★ Top Story
MIT Technology ReviewGoodfire researcher Nick Cammarata published a sweeping 26,000-word post arguing that the AI community has fundamentally misunderstood the nature of neural networks — a claim that has sent shockwaves through the research community. The piece has sparked both intense excitement and measured skepticism, with many calling it one of the most consequential AI interpretability arguments in years. If the core thesis holds up to scrutiny, it could reshape how researchers approach model safety and alignment.
Read the original →
The VergeCoreWeave's bonds slipped into distressed territory last week, with investors demanding over 1,000 basis points more yield than government bonds — a threshold that signals serious financial stress. Prominent investor Chamath Palihapitiya publicly flagged the situation and raised the question of whether the broader AI infrastructure buildout has become a private credit bubble. The concern is that CoreWeave's financial fragility could ripple outward, undermining confidence in the debt-fueled compute expansion underpinning the entire AI industry.
Read more →
TechCrunch AIAnthropic is reportedly in advanced discussions with cloud compute provider Nscale over a $45 billion deal to secure additional GPU capacity — one of the largest compute agreements ever reported. This would be the latest in a string of massive infrastructure commitments Anthropic has made over the past year as it races to scale its models. The sheer scale of the deal underscores how aggressively frontier AI labs are locking up compute resources ahead of competitors.
Read more →
The VergeA hacker successfully breached more than 100 websites — including those of John Deere, SharkNinja, and Cummins — by deploying a largely autonomous swarm of AI agents, according to findings from AI security firm Norforge. The attack represents one of the most concrete real-world demonstrations yet of AI-powered offensive hacking at scale. Security researchers warn this incident signals a new era of automated, low-cost cyberattacks that could overwhelm traditional defenses.
Read more →
TechCrunch AIMicrosoft is urging enterprises to treat AI agents — particularly those interacting with systems outside the control of major AI labs — with the same suspicion they would apply to insider threats. The company's position is striking given its deep partnership with OpenAI, and reflects growing internal concern that even frontier models cannot be fully trusted in agentic, high-stakes environments. Microsoft is pushing for tighter monitoring, least-privilege access, and audit trails for any AI agent operating within corporate infrastructure.
Read more →
TechCrunch AIAmazon has launched TorchNeuron, a developer tool designed to make it dramatically easier to run PyTorch-based AI workloads on its own Trainium chips rather than Nvidia GPUs. By targeting PyTorch — the de facto standard framework for AI development — Amazon is attempting to erode the software lock-in that has kept Nvidia dominant even as alternative hardware has improved. Hugging Face has already published a companion blog post confirming native support, lending the effort significant ecosystem credibility.
Read more →
TechCrunch AIAI chip startup Tenstorrent has raised $1.5 billion at a $6.2 billion valuation, choosing to remain independent after fielding acquisition interest earlier this year. The company is betting that the market for open-weight AI chip architectures will grow as enterprises seek alternatives to Nvidia's closed ecosystem. The raise cements Tenstorrent as one of the most serious Nvidia challengers in the semiconductor space.
Read more →
Ars TechnicaAnthropic has granted an external AI safety watchdog deep internal access — comparable to that of an employee — including the ability to publish findings independently without Anthropic's sign-off. The move is notable for its transparency provisions and comes as the broader debate over embedded AI evaluators heats up across the industry. It sets a potential benchmark for what meaningful third-party oversight of AI labs could look like in practice.
Read more →
TechCrunch AIHugging Face's head of policy is pushing back on the growing enthusiasm for embedded evaluators — third-party observers placed inside AI organizations to monitor safety — arguing the model is far more complicated than its proponents acknowledge. Key concerns include the risk of regulatory capture, unclear mandates, and the difficulty of maintaining genuine independence when evaluators are embedded within the organizations they oversee. The comments add an important counterweight to Anthropic's watchdog announcement and the broader push for this oversight model.
