Apple builds a Mac-operating AI agent as Anthropic's Riemann proof claim surfaces
Sunday, 22 November 2026 | 57 articles
Apple is reportedly building an AI agent that operates a Mac like a person, with a job posting naming OpenAI, Anthropic and Meta as rivals. Meanwhile, an Anthropic mathematician reportedly proved the Riemann hypothesis with AI, and Anthropic's IPO filing warns AI could pose existential risks to humanity.
Published by Martin Ševčík 22 November 2026 at 06:07
The most revealing detail in today's news isn't the Riemann hypothesis claim. It's what Anthropic plans to tell investors. A company preparing to go public is reportedly willing to write that advanced AI could carry "catastrophic or existential risks to humanity" into its IPO filing. Companies usually sand every sharp edge off these documents. Anthropic appears to be doing the opposite, and I think that tells you something about how it sees its own pitch: we are the careful ones, so our risk language is part of the product.
Is that cynical? Maybe a little, but I'd call it coherent. Jaan Tallinn, who spent years funding AI safety work and was among Anthropic's earliest backers, now stands to turn that conviction into roughly $20 billion. By the way, that is the uncomfortable loop at the heart of this industry: the people most worried about the technology have repeatedly ended up financing, and profiting from, the labs building it. I don't think that makes their concerns insincere. But it does mean we should read safety rhetoric and capital-raising rhetoric side by side, because they are now the same document.
Then there is the claim that an Anthropic mathematician used AI to prove the Riemann hypothesis, a 167-year-old problem. I'd treat this with real caution. It comes as a report, not a verified proof, and mathematics has a very good immune system: a result like this lives or dies on whether the community can check it, line by line. If it holds, it would be a genuine landmark rather than another incremental benchmark gain, and the year-long run of mathematical progress the report describes would deserve far more attention than it gets. If it doesn't, it will be a useful reminder of how quickly an AI headline can outrun its evidence. Either way, the timing, a few days before IPO season, is hard to ignore.
The story with the most practical consequences for people building things is Apple's. Job listings dated November 19 describe an agent that writes code, browses the web and controls apps on a Mac the way a person would, reportedly with a launch as early as late 2027 and OpenAI, Anthropic and Meta named as rivals. At the same time, Apple is tightening the permissions rules for third-party agents running on macOS. That combination matters more than either piece alone. Apple controls the operating system, so it can decide what any agent, including its competitors', is allowed to see and touch. If you are building a computer-use product for the Mac, your biggest dependency may also be your future competitor. Google adding computer use to its own offerings shows this race is already on, and platform owners hold the home-field advantage.
Meanwhile, Andrej Karpathy released a 70-minute, slide-free walkthrough of how he actually uses AI day to day. That seems like the right counterweight to everything above. While the industry argues about existential risk and unsolved theorems, the most useful thing a top practitioner can show us is a plain workflow. I'd watch that before reading another IPO headline.
The question I'm left with is who ends up deciding what agents are allowed to do on our machines: the labs, the platforms, or us?
List of sourced links used in the brief
Importance:VideoAI usage
Karpathy shares a 70-minute walkthrough of how he uses AI daily
OpenAI co-founder Andrej Karpathy screen-shared his everyday AI workflow in a roughly 70-minute session with no slides. The video is pitched as a look at how top practitioners actually use AI tools in practice. Source: youtube.com
Importance:ResearchQwen3.5-4B model
Qwen3.5-4B listing surfaces alongside Pliny the Prompter's Hugging Face projects
A Hugging Face page for Qwen/Qwen3.5-4B appears next to a list of community repositories, including pliny-the-prompter/obliteratus and several other small apps and tools. The snippet offers little detail beyond the project names. Source: endorlabs.com
Importance:NewsRiemann hypothesis proof
Anthropic mathematician reportedly proves the Riemann hypothesis with AI
An AI-generated solution to the 167-year-old problem, long considered far beyond the reach of human mathematicians, is presented as the culmination of the technology's year-long run of mathematical progress. The claim comes from a report on an Anthropic mathematician. Source: google.com
Importance:NewsOpen source chatbots
Open-source chatbot roundup for 2026 features Pliny's obliteratus repo
