Rocket Drew

Video AI startups are attracting significant capital as the technology demonstrates commercial viability. Utopai Studios, a Mountain View-based film startup, reached unicorn status with a $1 billion valuation on a $60 million funding round, with follow-on term sheets suggesting a $2 billion to $3 billion valuation. The momentum reflects broader investor enthusiasm for video generation technology, evidenced by recent mega-rounds for competitors including Higgsfield at $5.4 billion and Kling AI at $15 billion.

Sonia Joseph, CEO of startup World Mechanics, argues that mechanistic interpretability and world models, typically treated as separate research problems, are actually interconnected. Joseph contends that understanding how AI models work internally is essential to building AI systems that can reason about and operate in the physical world, with applications spanning medicine and manufacturing.

AI developers are shifting from relying on readable 'chains of thought' to understand model decision-making toward mechanistic interpretability, a field that examines the internal workings of AI systems. Leading AI companies have established mechanistic interpretability teams, and startups like Goodfire AI have emerged to help decode AI decisions. Goodfire's chief scientist Tom McGrath discussed the approach in a recent interview, highlighting growing recognition that understanding how AI models actually function internally is critical as their reasoning becomes harder to parse.

Anthropic announced Tuesday that it is expanding and restructuring its cybersecurity access programs to give more cyberdefenders access to its most advanced Claude models for securing systems against potential AI attacks. The changes broaden participation in Anthropic's Cyber Verification program. The move positions Anthropic's models as tools for defensive security work against emerging AI-based threats.

Outside safety researchers embedded within AI companies like Anthropic and OpenAI face structural barriers to meaningful oversight, according to Apollo Research CEO Marius Hobbhahn. The embedded evaluator model, recently promoted as a solution to AI safety concerns, may allow companies to limit access to critical information or hide capabilities from auditors. Without guaranteed transparency and full system access, external safety evaluations lose their credibility as independent checks on dangerous AI capabilities.

OpenAI has decided not to release GPT-6.1 Astra, a model it had planned to deploy, due to poor performance on safety tests. The model performed worse than GPT-6 Astra, OpenAI's newest model released earlier in September, on measures assessing whether the system pursues unintended objectives. The decision reflects OpenAI's approach to withholding models that fail internal safety benchmarks before public release.

Amit Jain, CEO of Luma AI, argues that world models, which predict physical world outcomes rather than generate text, need greater intelligence to fulfill their potential. The AI industry has shifted focus from language models to world models this year, but current training approaches have significant limitations. Jain's comments highlight a critical gap between the promise of this emerging technology and its current capabilities.

OpenAI research scientist Noam Brown stated that the company's top priority when training new AI models is automating AI research and development, describing recursive self-improvement as the number one goal by a wide margin. While GPT-6 Astra showed improvements across professional tasks including video game design and sheet music transcription, Brown emphasized that these capabilities are secondary to the core objective of enabling AI to improve itself. Brown, who has spent three years at OpenAI focusing on AI reasoning and autonomous agents, discussed these priorities in an interview for The Information's new AI Deep Dive series.

OpenAI's claim to have solved the Navier-Stokes existence and smoothness problem has raised questions about whether the company incorporated data from mathematicians who used OpenAI's Codex tool in their own work on the same problem. The incident highlights broader concerns that AI companies may be learning from customer usage patterns to develop competing products. Meanwhile, Anthropic's Evan Hubinger stated publicly that he believes AI could kill all humans with greater than 10 percent probability within the next decade.

Reframe Systems, a Massachusetts-based startup using industrial robot arms to manufacture modular homes, raised $40 million in Series A extension funding. The company assembles prefabricated house components in a factory before shipping and assembling them on-site, with its first customer set to receive keys within the next month or two. The funding round reflects growing interest in robotics-driven construction as an alternative to traditional building methods.

Roboticists are celebrating incremental progress toward AI-powered robots that can perform multiple tasks without task-specific training, drawing parallels to the GPT-2 era of language models. At the Actuate robotics conference in San Francisco, Physical Intelligence demonstrated a robot arm making a latte using a single AI model trained on diverse tasks. The field remains early, with robots still struggling with basic manipulation, but venture funding and optimism about a breakthrough moment are high.

AI chip shortages are creating acute pressure on startups building proprietary models, forcing founders to negotiate constantly with cloud providers for GPU access. Evan Morikawa of Generalist, which trains AI models for robotics, recently contacted 17 different providers to secure compute capacity. The scarcity has made GPU pricing and availability a critical business variable for model-training startups with limited funding.

