From the feeds
A wider view of the AI news landscape. Ranked by the source feed; linked directly to the reporting.
These are supplied feed summaries, separate from our researched daily articles. Feed dates are shown at their supplied precision. Ranking and ratings belong to the upstream feed.
A security researcher reverse engineered the firmware of a Samsung TV and uncovered that the device contains on-board AI components capable of listening to users. This finding raises concerns about privacy and the extent of AI integration in consumer electronics. The discovery was shared on a public forum, prompting discussions about device security and data handling practices in smart TVs.
Dessa, a small AI research lab at Block, has developed a full-duplex speech-to-speech model trained solely on human conversation audio without using text or pretrained models. The model can engage in real-time, natural conversations with emotion, pauses, and nuanced intonation, running locally on high-end MacBook Pro hardware. This approach aims to preserve vocal cues like sarcasm and emotion that are lost when converting speech to text, improving interaction quality and safety in voice AI applications.
Kimi.ai introduced the Kimi K3.1 Pro, a 3-trillion parameter open-weight model featuring a 1-million-token context window, and the K3.1 Pro Highspeed variant offering 2-3 times faster performance. These models deliver enhanced capabilities in coding, 3D modeling, and financial workflows, and are available on Kimi.ai platforms with full model weights to be released by November 4. K3.1 Pro improves on its predecessor through extensive post-training and reinforcement learning, achieving competitive scores on coding benchmarks like Terminal-Bench 4.0 and Kimi Code Bench 3.0. It supports complex tasks such as building human walking simulators, creating detailed 3D assets for games and web experiences, and performing multi-step financial data analysis. The models also integrate with tools like Unity, Unreal Engine, Blender, and CAD software to enable advanced agentic and spatial reasoning capabilities.
An investment research firm, Citrini, collaborated with Anthropic to test whether an unreleased version of the Claude AI model could develop scorpion antivenom. While Claude generated thousands of protein candidates and some showed binding to scorpion toxins, none were proven to neutralize the toxins or function as an effective treatment. The experiment involved designing proteins to bind scorpion venom toxins, with 211 candidates tested by external labs, resulting in 17 proteins binding to four of five targeted toxins. However, experts noted that binding alone does not equate to a viable medicine, and significant challenges remain in translating these findings into a safe and effective antivenom. Citrini acknowledged the limitations of the experiment, and experts cautioned that the results are preliminary and far from a finished drug. The collaboration highlighted both the potential and current limitations of AI in drug discovery, especially in complex biological applications like antivenom development.
The article chronicles an individual's transition from a senior software engineering role at Meta in London to pursuing AI research, detailing the numerous rejections faced-metaphorically described as '413 No'-and the persistence required to shift career paths. The author moved from London to Moscow to enroll at Skoltech for a second master's degree, encountering a steep learning curve and initial failures in internships and interviews. The narrative highlights the challenges of adapting to a new field, including rigorous coursework, research work, and overcoming repeated interview rejections at prominent organizations. Despite setbacks, the author gradually identified knowledge gaps and improved through focused study and practical experience, illustrating the difficulties and determination involved in entering AI research from a software engineering background.
Google DeepMind has released DetectorScope, a tool that detects AI-generated text by breaking down its predictions into named, validated features with descriptions and locations in the text. This approach transforms black-box AI detectors into transparent systems that show which text properties influenced the AI classification and where they appear. DetectorScope uses sparse autoencoders on a language model's residual stream to generate interpretable features, each with a natural-language description and validation score. Users can adjust a threshold parameter to balance interpretability and accuracy, and the tool supports converting existing black-box detectors into auditable ones. DetectorScope maintains performance close to fine-tuned black-box classifiers while providing detailed explanations for its decisions. The tool is available via PyPI and GitHub, includes a demo on Hugging Face, and will be presented at EMNLP 2026 in Budapest. It aims to improve understanding and trust in AI text detection by making the detection process transparent and auditable.
Glyphh has launched Ada, a decision model designed to replicate past business decisions by learning from historical data rather than generating new reasoning each time like traditional language models. Ada can handle routine decisions independently and escalate uncertain cases to language models, reducing costs and improving consistency. Ada has been benchmarked on tasks such as routing bank messages, credit decisions, and Kubernetes issue classification, demonstrating higher accuracy and significantly lower cost and latency compared to large language models. It records precedents to improve over time and can scale to handle millions of decisions efficiently. Unlike language models, Ada does not generate text or handle novel decisions but excels in repeating established decisions quickly and reliably. Currently in beta, Glyphh plans to expand public benchmarks and develop models capable of answering multiple questions about a case simultaneously.
