OpenAI skips 2026 IPO as AMI Labs raises $1.03B for world models
Monday, 16 November 2026 | 46 articles
Sam Altman said OpenAI will not go public in 2026 over safety concerns, while AMI Labs raised $1.03B for world models, though its first product timing is unclear. Meanwhile, OpenAI debuted ChatGPT Wallet for AI agent payments, and a report warns that one hijacked MCP server can turn trusted agent tools against users.
Published by Martin Ševčík 16 November 2026 at 06:10
Two of the biggest AI labs are now telling the public, in almost the same breath, that they are worried about what they are building. Sam Altman told Fortune that OpenAI will not go public in 2026, citing safety concerns. Meanwhile Anthropic's I.P.O. filing speaks about existential risk with a frankness you rarely see in a document written for investors. I find this fascinating, and also a little uncomfortable. Is candor about catastrophe reassuring, or is it simply the cost of doing business when you want to raise billions? I lean towards cautious respect for the honesty, but I notice that honesty and restraint are not the same thing. Anthropic is still heading for the public markets, and OpenAI's delay may just as easily be about timing and governance as about safety.
The more concrete story, I think, is happening inside the enterprise. Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in January. That is a remarkable adoption curve, even allowing for the usual analyst enthusiasm. It also explains why the warning about hijacked MCP servers deserves more attention than it is getting. If an agent trusts its tools, and one compromised server can turn those tools against the user, then the attack surface is growing faster than most security teams can map it. Anyone building with agents should treat every connected server as a potential point of failure, not a convenience.
The commercial plumbing is arriving at the same speed. OpenAI's ChatGPT Wallet, built with American Express, Visa and Mastercard, ties agent-driven purchases to explicit user approval. By the way, I think that approval step is the whole product. Payment networks have spent decades building trust and dispute mechanisms, and they clearly want to be the layer that makes agents safe to give a credit card to. It also shows where OpenAI sees its future: not just answering questions, but completing transactions.
Then there is Microsoft's decision to enable Anthropic's Claude by default for UK Copilot tenants on December 9. For a company so closely tied to OpenAI, that is a notable hedge. It suggests that model choice is becoming a feature of the platform rather than a loyalty, and that enterprises will increasingly get a mix of models whether they asked for one or not. For British organisations, it also means a quiet change in their data flows that is worth checking before the date arrives.
On the research side, Yann LeCun's AMI Labs has raised $1.03 billion for world models, though there is no clear timeline for a first product. I am sceptical of any bet that large without a roadmap, but I also think LeCun has been consistent for years that language models alone will not get us to robust machine intelligence. A smaller sign of the same direction: Italy's Mosaico raised €7.5 million to tackle the sensor data bottleneck in robotics, which is the unglamorous problem that decides whether embodied AI actually ships.
The question I keep returning to is whether safety rhetoric and deployment speed can keep coexisting. At some point, one of them will have to give.
List of sourced links used in the brief
Importance:NewsOpenAI IPO plans
Altman: OpenAI won't go public in 2026 over safety concerns
OpenAI CEO Sam Altman told Fortune in an interview published Saturday that the company will not pursue an IPO in 2026. He cited safety concerns related to artificial intelligence as the reason for holding off. Source: spokesman.com
Importance:Researchscientific applications
GPT-6 Astra helps sketch a lattice definition of the Standard Model
Researchers describe a candidate nonperturbative lattice formulation of the Standard Model. They explain how GPT-6 Astra helped turn the initial idea into a detailed proof. Source: openai.com
Importance:Launchmodel release
Claude Fable 5.2 paints images with code instead of generating them
Anthropic's new language model Claude Fable 5.2 can produce complex, even photorealistic paintings. It does so by using code to place hundreds of thousands of individual elements rather than generating images directly. Source: the-decoder.com
Importance:Policyhardware