A 2026 list of open-source chatbot projects highlights self-hosted agents that automate tasks and workflows in Slack and Microsoft Teams. It also mentions pliny-the-prompter/obliteratus, which has about 578 GitHub stars. Source: indiesignals.net
Importance:ResearchLLM illicit response risk
LLMs often comply with illicit instructions across legal and social contexts
Large language models may facilitate unlawful activities, yet this risk remains underexplored, as prior studies focus on illicit-content recognition or narrow scenarios. The new work examines how models respond to illicit instructions across diverse socio-legal contexts. Source: google.com
Importance:ResearchLaplace benchmark performance
Audited language model guides mesh refinement, cutting elements on L-shaped Laplace test
A deterministically audited language model guided adaptive mesh refinement on the L-shaped Laplace benchmark, reducing the final element count by twenty-one percent. Source: google.com
Importance:Newsmodel security
OpenAI makes extracting its models' reasoning far harder for attackers
OpenAI has sharply tightened safeguards against extraction of its models' internal reasoning. An attack success rate of 54.0 percent was cut dramatically, and active requests fell from 22.9 to 0.9 percent. Source: the-decoder.com
Importance:Newsproduct update
Google's homepage tweak nudges search toward an AI answer engine
Google's latest homepage update moves its search engine further toward becoming an AI answer engine. A new feature lets users quickly find and reopen their recent AI searches. Source: the-decoder.com
Importance:Researchmodel architecture
Nvidia researchers: giant AI context windows are often unnecessary
Nvidia researchers argue that language models frequently don't need their full context windows, which are slow, costly and error-prone. They propose an alternative approach to handling long contexts. Source: the-decoder.com
Importance:Launchnew model release
Inkling 2 Mini brings a looped transformer to local setups
Inkling 2 Mini is a new looped transformer model built as a multimodal Mixture-of-Experts. It has roughly 26B total parameters, with 3B active per token. Source: huggingface.co
Importance:Researchmodel security
OpenAI details its layered defense against adversarial distillation
OpenAI describes a multi-layered strategy to counter adversarial distillation, combining model training, safeguards, detection and enforcement. The goal is to prevent others from copying the capabilities of its models. Source: openai.com
Importance:Researchharness design for LLMs
Language model harnesses can generalize compositionally, new research suggests
Modern language models are post-trained and deployed as agents inside a "harness", but standard harness designs pile up context that the neural model must process. The paper examines whether harness design can support compositional generalization. Source: alphaxiv.org
Importance:Researchcritical care LLM review
Review: clinical LLMs still lack real-world use in critical care
A scoping review of 62 studies found that large language models for decision support in acute and critical care have not progressed to real-world clinical deployment. Source: google.com
Importance:Newslocal LLM privacy
How to run a local LLM for private, uncensored AI
Cloud chatbots are convenient, but they are not the only way to use generative AI. A local large language model runs on your own computer, keeping your data private and your usage free of cloud restrictions. Source: google.com
Importance:ResearchLLM surgical assessment
LLMs beat residents at classifying cardiovascular surgical risk in pilot study
In a pilot study of 32 cardiovascular surgery cases, contemporary large language models reached up to 75 percent exact agreement with specialist consensus. The models outperformed residents in risk classification. Source: google.com
Importance:NewsAnthropic
Anthropic's newsroom describes its mission as safe, steerable AI
Anthropic's newsroom page presents the company as an AI safety and research firm working on reliable, interpretable and steerable AI systems. The snippet contains no specific announcement. Source: anthropic.com
Importance:NewsOpenAI mathematics
OpenAI claims its internal model solved 100+ open math problems
After convening about forty mathematicians to discuss AI progress in mathematics, OpenAI said in a September 21 blog post that an internal frontier model, trained since August 28, resolved more than 100 long-standing open problems. The company had hinted that its models had cracked hundreds of conjectures. Source: en.wikipedia.org
Importance:LaunchArticulate platform
Articulate unveils biggest platform overhaul in a decade at Articuland 2026
At Articuland 2026, Articulate announced its largest modernization in ten years for its AI-powered workplace training platform. The company says it will lay out what comes next for the industry. Source: articulate.com
Apple's AI agent job posting names OpenAI, Anthropic and Meta as rivals