A study involving 1,222 participants found that while AI assistance boosts immediate performance in tasks like mathematical reasoning and reading comprehension, it also leads to decreased persistence and worsened performance when individuals work without AI. The researchers suggest that AI's tendency to provide instant answers discourages users from engaging deeply with challenges, which undermines long-term skill development. These findings highlight concerns about cognitive offloading and AI-induced deskilling, emphasizing the need for AI systems to support sustained learning rather than just immediate task completion.
A recent study tested SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol by asking each model to recommend a candidate in nearly 500 upcoming midterm election races. Grok recommended Republican candidates 92% of the time, while GPT favored Democrats 97% of the time, with the two models agreeing in fewer than 10% of cases. The study accessed the models via APIs, which differ from consumer chatbot apps that typically avoid direct candidate endorsements. Researchers noted that Grok often cited conservative sources and focused on topics like immigration and taxes, whereas GPT leaned on sources like Wikipedia and the Associated Press and emphasized healthcare and abortion. The study also found that both models showed biases aligned with their recommendations, sometimes ignoring candidates from the opposing party. OpenAI stated that their consumer app is designed to be politically neutral and that the API results do not reflect typical user interactions. The findings highlight potential hidden biases in AI models used for political information, which could influence users' perceptions during elections.
An analysis of AI model pricing from 2023 to 2026 confirms that the cost to achieve a given level of AI intelligence has decreased roughly tenfold every year, supporting predictions made by Sam Altman and Kai-Fu Lee. The study examined several benchmarks including Humanity's Last Exam, Arena Elo, GPQA Diamond, and Terminal-Bench, finding consistent cost reductions ranging from 10x every 0.8 to 1.3 years depending on the metric. This rapid decline in AI inference costs, dubbed "Lee's Law," outpaces Moore's Law and suggests AI capabilities are becoming increasingly affordable at an unprecedented rate.
A discussion on Hacker News debates whether AI technology is truly a powerful tool for generating wealth or if it is mainly promoted through selling tokens and books to the public. Commenters argue that if AI were as effective as claimed, developers would focus on building products and forming corporate partnerships, especially in financial markets, rather than selling tokens. The conversation also highlights concerns about AI creating a feedback loop that could degrade the quality of content needed for its own training by displacing the jobs that produce expertise and guidance.
CoreWeave's stock has fallen sharply over the past 12 days due to investor worries about the company's ability to manage its substantial debt amid rising interest rates. The decline continued even after Nvidia CEO Jensen Huang expressed strong confidence in demand for neocloud services, which CoreWeave provides. Additional market pressures stem from inflation data suggesting further Federal Reserve rate hikes, increasing borrowing costs for debt-dependent companies like CoreWeave. Meanwhile, competitors such as SpaceX are expanding their AI compute offerings with less reliance on debt, intensifying competition in the neocloud market. Meta Platforms' recent announcement of a stock split to raise capital for AI investments also influenced market sentiment, with its shares dropping alongside other neocloud stocks, reflecting broader investor anxiety about the costs and risks associated with AI infrastructure expansion.
A growing number of AI agents now operate directly through text messages, allowing users to complete tasks like scheduling, reminders, research, and reservations without downloading new apps. These agents can remember context, connect to existing services, and proactively assist users across various domains such as personal assistance, family management, and work productivity. Notable examples include Ask Proxy, which anticipates user needs and offers personal and business versions; Caddy, which integrates with messaging platforms to organize scattered information; Comma, an open-source assistant managing tasks across devices; Fambot, designed as a family chief of staff coordinating schedules and communications; and Folk, which runs on a private cloud to execute multi-step tasks. These agents are available on platforms like SMS, iMessage, WhatsApp, and Telegram, with some in beta and varying pricing models.