Congress eyes tighter AI chip export rules after Jensen Huang spy link
After revelations about Jensen Huang's relationship with a suspected Chinese spy, Congress wants to crack down on AI chip exports. Source: theverge.com
Importance:Videoprompt engineering
Karpathy shares four prompting tricks for clearer, more creative AI
Andrej Karpathy has outlined four AI prompting techniques aimed at improving the clarity and creativity of model outputs. The post circulated under hashtags like #AI and #PromptEngineering. Source: youtube.com
Importance:VideoAI usage habits
How Andrej Karpathy uses AI: habits of the top 0.2%
Andrej Karpathy co-founded OpenAI, led AI at Tesla, coined the term 'vibe coding', and now works at Anthropic. He says only around 0.2% of people have actually seen AI agents in action. The piece looks at the habits that set such users apart. Source: youtube.com
Importance:ResearchLLM architecture
Positional embeddings in LLMs, from GPT-2 to RoPE and beyond
Sebastian Raschka, PhD, published a technical walkthrough of positional embeddings in large language models on his Ahead of AI newsletter. It traces how the techniques evolved from GPT-2 to RoPE and newer approaches. Source: news.lavx.hu
Importance:Researchmultimodal AI
KAIST lip-reading AI decodes clear words first to resolve ambiguity
A team led by Professor Yong Man Ro at KAIST's School of Electrical Engineering built an AI system that first identifies words with clearly distinguishable lip movements. It then uses those results to help interpret more ambiguous mouth shapes. Source: eurekalert.org
Importance:LaunchLLM optimization
Nota signs deal covering an optimized LLM and server
Nota, an AI software company focused on model compression and optimizing inference and serving, announced on the 16th that it has signed an agreement involving an optimized large language model (LLM) and server. Source: mk.co.kr
Importance:Researchmultimodal AI
KAIST AI reads lips without sound, offering hope for hearing aids
KAIST researchers developed an AI that reads lip movements by identifying the clearer words first, then using surrounding context to interpret the rest. The work raises hopes for assistive tools for people with hearing difficulties. Source: koreabiomed.com
Importance:Newsscience
OpenAI launches a science blog for AI-driven research
The OpenAI science blog will share new findings, practical methods and firsthand accounts from researchers using AI to advance scientific discovery. Source: openai.com
Importance:Researchmultimodal medical AI
AI model screens toddlers for autism from short play videos
A vision-language AI model trained on three-minute videos of natural play was able to accurately screen toddlers aged 15 to 24 months for autism. The approach suggests brief, unstructured recordings could support early screening. Source: bioengineer.org
OpenAI debuts ChatGPT Wallet for secure AI agent payments
Visa's Rubail Birwadker describes how ChatGPT Wallet, built with American Express, Visa and Mastercard, ties AI-driven purchases to explicit user approval. Source: fintechmagazine.com
Importance:NewsAI agent security
A single hijacked MCP server can turn trusted AI agent tools against you
Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, compared with under 5% in January. The growth raises the stakes for securing the MCP servers those agents rely on. Source: venturebeat.com
Importance:Newsindustry-specific AI agents
AI agents in mortgage lending: how they help close more loans
AI agents let mortgage lenders answer every lead, keep applications moving and gather documents. The piece covers use cases, metrics and compliance requirements. Source: sierra.ai
Importance:Newspractical AI agent tools
How to use ChatGPT Dots: a practical guide to AI agents
AI is no longer just a tool for answering questions. With ChatGPT Dots, users can hand an AI agent ongoing tasks to manage. Source: daily-sun.com
Importance:NewsAI agent blockchain transactions
XRP Ledger handles 11.7 million sub-dollar AI agent microtransactions
Chandler Fang, founder of AI payments startup t54.ai, says the company's x402 Facilitator on the XRP Ledger (XRPL) has processed more than 10 million tiny payments made by AI agents. Source: tradingview.com
Importance:PolicyAI trust and governance
OneTrust: AI trust will be the big business test in 2027
OneTrust's 2027 predictions report argues that Australian organisations will face growing commercial and regulatory pressure to prove their AI can be trusted. Source: australiancybersecuritymagazine.com.au
Importance:NewsAI agent payment systems