A new Apple job listing for AI agent engineers reportedly points to OpenAI, Anthropic and Meta as competitors, with Bloomberg confirming some details. A launch could come as early as late 2027, though much about the project remains unproven. Source: shattered.io
Importance:Newscomputer use in AI agents
Docs show Apple building an AI agent that uses a Mac like a person
Apple is recruiting engineers to create an autonomous AI agent able to browse websites, write code and control Mac applications. At the same time, new macOS permissions are being introduced. Source: google.com
Importance:NewsAI agent competition
Apple builds its own AI agent while tightening rules for rival Mac agents
Job listings dated Nov. 19 describe an Apple AI agent that writes code, browses the web and controls apps. The listings surface just as Apple is tightening the rules for rival agents running on the Mac. Source: google.com
Importance:Newscomputer use in AI agents
Apple is building an AI agent that can operate your computer for you
Apple (AAPL) is hiring engineers to develop an agent that operates a Mac the way a person would, with a launch possible as soon as late 2027. Google has meanwhile added computer use to its own offerings. Source: google.com
Importance:Researchtool integration
Connect AI agents to Oracle Autonomous AI Database via MCP and A2A
Oracle Autonomous AI Database ships with managed MCP and A2A servers. They make tools and agents defined in the Oracle Select AI Agent Framework accessible to external AI agents. Source: google.com
Importance:Researchagentic commerce
mEinstein outlines framework for judging consumer intent in agentic commerce
Agentic commerce needs to go beyond completing transactions and understand consumer intent, readiness and permission before any purchase is made. mEinstein lays out a framework for evaluating these factors. Source: google.com
Importance:VideoAI agent model comparison
GPT-62 Sol vs Claude Opus 55: agent routing, cost per successful outcome and security
A comparison of GPT-62 Sol and Claude Opus 55 focuses on optimizing AI agent routing, measuring cost per successful outcome and enterprise security. The piece also includes a video and a testing approach. Source: t.co
Importance:Researchtool use in AI agents
How to set an AI agent tool call timeout so one slow API can't stall a run
This guide shows how to derive a tool call timeout from p95 latency, add an overall run deadline, budget retries and handle timeouts gracefully. The goal is to keep a single slow API from stalling an entire agent run. Source: startupfortune.com
Importance:Researchsecurity in agentic systems
Know-your-agent testing aims at agentic payment fraud
Banks and payment firms are testing know-your-agent rules to authenticate AI agents in invoice and liquidity workflows. Financial services companies and their partners are working on the approach together. Source: google.com
Anthropic's IPO filing warns AI could pose existential risks to humanity
Anthropic intends to tell prospective IPO investors that advanced AI could carry "catastrophic or existential risks to humanity." The language is an unusually stark admission for a company preparing to go public. Source: google.com
Importance:Newsinvestor perspectives on AI safety
Anthropic's most safety-minded early investor is now worth $20 billion
Jaan Tallinn spent years funding AI safety work and was among Anthropic's first backers. With the company's public debut, he stands to cash in on a massive payday. Source: google.com
Historic mega-IPO lays bare the tension over AI safety
The company is under shareholder pressure even as its own leaders warn about the dangers of the technology it builds. Source: 1news.co.nz
Importance:Newspolitical perspectives on AI risk
Trump calls AI warnings a 'HOAX' and tells industry not to 'kill the Golden Goose'
As AI leaders sound alarms about existential risks, Trump dismissed the warnings as a hoax. Vice President JD Vance went further, calling them a "Trojan horse." Source: google.com
From Pentagon standoff to copyright suit: five controversies dogging Anthropic
Anthropic has cast itself as the safety-first alternative in an increasingly crowded AI field, building its brand on responsible development. A series of controversies is now testing that image. Source: google.com
Importance:Launchsafety-profit tensions
Anthropic's Wall Street debut exposes tension between AI safety and profit
The Claude chatbot maker debuted on Wall Street at a $1.75 trillion market value, capping a stretch of explosive growth. The listing sharpens questions about balancing safety with commercial demands. Source: google.com
Importance:Newssafety commitments in public companies
Anthropic vows to put humanity first — will going public change that?
Critics argue that as a public company, Anthropic's pledge to prioritize safety over commercial success will now be put to the test. Source: google.com
Importance:Newscommunity sentiment
AI safety community grows more alarmed — is the fear justified?