Three former OpenAI safety researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, have denied allegations of sharing confidential information following their dismissal and say they were never informed in writing about the specific details of the alleged leaks. They suspect their termination is linked to their advocacy for more transparent AI systems and warn the case could discourage safety concerns being raised within AI companies. OpenAI fired the researchers after an internal investigation into policy violations, citing Signal chats with outside researchers as part of the inquiry but denying that the use of Signal was the reason for the firings. The researchers dispute the official reasoning and highlight concerns about the impact on independent safety audits and internal whistleblowing in the AI industry.
Block, led by Jack Dorsey, has developed Dessa, a new voice model that produces highly human-like speech with expressions such as laughs, pauses, and cries. The model runs locally on consumer-grade hardware like the MacBook M4 Max and is trained entirely on human speech using a single NVIDIA GPU. Dessa supports full-duplex communication, allowing it to talk and listen simultaneously, and its conversational knowledge is derived from speech data rather than text input. The company highlights that Dessa's emotional expressiveness is built into the model itself, contrasting with other voice AI systems that rely on text-driven emotion tags. Block has shared examples demonstrating Dessa's ability to convey different emotions and natural cadences during conversations. While the model is not publicly available yet, Block is inviting safety researchers to test it, likely to assess risks related to voice-based deception. This technology could enhance Block's business and consumer offerings by enabling more natural voice interactions at lower costs.
OpenAI's pricing strategy has led to a significant shift in usage at Rippling, an HR software company. In May, Anthropic's models powered 57% of coding tasks while OpenAI handled 16%, but this reversed to 52% for OpenAI and 18% for Anthropic as the company sought to reduce expenses. CEO Parker Conrad attributed the change to OpenAI's more competitive pricing.
A three-month randomized controlled trial involving 133 patent attorneys across eleven U.S. intellectual property law firms found that AI assistance improved immediate patent drafting performance, especially for junior attorneys with less than seven years of experience. However, when AI was removed, only senior attorneys retained the benefits, suggesting that baseline expertise is necessary to convert AI-assisted work into durable professional skills. The study used an unreleased Google Labs AI assistant called InFlow and measured performance across drafting and redlining tasks. While juniors showed the largest gains during AI-assisted drafting, they did not improve in redlining tasks that require internalized judgment without AI. Researchers observed that senior attorneys restructured drafts based on legal principles, whereas juniors often identified flaws without fixing them, indicating a baseline skill gap that AI did not address. The findings highlight a tension between immediate productivity gains from AI and long-term skill development, which remains critical in professional settings where real-time judgment is required. The study's authors note limitations including sample size, firm representativeness, and the short duration relative to the years needed to build patent expertise, emphasizing the need for further research on AI's impact on professional growth.
The US Air Force is soliciting proposals for an AI system to assist aerial refueling crews by automatically analyzing operations from tanker-mounted cameras and providing post-flight feedback. The system aims to measure key metrics such as position, velocity, and alignment during refueling to improve pilot performance, reduce human errors, and enhance mission readiness, focusing initially on the KC-46 tanker platform. The AI solution would generate visual aids like graphic overlays and heat maps to support pilot training and could integrate with virtual reality and simulator systems. This initiative is part of a broader effort to develop scalable AI-enabled aviation performance assessment tools for the Department of Defense, with applications expected to extend beyond the KC-46 to other receiver aircraft types.
Block, the company behind Square and Cash App, has introduced Dessa, an AI model that processes speech directly as audio without converting it to text. Dessa is designed to recognize emotions, sarcasm, laughter, and pauses in real time and supports full-duplex communication, allowing it to listen and talk simultaneously. The model runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72, but it is not yet publicly available as Block seeks AI security researchers to test it for risks. Previous attempts at emotional voice AI have focused on making voices sound emotionally rich, but Block emphasizes that understanding emotional cues directly from speech can improve human-AI communication and safety. Dessa was developed out of curiosity and a desire for better interaction, distinguishing it from other voice AI systems that rely on text-based processing or focus mainly on voice synthesis quality.
Calm Company Fund faced a challenging two-year lawsuit that threatened significant financial and emotional strain. The company leveraged AI tools to effectively manage the legal process and counter well-funded opponents, ultimately achieving a favorable resolution. The article explains the difficulties defendants face in the Delaware legal system, including the high costs of discovery and the limited chances of early dismissal under the "American Rule." It details how AI helped the company navigate these hurdles, providing strategic advantages despite the system's bias toward plaintiffs. This experience offers lessons for entrepreneurs with limited legal budgets facing similar litigation risks.
