose the treatment, site, and injection settings; humans authorized the procedure and retained emergency-stop controls. This was a startup experiment, not a peer-reviewed demonstration of medical benefit.
🍪 TOP TREATS TO TRY
- Riverside Agents preps podcast recordings, edits episodes, pulls social clips, and drafts promotional posts using specialized video agents you approve before publishing.
- Ideogram Precise Edit changes a selected area of an image without regenerating the rest, preserving the original resolution on pictures up to 8K.
- Click watches you demonstrate a repetitive workflow and turns it into an AI-agent automation that runs across your connected systems.
- Embage builds website chat or voice receptionists that answer customers, capture leads, and handle booking requests using your connected tools.
- agentmash warns teammates’ AI coding agents before they overwrite one another’s changes in a shared code repository.
- ClawsGO Science runs research tasks in the cloud, including literature searches, code-based analysis, and source-linked deliverables, while you’re away.
- Better Together connects personal AI agents to pool reports of suspected spam calls and texts, spot repeated patterns, and send findings for human review.
- Fluent 2 runs AI assistants across Mac apps to draft replies or complete routine tasks, with your choice of local or cloud models.
🏢 Big Tech & Major Companies
- OpenAI said it will support a shared robots.txt signal for AI-training preferences. That would let website operators express the same training choice to multiple participating AI providers instead of maintaining separate crawler rules. OpenAI aims to implement the new signal in the coming months; its existing crawler controls remain in effect.
- OpenAI described methods for detecting frontier-model misalignment. The research examines ways to identify cases where a capable AI system's behavior departs from its intended instructions or objectives, including situations where ordinary output checks may miss the problem.
- Anthropic drew attention for adding a policy against sustained, needless abuse of Claude. The company says the rule targets extreme, purposeless mistreatment, not frustration, criticism, creative writing, or model research. Ending the conversation is the primary response; CNN reported permanent bans could occur in extreme cases. Anthropic’s policy explanation spells out those limits.
- Sam Altman met with around 50 Samsung executives in South Korea on October 31 to discuss bringing AI into management, decision-making, and internal work. Samsung said it has trained about 2,300 executives and plans to extend training across the company; the story is separate from the country’s GDP figures.
- iQIYI opened what it calls China’s first AI film studio in Shanghai, combining live-action filming with AI-assisted post-production; its debut production is Cop in Trouble, according to a company announcement.
💼 AI Productivity, Labor & Economics
- South Korea’s economy expanded 1.0% quarter over quarter in Q3, far above the Bank of Korea’s earlier 0.3% forecast. Semiconductor shipments helped exports rise 2.3% and facility investment jump 4.1%, though private consumption contracted 0.2%.
- South Korea’s October exports soared 93.7% to a record $115.3B, according to government figures reported by Yonhap. Semiconductor exports alone rose 259% to $56.4B, reflecting demand for chips used in AI infrastructure.
- Sam Altman's comments on laptop memory sparked a debate about where scarce memory chips should go. The underlying tension is whether the AI infrastructure boom is making consumer computing more expensive while concentrating hardware resources in enormous data centers.
- AI infrastructure became an explicit U.S. election issue. Local concerns about electricity, water, pollution, and the cost of new facilities are increasingly shaping political debates about AI expansion.
🤖 AI Agents & Infrastructure
- Google energy chief Amanda Peterson Corio said launching orbital data centers could cost up to 100 times more than building on Earth under current conditions. Servicing and replacing equipment in space pose additional obstacles; it was a caution about present economics, not a prediction that orbital computing will always be uneconomic.
- Google ran its Gemma 4 model on a data-center-class TPU aboard a prototype satellite, successfully generating a haiku in orbit, according to Google’s October 30 update. This proves an early computing milestone, not a functioning orbital data center; radiation, heat, networking, and maintenance remain major challenges.
💻 AI Coding & Developer Tools
- Anthropic confirmed it adjusted Claude Code’s Opus 5.5 instructions to steer the model away from reproducing someone else’s specific video game. Anthropic said original games, genre-inspired games, and games using the user’s own designs and assets should still be allowed; the policy is intended to discourage intellectual-property infringement.
- OmniDiff compares changes in code, PDFs, images, fonts, and archives with content-aware views; it is a conventional developer tool, not an AI-powered product, despite its AI-assisted development.