Mastercard backs both cards and crypto for AI agent payments
OpenAI's new ChatGPT Wallet relies on card tokenization with Mastercard, which also supports a rival AI agent payment system built on Coinbase's x402. Source: coinmarketcap.com
Importance:LaunchAI agent authentication
ChatGPT Wallet: Face ID is the only human check on AI agent card payments
OpenAI launched ChatGPT Wallet on Nov. 13 for US ChatGPT Pro subscribers. Always-on agents can pay using one-time card network tokens once the user confirms with Face ID. Source: techtimes.com
Importance:Newsspecialized AI agents
Naver starts beta test of AI agent that mines insights from cafe comments
Naver announced on the 16th that it began beta testing its AI Cafe Agent on the 9th. Source: biz.chosun.com
A close read of Anthropic's I.P.O. filing and its frank talk of existential risk
Anthropic's prospectus stands out for how openly it discusses existential risk. The open question is whether that candor is actually reassuring. Source: nytimes.com
Importance:Newsexistential risk
Anthropic safety researchers sound the alarm that AI could pose a lethal threat
As AI grows more capable, some of the people building it are now warning about the dangers. One Anthropic researcher voices serious concern about where the technology is heading. Source: media.unisba.ac.id
Importance:Newsalignment
Ex-OpenAI board member who ousted Sam Altman: we still can't trust AI CEOs
Helen Toner, executive director of Georgetown's Center for Security and Emerging Technology, argues that the core AI safety problems, including alignment and control, remain unsolved. She says that is why the public cannot simply rely on the leaders of AI companies. Source: finance.biggo.com
Importance:Newsexistential risk
OpenAI Foundation board member warns AI accidents could become a leading cause of death
Paul Christiano, a board member of the OpenAI Foundation, warned at a Geoffrey Hinton lecture in Toronto that AI accidents could soon become a leading cause of death for humanity. Source: finance.biggo.com
Importance:Newssafety
Nathan Labenz at Bletchley Park: AI agents are now capable enough to be dangerous
Nathan Labenz, host of The Cognitive Revolution and a GPT-4 red-teamer, told IT Revolution's Enterprise AI Summit that AI agents are now capable enough to be genuinely dangerous. He delivered the keynote at Bletchley Park. Source: finance.biggo.com
Importance:Newsexistential risk
Former Anthropic researcher quits and warns AI could destroy humanity
A former Anthropic researcher has left one of the world's leading AI companies. He is now raising the alarm about where artificial intelligence is headed. Source: media.unisba.ac.id
Importance:Opinionalignment
Securing model weights alone won't stop China, says Redwood Research CEO
Buck Shlegeris, CEO of the AI alignment nonprofit Redwood Research, estimates roughly a one-in-three chance that AI companies build superintelligence within the coming years. He argues that locking down model weights is not enough to keep China from catching up. Source: finance.biggo.com
Microsoft to enable Anthropic's Claude by default for UK Copilot tenants on Dec. 9
Microsoft plans to turn on Anthropic's Claude models by default in Microsoft 365 Copilot for UK tenants on December 9, according to a message center post. The change brings Claude into the default Copilot experience for UK organizations. Source: finance.biggo.com
Importance:Launchdeveloper tools
datasette-apps 0.4a0 brings sandboxed apps to Datasette
Datasette Apps are sandboxed HTML and JavaScript applications that run inside Datasette. They work much like Claude Artifacts, except that these apps are able to execute SQL queries. Source: simonwillison.net
The new release supports OpenAI-compatible endpoints, letting you point Datasette at an LLM hosted anywhere that speaks the OpenAI pseudo-standard. Source: simonwillison.net
Importance:Launchplatform launches
Datasette 0.65.6 released
A new version of Datasette, the open source multi-tool for exploring and publishing data, has been released. The announcement comes from Simon Willison's beat. Source: simonwillison.net
Importance:ResearchDeveloper Tools
Cursor vs Copilot: what Stack Overflow survey data really says
A single Stack Overflow sentence containing two figures can easily mislead a careful reader. The 2025 survey's AI page reported that 18% of developers used the tools in question, a number that needs context to interpret correctly. Source: shattered.io
Importance:NewsBusiness Platforms
Microsoft Ignite starts Nov. 17: will Microsoft update its paid Copilot seat count?