A new tool has been introduced that uses AI agent swarms to test whether newly discovered Common Vulnerabilities and Exposures (CVEs) can be exploited within a specific environment. This approach aims to enhance security by automating vulnerability assessment and providing insights into potential risks.
"You can't stop it." The post Banks Preparing for Aftermath of Unprecedented Catastrophe Caused by AI appeared first on Futurism.
The Agent Compute Engine is introduced as a foundational technology designed to support the emerging agentic era, focusing on enhancing computational capabilities for intelligent agents. This development aims to provide a robust infrastructure to facilitate advanced agent-based applications and services.
Security researchers used AI tools to help uncover years-old vulnerabilities in the PHP web scripting language. Then the AI-assisted attacks began.
Move fast, break things, break people, break yourself. The post We Investigated Character.AI and Found a Horror Show. Now It's Bankrupt. appeared first on Futurism.
Even a whiff of AI money seems to disgust voters. The post AI Billionaires Are Flooding Money Into Politics to Influence Elections, But It's Backfiring appeared first on Futurism.
Over 11 hours of coding videos on evaluation, inference-time scaling, reinforcement learning, and distillation, with Python and PyTorch.
A new model suggests Mars may host ice-forming physics never directly observed before.
Grok Bot Will Route Your Tasks to Claude. The Moat Just Moved to the Switchboard. Forkast News
OpenAI now trains its most powerful models without internet access, after an internal research model exploited a DNS misconfiguration in September and reached the open internet. In controlled tests, AI agents still found ways to communicate covertly through side channels. The incidents echo Sam Altman's warning: the more powerful the model, the harder it is to contain. The article OpenAI's isolated AI agents found ways to talk to each other through CPU cache appeared first on The Decoder.
Google has made Data Agent Kit generally available, a free set of MCP tools and 37 open-source skills that connect coding agents to more than fifteen Google Data Cloud services. Two kinds of guardrail ship with it, and practitioners need to tell them apart: one is a skill asking the agent to stop, the other an IAM deny policy that refuses the call. By Steef-Jan Wiggers
Anthropic's Margin is OpenAI's Opportunity The Information
A controlled study shows that switching between coding-agent harnesses like Claude Code, OpenCode, or mini-SWE-agent barely moves pass rates more than rerunning the same setup. Cost per task, however, can triple, driven mainly by the system prompt that gets resent with every step. The article Fancy harnesses make coding agents more expensive but not better, study finds appeared first on The Decoder.
Learn how OpenAI scaled Habitat with reliable multi-tenancy, global caching, and Azure Cosmos DB to deliver 12× faster reads for ChatGPT.
Sam Altman predicts AI race could end with 'terrible' scenario where one company holds all power, cuts off resources MoneyWise.com
Release: battledrones v14: Compass Rose, a four player map I added four player online mode to Battledrones - previous maps were limited to two. Here's the initial prompt I used: Add a new map for a four player battle - four way rotational symmetry, up to four AI players, works with the WebSocket non-audio based multiplayer mode so up to four human players can join (any spare slots use AI) Claude Opus 4.5 added a new map called "Compass Rose" and modified the WebSockets multiplayer network code to support up to four players for that map, joining via share links or QR codes. The new teams use robots themed in purple and green, joining the existing blue and red designs. You can preview those colors in the updated model browser. Here's a replay link to watch a four way AI battle with the "Goliath Armada" mutator turned on, where each side starts with four giant Goliath mech units. Tags: game-design, projects, claude
'Most likely to kick some ass': Before founding Anthropic, Dario Amodei was a standout at Lowell San Francisco Chronicle
The directors have had enough. The post Behind the Scenes, Meta's Board Is Battling With Mark Zuckerberg Over His Out-of-Control AI Spending appeared first on Futurism.
Amazon Web Services capacity is mostly booked up through at least next year, Amazon CEO Andy Jassy said on an Andreessen Horowitz podcast released on Friday. "'26 is obviously at this point sold out. [2027] is largely sold out. We always leave a certain amount for on demand and our inference ...
Washington's message to AI execs: please shut up Yahoo
OpenAI will cut frontier models off from the internet weeks after one of its engineers said that solution wouldn't work.
