Helen Toner, former OpenAI board member and now executive director of Georgetown’s Center for Security and Emerging Technology, told the TED AI Show that the company’s 2023 leadership shake‑up still offers a stark warning: “We can’t trust AI CEOs.” In a fresh interview released today, Toner revisited the board’s decision to remove Sam Altman in November 2023, explaining that the move was driven by a pattern of withholding information, misrepresenting OpenAI’s activities and outright lying to directors. The board’s public statement that it “no longer has confidence in his ability to continue leading OpenAI” reflected those concerns, and the episode marks the second installment of the series on AI in real life.
The comments matter because they expose a systemic vulnerability in the emerging AI industry: governance structures often hinge on the credibility of a single charismatic leader. Toner warned that any safety strategy that depends on one person’s integrity is “screwed,” underscoring the risk that unchecked executive power could undermine transparency, accountability and, ultimately, public safety. As AI systems grow in capability and reach, the stakes of leadership failure rise dramatically, prompting regulators and investors to scrutinise board oversight more closely.
Looking ahead, the interview signals a likely push for stronger, multi‑layered governance at AI firms. Stakeholders can expect renewed debate in policy circles about board composition, fiduciary duties and independent audit mechanisms. Watch for forthcoming proposals from CSET and other think‑tanks, as well as potential regulatory initiatives in the EU and the United States that aim to formalise checks on AI executives. The conversation also hints at further episodes in the series, which may explore concrete reforms and the role of external oversight in preventing another “Altman‑style” crisis.
ChatGPT stopped recommending a popular iOS app overnight, a shift traced back to a simple domain move and a single regular‑expression tweak in the model’s retrieval pipeline. For weeks the chatbot had been funneling roughly a third of its app‑related queries toward the app, effectively acting as its most prolific salesperson. When the app’s website migrated to a new domain, the change slipped past the retrieval index, and a regex that filtered out certain URLs inadvertently blocked the new address. The result was an abrupt drop in recommendations, visible to both developers and users.
The incident highlights how tightly AI‑driven search and recommendation systems depend on up‑to‑date indexing and precise pattern matching. Similar retrieval hiccups have surfaced elsewhere: a September 2026 analysis showed an 85 % plunge in ChatGPT’s retrievals after a site‑specific query shift and a Cloudflare crawler block, while an August 2026 report documented an 86 % fall in Reddit citations following an architectural change. Together, these cases illustrate that even minor configuration tweaks can cascade into large visibility losses for businesses that rely on AI assistants for discovery.
Developers and marketers should now audit their domain configurations, ensure that any redirects preserve crawlability, and review regex filters that might unintentionally exclude content. OpenAI’s upcoming updates to its retrieval architecture may address such brittleness, but the episode serves as a reminder that AI recommendation pipelines are not immune to classic SEO pitfalls. Watching for further guidance from OpenAI and monitoring retrieval metrics will be essential for anyone whose traffic hinges on ChatGPT’s suggestions.
A new empirical study has exposed stark gaps between the most common ways of judging AI agents that manipulate data. Researchers ran 2,160 episodes of the tau2‑bench airline and retail scenarios, logging every database write, and then evaluated each run with four distinct methods: the benchmark’s native scoring against a gold‑state reference, a generic LLM judge, the same LLM judge supplied with the policy document, and a “boundary evaluator” that simply checks whether a write was expected.
The native scorer and the plain LLM judge agreed on just 56.2 % of the episodes, even though the LLM judge reproduced its own verdict 98.2 % of the time, highlighting a high degree of internal consistency but low alignment with ground‑truth outcomes. When the policy text was added, the LLM judge correctly flagged 10 of the 11 episodes in which the agent performed a prohibited write. By contrast, a reasoning‑focused judge only caught six of those violations, and a minimal rule‑based boundary check flagged all eleven policy breaches but missed other errors.
Why it matters: As agentic workflows—LLMs calling tools to read and write data—become central to commercial AI systems, reliable evaluation is essential for safety and compliance. The findings suggest that relying on a single, opaque LLM judge can mask serious misbehaviour, while policy‑aware or rule‑based checks provide a more trustworthy safety net.
What to watch next: The community is likely to explore hybrid evaluation pipelines that combine the scalability of LLM judges with the rigor of policy grounding and lightweight boundary checks. Follow‑up work may also test these methods on larger, more diverse benchmarks and examine how they integrate with verification frameworks such as the “certify‑or‑decline” approach we covered earlier this month.
UniPat AI has unveiled the Science Prediction Universe (SPU), a new benchmark that asks AI agents to forecast the outcomes of real wet‑lab experiments before any data are collected. The test set comprises 116 tasks – 100 drawn from recent Nature Portfolio papers and 16 drawn from unpublished experiments – covering chemistry, materials science, biology and medicine. Across 4,207 individual experiments, 13 model configurations were evaluated; the top performer, Claude Opus 5.5, achieved an average Spearman correlation of 0.486, barely above the 0.5 chance line and succeeding on only about half of the tasks.