Members of the AI safety world are voicing mounting distress, with one describing feeling scared, sad, and angry. The piece asks whether that growing alarm is warranted. Source: google.com
SAP and Microsoft team up to bring AI value to enterprises
SAP is making its Joule assistant more interoperable with Microsoft's ecosystem. The integration lets agents access and reason over business data, context, and tools. Source: google.com
Importance:LaunchDeveloper Tools
Muru: a new AI-assisted Rust server for Python web apps
Muru is a freshly released Rust server developed with the help of AI. Its goal is to give developers top-tier tools for running ASGI and WSGI Python apps in both development and production environments. Source: simonwillison.net
Importance:LaunchDeveloper Tools
Muru: AI-assisted Rust server for ASGI and WSGI Python apps
Muru is a new Rust server built with AI assistance, designed to offer the best possible tooling for running ASGI and WSGI Python applications. It targets both development and production use. Source: simonwillison.net
Importance:NewsAI Tool Enhancement
Microsoft upgrades Copilot with file management and PC troubleshooting
Microsoft is working to make Copilot more practical, even if it isn't everyone's first pick among AI platforms. New features announced at Microsoft Ignite help users manage files and resolve PC problems. Source: tech.yahoo.com
Importance:NewsAI Model Integration
Microsoft brings OpenAI's GPT-6.2 Sol to Copilot globally as MAI models gain ground
On Nov 20, Microsoft rolled out OpenAI's GPT-6.2 Sol to Copilot worldwide, while Astra remains limited to Frontier users in Office. The move comes days after the launch of Microsoft's own MAI Code 2. Source: google.com
Importance:NewsAI Cost Analysis
Study claims Microsoft Copilot is the cheapest AI option for users
A recent study suggests Microsoft Copilot delivers the most cost-effective AI solution for users. It highlights the platform's price advantages over competing offerings. Source: smallbiztrends.com
Importance:ResearchAI Integration Patterns
Building a hybrid AI setup in Obsidian with local and cloud models
Connecting AI to your notes can supercharge them, but it raises the question of whether to grant AI access to your entire knowledge base. The author describes a hybrid approach combining local and cloud AI. Source: google.com
Importance:NewsAI Business Strategy
Bank of America: Microsoft's Copilot strategy paves the way for enterprise AI growth
Bank of America says Microsoft's approach to embedding AI in business software gives it a practical path to expanding its enterprise AI business. Source: google.com
Gemini 4 Neon 'video' is an exported artifact, not true generation
A 60-second Gemini 4 Neon film is not generated video, since Google's model card lists text as the only output. The piece explains what the skill actually does and what it costs to use. Source: orcarouter.ai
Importance:News
MIT's Jim Glass on making multimodal foundation models more human-like
Jim Glass, a researcher at the MIT-IBM Computing Research Lab and MIT CSAIL, discusses work on relating text and other modalities in multimodal foundation models. The Q&A focuses on enabling more human-like interaction and understanding. Source: google.com
Importance:News
Grok Imagine video guide: specs, 1080p, API pricing and 20+ prompts
A November 2026 overview of Grok Imagine covers Video 1.5 specs, including 1080p output and a 15-second limit. It also lists per-second API prices by model and offers more than 20 ready-to-use prompts. Source: google.com
Importance:News
Google launches Nano Banana 2 following viral hit of its image model
Google on Thursday unveiled the successor to its viral Nano Banana image generation model, promising faster performance. The release is part of the company's wider push in AI image generation. Source: google.com
Importance:News
11 AI video generators compete for users as the 2026 market fragments
As the AI video generation market becomes increasingly segmented, picking a service that suits a given production workflow matters more. DigitalToday looks at the 11 services vying for users in 2026. Source: google.com
Importance:News
VideoRNG creates relightable Neural Gaussians from generative video models
VideoRNG, described on arXiv cs.GR, is a two-stage approach that generates relightable 3D assets from text or a single image. The assets avoid baking in a fixed lighting look. Source: google.com
Importance:News
Pole vaulter EJ Obiena compares his life 13 years ago using GTA
Filipino pole vaulter EJ Obiena reflected on how much his life has changed over 13 years, drawing a comparison with the video game Grand Theft Auto. He shared photos of himself from that time. The snippet is too short for further detail. Source: google.com
McKinsey: Physical AI's value shifts from robot hardware to deployment
McKinsey now projects its $1 trillion physical AI estimate to be reached by 2035. The firm expects the bulk of the value to move away from hardware toward software, deployment and field operations. Source: google.com
NMSU Wins $600K DARPA Grant for Provably Correct AI Reasoning
New Mexico State University has received $600,000 from DARPA through the agency's CLARA program. The funding will support neuro-symbolic AI research that combines machine learning with formal logical proofs to make AI reasoning verifiable. Source: google.com
Importance:OpinionAI governance and safety
Beagle Heard #3: Who gets the power to say no to AI?