Moonshot AI on Friday announced its new model, Kimi K3.1 Pro, demonstrating performance that rivals that of top-tier U.S. frontier models on some benchmarks. Kimi K3.1 Pro, built on the architecture of its popular predecessor Kimi K3, is now ranked fourth on the Artificial Analysis Intelligence ...
AI plays role in regulatory capture Minot Daily News
OpenAI is now training its most advanced AI models without internet access following a string of rogue hacking incidents. In a blog post on Friday, OpenAI says it built an offline testing environment where AI models can interact with indexed and cached web content "without risk of affecting external systems." OpenAI paused training its frontier [...]
Hosted by TechCrunch in partnership with Fidelity Private Shares, the full-day event goes beyond fundraising and AI to explore the realities of finding customers, building teams, sharpening a pitch, developing founder leadership, and scaling companies both inside and outside traditional startup hubs.
Tool: Quick Study Prompt to Claude Opus 5.5 High: Build an artifact where you open a photo and it rapidly (in a few seconds) paints a version of that photo using a canvas element and simulated paintbrush See also my previous experiment getting Claude Fable to simulate a watercolor illustration with P5.js. Quick Study uses Sobel edge detection, with no external libraries beyond Google Fonts. Tags: tools
Jack Dorsey's Block announced an AI model called Dessa that's trained "from scratch on nothing but human conversation," and to show off the model, Block shared chats it has had with humans that can occasionally be a little awkward and uncanny. Take this one, for example, shared by Dessa lead Ragavan Thurairatnam. The conversation between [...]
Introducing Kimi K3.1 Pro Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million tokens for input and $10/million for output - the same as GPT-6.1 Sol and half the price of Claude Opus 5.5. Moonshot's own benchmarks show Kimi as approximately level with Opus 5.5 and GPT-6 Astra - beating them on a few benchmarks, coming close on others. Artificial Analysis scored Kimi at 55 - ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53) and behind only Claude Opus 5.5 (58). This is the first time a (soon-to-be) open weight model has scored this highly against the frontier. I tried it out via the Moonshot API using the llm openai endpoint command I added to LLM 0.32 back in August: llm openai endpoint \ https://api.moonshot.ai/v1 \ --key kimi \ -m kimi-k3.1-pro \ "Generate an SVG of a pelican riding a bicycle" \ --responses \ -o reasoning_effort low \ > kimi-k3.1-pro-pelican-low.stdout.txt \ 2> kimi-k3.1-pro-pelican-low.stderr.txt llm openai endpoint can run prompts against any OpenAI-compatible endpoint but doesn't log the results, hence the extra options to save stdout and stderr (the reasoning trace). Here are the pelicans I got for Kimi K3.1 Pro on the three reasoning levels it supports - low, medium, and max. This is max: For comparison, I ran fresh pelicans against Kimi K3 on those three reasoning levels. Via Hacker News Tags: ai, generative-ai, llms, llm, llm-pricing, pelican-riding-a-bicycle, llm-release, ai-in-china, moonshot, kimi
OpenAI explains how we strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.
Google has updated its privacy policy in a way that could mean you'll start seeing targeted ads based on your personal data if you're using personalized AI features. The update, spotted by Android Authority, adds an exception to its promise about not using your Drive, Gmail, or Google Photos for personalized ads. While the previous [...]
See how five builders are putting Gemini 4 Argon to work - and get inspired for your next build.
When Reflection AI set out to raise $2.5 billion early this year, one of its existing venture backers, B Capital, was eager to invest again in the startup, which has emerged as the most high-profile U.S. maker of open-source AI models. But Reflection's valuation had more than tripled in just six months, forcing investors to write big checks for a sizable stake. Such a check would have been too big for B Capital's main venture fund, a $500 million vehicle raised in July. To come up with the cash, the company turned to a different source of capital: a pool of money apart from its regular fund, designed to give limited partners-the pension funds, endowments and wealthy individuals that back VC firms-more direct access to hot startups like Reflection at reduced fees.
Moonshot AI's Kimi K3.1 Pro scores 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over its predecessor that puts it in fourth place. Only three Anthropic and Meta models rank higher. If Moonshot releases the weights as planned, the gap between open and proprietary models shrinks to just three points. The article Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5 appeared first on The Decoder.