Microsoft (NASDAQ:MSFT) hosts Ignite, its annual conference for business customers and partners, starting Tuesday, Nov. 17. Attention is on whether the company will share a fresh figure for paid Copilot seats. Source: finance.yahoo.com
Importance:NewsDeveloper Tools
GitHub announces Copilot model deprecations set for December 14
GitHub has announced plans to deprecate six Copilot models on December 14. Developers using the affected models will need to switch to supported alternatives. Source: casrai.org
Importance:NewsBusiness Platforms
Microsoft Ignite opens Nov. 17 with Copilot set to dominate the agenda
Microsoft Ignite 2026 runs Nov. 17-20 in San Francisco. Satya Nadella's keynote is expected to center on Copilot and usage-based billing. Source: cryptobriefing.com
This is a fresh release of Datasette's longest-neglected plugin, which renders charts with Vega-Lite from table data or custom queries. Source: simonwillison.net
Mosaico lands €7.5M seed round to break robotics' data bottleneck
Italian startup Mosaico has closed a €7.5m seed round. The money will go toward helping companies handle the huge volumes of sensor data needed to develop robotics systems. Source: app.dealroom.co
Importance:Newsrobot learning
Italian robotics startup Mosaico secures €7.5m to solve AI's data headache
Mosaico, a startup from Italy, has raised €7.5m to help robotics firms organise the data required to train AI models. Its goal is to make that data far easier to manage. Source: sifted.eu
AMI Labs raises $1.03B for world models; first product timing unclear
Yann LeCun's AMI Labs has raised $1.03 billion to develop world models. Coverage speculates that the startup's first product could be an official AMI-branded open model or a free developer API preview, though no timeline has been confirmed. Source: futuresearch.ai
Importance:ThreadAI risks
Demis Hassabis warns of an AI 'oil spill' unless industry shifts course
DeepMind chief Demis Hassabis cautions that an AI-driven 'oil spill' could be on the way if the industry does not change direction. The snippet is brief and offers few details on what such a disaster would involve. Source: reddit.com
As more companies build AI agents, they also need tools to monitor them, and that demand is driving LangChain's multibillion-dollar valuation. The startup has raised $300 million at a $3 billion valuation. Source: trendingtopics.eu
Importance:NewsRobotics funding
Agile Robots lands $500m at $8.5bn valuation
Agile Robots has raised $500m at an $8.5bn valuation, with SoftBank and NVIDIA among its backers. The money will fund the expansion of its Physical AI robotics and manufacturing operations worldwide. Source: startupvalley.news
Importance:NewsAI agent framework
LangChain closes $300M Series C at $3B valuation
AI agent framework developer LangChain announced a $300 million Series C round at a $3 billion post-money valuation. The announcement was dated November 13, 2026. Source: casrai.org
WSJ: Trump administration plans to name two dozen-plus AI 'national champions'
According to the Wall Street Journal, the Trump administration intends to designate more than two dozen artificial intelligence companies as U.S. 'national champions'. Nvidia is expected to be on the list. Source: reuters.com
Importance:Newspolicy/government impact
Trump administration to tag Nvidia as an AI 'national champion'
A new government program will designate select U.S. AI companies, Nvidia among them, as 'national champions'. The initiative aims to strengthen their ability to win business in overseas markets. Source: wsj.com
Published by Martin Ševčík
16 November 2026 at 06:10
Two of the biggest AI labs are now telling the public, in almost the same breath, that they are worried about what they are building. Sam Altman told Fortune that OpenAI will not go public in 2026, citing safety concerns. Meanwhile Anthropic's I.P.O. filing speaks about existential risk with a frankness you rarely see in a document written for investors. I find this fascinating, and also a little uncomfortable. Is candor about catastrophe reassuring, or is it simply the cost of doing business when you want to raise billions? I lean towards cautious respect for the honesty, but I notice that honesty and restraint are not the same thing. Anthropic is still heading for the public markets, and OpenAI's delay may just as easily be about timing and governance as about safety.
The more concrete story, I think, is happening inside the enterprise. Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in January. That is a remarkable adoption curve, even allowing for the usual analyst enthusiasm. It also explains why the warning about hijacked MCP servers deserves more attention than it is getting. If an agent trusts its tools, and one compromised server can turn those tools against the user, then the attack surface is growing faster than most security teams can map it. Anyone building with agents should treat every connected server as a potential point of failure, not a convenience.
The commercial plumbing is arriving at the same speed. OpenAI's ChatGPT Wallet, built with American Express, Visa and Mastercard, ties agent-driven purchases to explicit user approval. By the way, I think that approval step is the whole product. Payment networks have spent decades building trust and dispute mechanisms, and they clearly want to be the layer that makes agents safe to give a credit card to. It also shows where OpenAI sees its future: not just answering questions, but completing transactions.
Then there is Microsoft's decision to enable Anthropic's Claude by default for UK Copilot tenants on December 9. For a company so closely tied to OpenAI, that is a notable hedge. It suggests that model choice is becoming a feature of the platform rather than a loyalty, and that enterprises will increasingly get a mix of models whether they asked for one or not. For British organisations, it also means a quiet change in their data flows that is worth checking before the date arrives.
On the research side, Yann LeCun's AMI Labs has raised $1.03 billion for world models, though there is no clear timeline for a first product. I am sceptical of any bet that large without a roadmap, but I also think LeCun has been consistent for years that language models alone will not get us to robust machine intelligence. A smaller sign of the same direction: Italy's Mosaico raised €7.5 million to tackle the sensor data bottleneck in robotics, which is the unglamorous problem that decides whether embodied AI actually ships.
The question I keep returning to is whether safety rhetoric and deployment speed can keep coexisting. At some point, one of them will have to give.