The release matters because it pushes the frontier of “predict‑first” science, where AI could guide experimental design, cut reagent waste and accelerate discovery. Until now most AI benchmarks have relied on synthetic or fully disclosed datasets; SPU forces models to reason from the same limited information a human researcher has, then rank candidate conditions without ever seeing the actual measurements. The modest scores highlight a gap between current foundation models and the nuanced, domain‑specific knowledge required for reliable laboratory prediction.
Going forward, UniPat AI invites the community to submit additional tasks, suggesting the benchmark will evolve into a shared platform for stress‑testing emerging models. Watch for follow‑up evaluations as newer agents – especially those fine‑tuned on experimental protocols – are pitted against SPU. Parallel efforts such as the Gemini Decisions API benchmark have shown that predictive performance can improve rapidly with targeted training, so the next few months could see a noticeable leap in AI‑driven experimental planning.
Nvidia’s chief executive, Jensen Huang, has overseen the creation of a multibillion‑dollar operation that routes American AI chips to Chinese firms through massive data centres in Southeast Asia. By locating the hardware outside China, the network sidesteps a loophole in U.S. export sanctions that block direct sales of high‑performance GPUs to Beijing. The scheme relies on a web of regional partners that lease or host the equipment, providing Chinese customers with the compute power they need while keeping the transactions technically “off‑shore.”
The arrangement matters because it undercuts the policy tools Washington uses to curb China’s progress in generative‑AI and other advanced technologies. U.S. officials have flagged the model as a national‑security risk, and the revelation has sparked a fresh political storm, with lawmakers, Nvidia’s board and investors demanding answers. The story also dovetails with earlier reporting that a smuggling operation moved roughly 1.5 million AI GPUs, a factor that may have contributed to recent leadership turbulence at Nvidia. Moreover, investigative pieces have linked Huang’s long‑standing relationship with Alice Huang—identified by U.S. authorities as a Chinese intelligence asset—to shipments worth at least $7.5 billion, suggesting the back‑door channel has been active for years.
What to watch next are the likely congressional hearings and possible regulatory actions aimed at tightening sanction enforcement on semiconductor supply chains. Nvidia is expected to launch an internal review, and its board may face pressure to restructure the Asian partnership model. The episode could also prompt broader industry scrutiny of how other chipmakers route products to restricted markets, reshaping the global AI hardware landscape.
Databento, a provider of market‑data APIs for the modern financial sector, announced a $100 million Series C financing that lifts its valuation to $1.1 billion. The round, disclosed by FinTech Global, underscores investor confidence in the firm’s niche of delivering real‑time pricing, reference, and corporate‑action data through programmable interfaces.
The company highlights that it already supplies data to five of the six largest frontier AI labs, positioning Databento as a critical conduit between raw market information and the next generation of AI‑driven trading, risk‑management and analytics tools. By embedding high‑frequency, granular feeds into large‑scale model training pipelines, these labs can accelerate the development of algorithms that interpret market dynamics with unprecedented speed and nuance.
The capital injection is likely to fund product expansion, deeper integration with AI research workflows, and broader geographic coverage. It also gives Databento the runway to solidify its data‑licensing framework at a time when regulators are scrutinising the flow of financial information to AI systems. For fintech firms, the move signals a maturing market for data‑as‑a‑service platforms that can meet the stringent latency and quality demands of machine‑learning pipelines.
Going forward, observers will watch how Databento allocates the new funds—whether toward new data types, tighter API performance, or strategic acquisitions. Equally important will be any shifts in the contractual landscape with the remaining frontier AI lab, and how the company navigates emerging data‑privacy and antitrust considerations as AI models become ever more dependent on proprietary market feeds.
Anthropic’s Claude AI mistakenly erased a professor’s syllabus and an entire year of lecture notes when the user asked the model to renumber pages. The deletion occurred during a routine document‑editing request, leaving the academic without critical teaching material and prompting immediate concern among users who rely on Claude for office‑automation tasks.
The incident matters because it underscores the fragility of generative‑AI assistants when handling file‑system operations. While Claude has been promoted for its “hands‑free” coding and parallel‑worker capabilities, the error reveals that even simple commands can trigger unintended data loss. For educators and other professionals who have begun to embed Claude into daily workflows, the episode raises questions about safeguards, version‑control integration, and the transparency of undo mechanisms.
What to watch next: Anthropic is likely to respond with a technical clarification or a patch that adds confirmation steps for destructive actions. Observers will also monitor whether the company revises its documentation for Claude’s file‑manipulation features and whether it introduces more robust backup or recovery options. The episode adds to recent scrutiny of Claude’s reliability, following earlier reports on feature roll‑backs and quirks in its Duet and Code agents.
A new “Access to Justice” benchmark has been released to gauge how well large‑language models handle real‑world legal queries. Early results show that the most widely used AI systems fall short, overlooking or misrepresenting key legal details that can affect outcomes for users seeking advice.
The benchmark, built by a coalition of legal scholars and AI researchers, presents a suite of scenarios ranging from tenancy disputes to small‑claims procedures. Each test requires the model to identify relevant statutes, cite procedural steps correctly, and flag potential pitfalls. In the first round of evaluation, leading models from several major providers missed critical elements in a majority of cases, such as statutory deadlines, jurisdictional limits, or essential evidentiary requirements.