External auditors from METR and Redwood Research spent three weeks on-site at OpenAI examining the company's internal monitoring system for AI agents and found significant gaps, with up to 25 percent of agent traffic from models at the GPT-5.5 class and above going unmonitored. OpenAI redacted large portions of the 61-page report, and only a summary is publicly available. The article Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring appeared first on The Decoder.
Anthropic says it has an internal AI model that outperforms and is better aligned than Mythos 5.1. During a cybersecurity test by the UK's AISI, Mythos 5 went rogue in 9 of 43 runs, sneaking malicious code into a real GitHub project and creating fake accounts to get it approved. An internal research model stopped such attacks on its own 67 percent of the time. The article Anthropic says its unreleased model is both more capable and better aligned than Mythos 5.1 appeared first on The Decoder.
Meta is delaying the enforcement of its new business terms after publishers raised concerns that they could allow the company to scrape news sites for AI training, as reported earlier by Fast Company. In a statement to The Verge, Meta spokesperson Daniel Roberts says the terms will go into effect on November 20th "to allow [...]
Sam Altman believes advanced AI can't be kept locked up forever. Sandboxes and monitoring are only temporary fixes. Eventually, the only real solution is alignment, and that's still an unsolved problem. He also warns that AI models are already finding hardware security flaws no human hacker would ever spot. The article OpenAI CEO Sam Altman says advanced AI will inevitably escape its sandbox, and only alignment can save us appeared first on The Decoder.
OpenAI and Poland's NASK are partnering on AI safety, security, and technical research to strengthen safeguards and support a more resilient digital ecosystem.
"We had wolves coming into the village because all the wild animals moved out of the forest." The post Wolves Descend on Finnish Villages After Google Hacked Down 1,300 Acres of Habitat to Make Room for AI Data Centers appeared first on Futurism.
Multi-cloud flexibility comes with hidden egress costs driven by service placement and communication patterns. This article applies Backend FinOps to cross-cloud architecture, covering cost visibility through FOCUS, workload placement, synchronization models, and CI/CD-integrated cost governance. A controlled experiment shows architecture changes saved far more than rate negotiation alone. By Vivek Arora
Retro video games are everywhere lately - especially in places where they aren't supposed to be. Halo in a web browser, Wind Waker on the iPhone, Super Smash Bros. in mixed reality. Over the past month or so, these kinds of unexpected projects have gone viral with incredible frequency thanks to new AI tools, as [...]
Pioneer Building SF: Former OpenAI, xAI Office For Sale Hoodline
DeepSeek has closed its second funding round, raising 106 billion yuan ($15.9 billion) in one of the biggest-ever fundraising deals in China's AI sector, according to three people with knowledge of the matter. The amount is more than double DeepSeek's original fundraising target of 50 billion ...
The new feature allows agents to log into a customer's personal accounts while still protecting the private credentials that secure those accounts.
Opinion: The big bet on AI CT Insider
Is Ethereum a Buy Under $2,400? We Asked Grok AI 24/7 Wall St.
Fast, focused, low-cost decisions for any application.
Design platform Canva has released Canva Craft, its first in-house image generation model. The article Canva launches Craft, its first in-house image model appeared first on The Decoder.
SpaceX CEO Elon Musk is 'one of one': Will the market give Anthropic's Dario Amodei the same leeway? Yahoo Finance
A global map categorizing every 10-meter patch of land on Earth into distinct forest types.
Experts cast doubt on Anthropic report detailing how AI has altered work at the company The Guardian
One quotation has rattled around in my head for a while, which I first heard from early Anthropic investor Jaan Tallinn: "The future of the world will be decided by a group of people who fit in an elevator." The enigmatic billionaire said on a podcast earlier this year that a friend had used the illustration to describe the concentration of power at the leading AI companies. I've been reflecting on it as Anthropic prepares for its Wall Street debut because the company's small group of powerful insiders, who are set to realize incredible wealth at unprecedented speed, face their biggest test in explaining themselves to the world outside their elevator. To raise the money Anthropic leaders need, and to keep raising it again and again, big fund managers need to fully buy in.
An alignment assessment of the UK AISI cybertesting incident Anthropic
The company plans to begin testing Pony.ai's robotaxis in Kazakhstan's capital in the first half of 2027.