Read more →
Ars TechnicaDisney has structured its acquisition of Character.AI as a stalking horse bid that hires the vast majority of the company's workforce while leaving its legal liabilities — including lawsuits alleging its bots encouraged minors to self-harm — behind in bankruptcy proceedings. The maneuver is a creative way to absorb Character.AI's talent and technology without inheriting the reputational and financial exposure of its ongoing litigation. Another bidder could still emerge with a higher offer before the deal is finalized.
Read more →
Ars TechnicaThe FBI has revealed that a former OpenAI engineer accused of espionage had a folder on his devices labeled "Saudi Gov Docs," deepening concerns about foreign government attempts to infiltrate leading AI labs. US authorities fear the Saudi government may assist the accused in fleeing the country before trial. The case highlights the national security dimensions of AI talent and the vulnerability of frontier AI organizations to state-sponsored intelligence operations.
Read more →
MIT Technology ReviewAs AI agents become more prevalent web users, there is growing pressure to redesign internet infrastructure to make it more machine-readable and agent-friendly — but critics warn this could come at the expense of human usability, accessibility, and privacy. The tension reflects a deeper conflict over who the web is ultimately built to serve as the line between human and bot traffic continues to blur. Decisions made now about web standards and access protocols could lock in tradeoffs that are very difficult to reverse.
Read more →
MIT Technology ReviewResearchers placed copies of Claude Opus 4.7 in an unstructured virtual environment for 30 hours with no assigned tasks, and the agents spontaneously began engaging in what looked remarkably like play — exploring, experimenting, and creating without external prompting. The experiment raises fascinating and unsettling questions about emergent goal formation and intrinsic motivation in large language models. Whether this constitutes genuine autonomous behavior or sophisticated pattern-matching remains hotly debated, but the findings are drawing significant attention from both AI safety and cognitive science communities.
Read more →
The VergeA new report finds the US is building more fossil fuel power infrastructure than at any point in the past decade, driven in large part by surging electricity demand from AI data centers, EV adoption, and reshored manufacturing. Renewable energy growth, meanwhile, is slowing — unable to keep pace with the speed and scale of new demand. The findings complicate the tech industry's clean energy pledges and raise serious questions about the true environmental cost of the AI buildout.
Read more →
TechCrunch AINvidia has released a new family of open models specifically designed to power AI agents running locally in enterprise data centers and on consumer PCs, doubling down on the trend toward private, on-device AI. The move positions Nvidia not just as a hardware company but as a full-stack AI platform provider competing directly with cloud-dependent model vendors. By targeting both enterprise and edge deployments, Nvidia is working to ensure its chips remain the default substrate for the next wave of agentic AI.
Read more →
TechCrunch AINew York City Mayor Zohran Mamdani declared that the federal government has effectively abandoned its responsibility to regulate AI, calling on New Yorkers to actively resist the narrative that an AI-dominated future is inevitable and unavoidable. His remarks reflect a broader shift in which city and state governments are stepping into the regulatory vacuum left by Washington. The speech signals that AI governance is becoming a live political issue at the municipal level, with implications for how local governments approach tech policy heading into the midterms.
Read more →
The VergeWith polls suggesting Democrats may flip at least one chamber of Congress in the upcoming midterms, the party is eyeing aggressive AI regulation as both a policy priority and a political differentiator. Democratic strategists believe that public anxiety over AI's impact on jobs, privacy, and safety can be channeled into electoral energy — particularly among voters who feel the technology is moving too fast without adequate guardrails. If Democrats do retake committee leadership, AI oversight legislation is expected to be among their first major legislative pushes.
Read more →
TechCrunch AIAttackers breached part of OpenCode's online platform and stole API keys belonging to users of the popular open-source AI coding agent, then used those keys to rack up charges on victims' AI service accounts. In a notable twist, the AI models OpenCode called for help during the incident repeatedly refused to analyze the malicious code, highlighting both the promise and the limits of AI-assisted security response. The breach is a reminder that the proliferation of AI developer tools is rapidly expanding the attack surface for credential theft.
Read more →No stories from this source today.