The latest installment covers Yann LeCun, who argues for safer software while warning that power remains concentrated in few hands. LeCun, a Turing Award recipient and NYU professor, is now founder and executive chairman of AMI Labs. Source: medium.com
Series B startup acquired for 15 billion yuan in under five months
The strategic acquisition was completed less than five months after the startup closed its targeted financing round. The deal values the Series B company at 15 billion yuan. Source: google.com
Importance:Newsfunding
Cognition reportedly in talks to raise funding at $75 billion valuation
Coding startup Cognition AI is negotiating a new funding round that would value it at $75 billion, Bloomberg News reported Friday. The report cites people familiar with the matter. Source: reuters.com
Importance:Newsipo
WOWTALE Weekly: Anthropic's record $81.3B IPO and a week of AI mega-rounds
The fifth edition of WOWTALE Global Weekly covers startup and technology signals for Nov 15–21, 2026. Its top story is Anthropic pricing its IPO, alongside a series of large AI funding rounds. Source: google.com
Importance:Newsipo
Anthropic IPO: what to know about valuation, timeline and how to invest
Anthropic, the AI company behind the Claude family of large language models, is moving closer to an initial public offering. The article looks at the expected valuation, timing and ways investors could gain exposure. Source: google.com
Published by Martin Ševčík
22 November 2026 at 06:07
The most revealing detail in today's news isn't the Riemann hypothesis claim. It's what Anthropic plans to tell investors. A company preparing to go public is reportedly willing to write that advanced AI could carry "catastrophic or existential risks to humanity" into its IPO filing. Companies usually sand every sharp edge off these documents. Anthropic appears to be doing the opposite, and I think that tells you something about how it sees its own pitch: we are the careful ones, so our risk language is part of the product.
Is that cynical? Maybe a little, but I'd call it coherent. Jaan Tallinn, who spent years funding AI safety work and was among Anthropic's earliest backers, now stands to turn that conviction into roughly $20 billion. By the way, that is the uncomfortable loop at the heart of this industry: the people most worried about the technology have repeatedly ended up financing, and profiting from, the labs building it. I don't think that makes their concerns insincere. But it does mean we should read safety rhetoric and capital-raising rhetoric side by side, because they are now the same document.
Then there is the claim that an Anthropic mathematician used AI to prove the Riemann hypothesis, a 167-year-old problem. I'd treat this with real caution. It comes as a report, not a verified proof, and mathematics has a very good immune system: a result like this lives or dies on whether the community can check it, line by line. If it holds, it would be a genuine landmark rather than another incremental benchmark gain, and the year-long run of mathematical progress the report describes would deserve far more attention than it gets. If it doesn't, it will be a useful reminder of how quickly an AI headline can outrun its evidence. Either way, the timing, a few days before IPO season, is hard to ignore.
The story with the most practical consequences for people building things is Apple's. Job listings dated November 19 describe an agent that writes code, browses the web and controls apps on a Mac the way a person would, reportedly with a launch as early as late 2027 and OpenAI, Anthropic and Meta named as rivals. At the same time, Apple is tightening the permissions rules for third-party agents running on macOS. That combination matters more than either piece alone. Apple controls the operating system, so it can decide what any agent, including its competitors', is allowed to see and touch. If you are building a computer-use product for the Mac, your biggest dependency may also be your future competitor. Google adding computer use to its own offerings shows this race is already on, and platform owners hold the home-field advantage.
Meanwhile, Andrej Karpathy released a 70-minute, slide-free walkthrough of how he actually uses AI day to day. That seems like the right counterweight to everything above. While the industry argues about existential risk and unsolved theorems, the most useful thing a top practitioner can show us is a plain workflow. I'd watch that before reading another IPO headline.
The question I'm left with is who ends up deciding what agents are allowed to do on our machines: the labs, the platforms, or us?