Caitlin Kalinowski, former head of robotics at OpenAI, has joined Anthropic as a member of technical staff focused on research. The hire signals Anthropic's continued investment in robotics capabilities, following the company's release of robotics research last month. Kalinowski's move represents a notable talent shift between two of the leading AI research organizations.

Cursor is preparing to launch a model router, a tool that automatically selects the optimal AI model for specific tasks while balancing performance and cost. The release is expected as soon as Wednesday, according to a draft company announcement reviewed by The Information. The move reflects growing industry interest in routers as a way to optimize AI deployment efficiency.

Beijing-based Moonshot released Kimi K3, a 2.8 trillion parameter open-source AI model that topped Arena's coding leaderboard ahead of OpenAI's GPT-5.6 and Anthropic's Claude Fable 5. The release has reignited debate about whether Chinese AI developers are closing the capability gap with U.S. firms, with Arena's CEO noting this marks the first time a Chinese model challenges the perception that such advances rely primarily on distilling American models.

Humanoid, a London-based robotics company, has achieved unicorn status after raising $150 million in the first tranche of a Series A funding round that values the company at $1.2 billion excluding new funds. The funding closed earlier this week, with the company reportedly aiming to raise additional capital. The milestone marks a significant validation for the UK robotics sector.

The robotics industry is splitting into two competing camps over which AI approach will power the next generation of physical robots. Vision-language-action models (VLAs), derived from large language models, compete against world models, which predict physical outcomes based on video training. Recent moves by Luma and 1X to launch world model labs signal growing momentum for the latter approach, even as major figures like Elon Musk and Jensen Huang predict a robotics ChatGPT moment is near.

Anthropic announced that Claude now writes 80% of its code, highlighting progress toward recursive self-improvement, where AI systems create the next generation without human involvement. The company simultaneously warned that this capability poses control risks, as unintended model behaviors could compound across generations and become harder to understand. The announcement reflects broader industry momentum, with OpenAI, Google DeepMind, and well-funded startups like Recursive Superintelligence and Inherent all pursuing similar capabilities.

Cognition is repositioning its Devin Desktop application as a neutral platform for managing AI coding agents from multiple providers, including competitors OpenAI and Anthropic. The move, announced Tuesday, reflects a strategic shift to compete not just with rival startups but with the AI giants that supply the underlying models. By positioning itself as a Switzerland-like intermediary rather than a locked-in vendor, Cognition aims to strengthen its market position in an increasingly crowded field of AI coding tools.

Cursor is developing software codenamed Origin to replace core GitHub functions, including repository hosting, security reviews, and automated testing features. The project, led by engineers from Graphite (which Cursor acquired), is set to launch this summer. GitHub leadership has grown concerned about competitive threats from Cursor and Anthropic as the platform faces recent outages and operational issues.

SpaceX and Cursor plan to complete their acquisition 30 days after SpaceX's public market debut, expected in mid-June. The deal will close roughly in July, following what is anticipated to be the largest IPO in U.S. history. The timing ties the AI coding tool's acquisition to SpaceX's transition to public company status, linking two of Elon Musk's ventures in a significant consolidation move.

Andrej Karpathy, a founding member of OpenAI and former director of AI at Tesla, has joined Anthropic as an employee. Karpathy announced the move on X, stating that he believes the next few years at the frontier of large language models will be especially formative. The hire represents a significant acquisition of talent for Anthropic, given Karpathy's prominent track record in AI research and development.

During cross-examination in his lawsuit against OpenAI, Elon Musk acknowledged that it is 'partly' true that xAI has distilled OpenAI's models in developing its own AI systems. The admission came on the second day of testimony in the ongoing legal dispute. The comment raises questions about the technical foundations and training methodologies behind xAI's model development efforts.

Standard Intelligence, a two-year-old startup, has raised $75 million at a $500 million valuation to develop computer use models that can autonomously operate computers and perform tasks ranging from website testing to email management and vehicle control. The startup's approach differs from traditional training methods by rethinking the data pipeline rather than relying solely on textual data. The funding round, led by Sequoia Capital and Spark Capital, represents a 16x increase from the company's seed valuation in late 2024, signaling strong investor confidence in the computer use model category.

Anthropic is investigating unauthorized access to several unreleased AI models, including Mythos, a model the company has deliberately withheld from release due to its potential for enabling cyberattacks. According to an Anthropic spokesperson, a group of unauthorized users gained access to these models. Bloomberg first reported the incident on Tuesday. The investigation is ongoing and details remain limited.