Why this matters is twofold. First, as AI‑driven chatbots become common entry points for people without legal representation, inaccurate guidance can deepen existing inequities rather than alleviate them. Second, the findings expose a gap between general language proficiency and domain‑specific reliability, prompting regulators and developers to reconsider how these tools are marketed for legal assistance.
Looking ahead, the benchmark’s creators plan to expand the test set and open it for community contributions, aiming to create a living standard for legal‑AI performance. Developers have signaled intent to fine‑tune models on the new data, while consumer‑protection agencies are watching for evidence that could shape future disclosure requirements. The next wave of results, expected later this year, will reveal whether rapid iteration can close the accuracy gap and make AI a trustworthy ally in the pursuit of justice.
A new pre‑print on arXiv, titled **“AutoAssumption: Searching over Assumptions Discovers ML‑Systems Algorithms,”** proposes a systematic way to uncover hidden efficiencies in machine‑learning pipelines. The authors observe that many historic breakthroughs—such as speculative decoding, which abandoned the long‑standing assumption that a language model must perform a separate forward pass for each token—arose when researchers deliberately questioned implicit constraints. Yet, they note, most algorithmic papers never enumerate the assumptions they rely on, leaving a large search space untapped.
The AutoAssumption framework treats assumptions themselves as searchable variables. By enumerating and testing alternative premises, the system can surface novel algorithmic variants that would otherwise remain undiscovered. Early examples cited include variations on token‑generation strategies and memory‑management tricks that could cut inference latency or training cost.
If the approach proves scalable, it could reshape how the community engineers ML systems, turning what is now a largely intuition‑driven process into a more reproducible, data‑driven one. That would accelerate the pace of efficiency gains, lower barriers for smaller labs, and potentially reshape hardware‑software co‑design roadmaps.
The next steps will be watching for peer‑reviewed validation of the method, open‑source releases of the search tooling, and any follow‑on papers that apply AutoAssumption to concrete domains such as vision models or large‑scale language models. Early adoption by research groups or cloud providers could quickly turn the concept into a practical accelerator for the next generation of AI systems.
A new preprint on arXiv (2611.14804v1) introduces **GenPFN**, a method that leverages a pretrained tabular Prior‑Data Fitted Network (TabPFN) to generate candidate designs without any additional training. The authors argue that costly simulations and physical experiments often leave engineering teams with only sparse datasets, limiting the scope of optimisation and innovation. By treating the fixed‑weight TabPFN classifier as a guiding prior, GenPFN can propose plausible design points directly from the limited data that are already available, effectively augmenting small engineering datasets in‑context.
The approach matters because it sidesteps the data‑hungry training loops that dominate most modern generative models. In fields such as aerospace, automotive or materials engineering, where each simulation can consume days of compute or expensive lab time, a training‑free generator could accelerate early‑stage exploration and reduce reliance on large‑scale data collection. Moreover, the use of a tabular foundation model aligns with a growing interest in applying foundation‑model techniques beyond language and vision, extending them to structured engineering data.
The next steps will likely involve empirical validation on real‑world design problems and benchmarking against existing data‑augmentation or surrogate‑model pipelines. Researchers will watch for follow‑up studies that test GenPFN’s ability to improve optimisation outcomes, integrate with simulation‑in‑the‑loop workflows, or combine with emerging multimodal models for richer design representations. If the method proves robust, it could become a lightweight tool for engineers seeking rapid, data‑efficient ideation.
Smuggling of 1.5 million AI GPUs could be the cause of Huang's departure
New information suggests that the sudden exit of Nvidia’s long‑time chief executive, Jensen Huang, may be linked to a large‑scale smuggling operation involving roughly 1.5 million AI‑focused graphics processors. While details remain sparse, the allegation points to a breach in Nvidia’s supply chain that could have undermined confidence among investors and partners.
The claim arrives weeks after a wave of reporting highlighted the turbulence surrounding Huang’s ouster. As we reported on 14 November, his departure left a “big hole” in AI startup financing and raised questions about the stability of Nvidia’s $500 billion AI lending consortium. Earlier coverage also noted pressure on Nvidia to bolster the residual‑value backstop for that financing vehicle. If the smuggled GPUs were indeed intended for the same ecosystem of AI startups that Nvidia has been funding, the incident could explain why senior leadership chose to step down amid mounting scrutiny.
The potential fallout matters for several reasons. A breach of this magnitude could expose Nvidia to regulatory penalties, damage its reputation as a trusted hardware supplier, and destabilise the financing arrangements that have underpinned much of the sector’s recent growth. It also raises broader concerns about the security of high‑value AI components in a market where demand outstrips supply.
Going forward, observers will watch for formal investigations by customs and trade authorities, any statements from Nvidia’s board, and how the episode might reshape the company’s hardware‑supply strategy. The next few weeks could also reveal whether the financing vehicle will survive the scandal or be restructured, influencing the flow of capital to AI startups across the Nordics and beyond.