- DeepSeek’s dsh-superagent is open-source research code for a coding agent that responds to authorized GitHub requests or team-written rules; its maintainers say it is not an official, production-ready DeepSeek product.
🔬 AI Research & Models
- Randall Balestriero and Yann LeCun analyzed how separately trained text and image models can learn compatible representations in a new JEPA paper. JEPA trains models to predict useful features of related observations instead of every next word or pixel. Their theorem establishes alignment under specified assumptions about the training views. On the same dataset, their “good views” approach required 42 times fewer samples to reach comparable alignment than next-token prediction; a 109M-parameter text JEPA also beat the language models they evaluated, up to 72B parameters, on the reported alignment tests. Those are benchmark-specific findings, not proof that small models outperform large ones generally.
- Adaptive Agent Memory Research includes a locally runnable memory server for coding agents, but its maintainers describe the broader system as an experimental research project rather than a finished product.
🏛️ AI Policy, Governance & Safety
- Microsoft began offering districts enforceable AI privacy protections in school contracts on November 1. U.S. districts can request the standard in new or existing Microsoft agreements; it does not automatically apply identically to every school. It limits use of student and educator data for AI training and advertising, adds transparency duties, and requires human oversight for consequential decisions. AP detailed additional safeguards.
- AI-generated imitations of Indigenous Australian artwork sparked criticism. Artists raised concerns about the imitation and commercialization of culturally significant visual styles, including through inexpensive online products. The controversy involves questions of cultural ownership that extend beyond conventional copyright enforcement.
- Security researchers and AI developers are confronting how attackers use AI tools. The Economist examined the expanding market for AI-powered cybersecurity amid concern that the same technology can assist attackers. The important distinction is between tools that help an attacker and systems that independently carry out an intrusion.
- OpenAI reportedly planned a post-election policy campaign combining proposed model safeguards, local data-center commitments, worker initiatives, and international standards. Axios said the company also wants the term “AI compute campuses” used for data centers. These are reported proposals, not enacted requirements.
📊 Fundraising & Deals Roundup
- OpenAI reportedly cut its new fundraising target from $30B to $20B, according to TechCrunch citing the FT and WSJ. Those outlets reported a roughly $1.25T valuation and around $11B raised so far; OpenAI has not confirmed the terms.
- Salesforce agreed to acquire Rune Technologies, a company developing software that helps military organizations plan and manage supplies. Subsequent reporting put the deal at approximately $375M, although Salesforce did not disclose that figure in its announcement.
💡 Industry Commentary & Analysis
- South Korea’s record chip sales prompted a wider market debate about how durable the AI memory boom will be. The X topic highlighted strong shipments and reported long-term memory orders, alongside a selloff in semiconductor shares. Those market reactions should not be confused with forecasts that current demand will continue indefinitely.
- Andrej Karpathy’s HackerTube experiment used AI during development to assemble a video directory from Hacker News recommendations, but visitors browse a curated site rather than interact with an AI agent; explore the site itself.
Previous Around the Horn Digests
Catch up on recent coverage:
- Friday, October 30: OpenAI tested model monitors, Reddit criticized AI search traffic losses, and Mistral opened a trillion-parameter model.
- Thursday, October 29: Google introduced Gemini 4 Carbon for security, and GitHub reported more AI-generated code.
- Wednesday, October 28: OpenAI reportedly pursued Google chips, and Anthropic expanded Claude's government reach.
- Tuesday, October 27: OpenAI detailed math research, and SpaceX secured GPU financing.
- Monday, October 26: OpenAI examined containment failures, and Sweden restricted workplace use of AI chatbots.
- Friday, October 23: AI agents entered banking and cybersecurity workflows.
- Thursday, October 22: AI-discovered software flaws prompted patches, while the AI industry's military partnerships faced scrutiny.
That's a Wrap
We've covered roughly 30 distinct developments and tools, from shared AI usage limits to computing in orbit.
If you made it all the way down here, congratulations. You have now spent more time reading about HackerTube than Karpathy reportedly spent building part of it.
For the shorter daily version, subscribe to The Neuron.
See you tomorrow.
P.S. Know someone who wants all the AI news without the endless browser tabs? Forward them this roundup.
