We deploy a simple into our everyday lives as search engines (Jin et al., 2025; statistical framework based on risk control (Angelopoulos Xiong et al., 2024), coding assistants (Zhao et al., 2023), et al., 2022) that converts any safety signal into a binary and companions (Zhang et al., 2025a). As their applicability grows, so does the potential harm caused by malicious decision rule, and offers statistical guarantees on the false LLM outputs. Despite remarkable performance across a alarm or missed detection rate. The framework is universally applicable to different monitoring purposes and can leverage wide range of tasks, LLMs remain prone to generating halarbitrary proxy signals. Through experiments on mathematlucinated, factually incorrect (Ravichander et al., 2025), or ical problem solving and red teaming conversations, we harmful output (Yu et al., 2025) when deployed.

AI is'not smart' so what's next in artificial intelligence? We don't have robots that are nearly as good at understanding the physical world as a rat, says Yann LeCun, one of the leading figures in the world of artificial intelligence. He worked at Facebook-owner, Meta, for a decade, where he was chief AI scientist, but left in 2025 and founded Advanced Machine Intelligence Labs (AMI Labs). His goal is to move AI beyond current systems like ChatGPT, Claude and Gemini. They have their uses, he says, but will never be able to tackle complicated situations in the real world, like getting a robot to do household chores.

The drama reveals just how deeply Silicon Valley has sunk its claws into Hollywood. Enter your email to receive alerts for this author. You can manage your newsletter subscriptions at any time. You're already subscribed to the aa_Nitish_Pahwa newsletter. You can manage your newsletter subscriptions at any time.

The Download: a startup has a solution for AI's groupthink problem Plus: Scientists say they have built a cell from scratch for the first time. LLMs are stuck in a groupthink groove. This startup is trying to get them out. Open up your chatbot of choice--Claude, ChatGPT, Gemini--and type "Give me a random number between 1 and 10." You're going to get 7. Almost always. That won't work every time--but if it did for you, you may wonder if I have superpowers. The truth is that most large language models are stuck in a rut.

PCWorld examines whether Microsoft's core productivity apps like Word, Excel, and PowerPoint can withstand disruption from advancing AI technology. External AI applications such as ChatGPT and Claude now offer similar document formatting, content creation, and synthesis capabilities that rival Microsoft's own Copilot feature. The analysis suggests Microsoft's traditional productivity suite may become obsolete as AI chatbots increasingly handle tasks previously requiring dedicated office applications. Are Microsoft's core productivity apps -- Word, Excel, and PowerPoint -- endangered by the rise of AI? That's the point that Bloomberg and its sources addressed in coverage this week, noting that Microsoft is being buffeted by AI disruption as its stock plunges. "Whether Microsoft Word or Excel will be rendered obsolete by AI remains to be seen," said Jack Ablin, chief investment strategist at Cresset Wealth Advisors, which owns the stock, according to Bloomberg. "We don't know what the environment is going to look like in a few years, which opens up very real questions like, will we even use a Microsoft suite anymore?" Keith Fitz-Gerald, principal at the Fitz-Gerald Group, added.

Sam Altman is in talks with the US government in a bid to clear political hurdles, says the Financial Times. OpenAI's Sam Altman has reportedly been in talks with the US government to ensure his company's path towards achieving its goals remains free of political hurdles. According to the Financial Times, Altman has suggested giving the government a five percent stake in the company, in order to share the spoils of the AI boom with the public. But his idea doesn't only involve OpenAI: Under his proposal, other top AI companies like Google, Anthropic, xAI and Meta would have to agree to give the government a similar stake in their businesses. AI companies like Anthropic and OpenAI have recently encountered roadblocks from the US government when it came to releasing their latest AI models.

What if a reasoning model could decide when to decompose and parallelize independent subtasks, how many concurrent threads to spawn, and how to coordinate them based on the problem at hand? We provide a detailed analysis of recent progress in the field of parallel reasoning, especially adaptive parallel reasoning. Disclosure: this post is part landscape survey, part perspective on adaptive parallel reasoning. One of the authors (Tony Lian) co-led ThreadWeaver ( Lian et al., 2025), one of the methods discussed below. The authors aim to present each approach on its own terms. Recent progress in LLM reasoning capabilities has been largely driven by inference-time scaling, in addition to data and parameter scaling ( OpenAI et al., 2024; DeepSeek-AI et al., 2025). Models that explicitly output reasoning tokens (through intermediate steps, backtracking, and exploration) now dominate math, coding, and agentic benchmarks.

OpenAI CEO Sam Altman has been in talks about public ownership with Donald Trump, according to the report. OpenAI CEO Sam Altman has been in talks about public ownership with Donald Trump, according to the report. OpenAI'in early talks to give 5% stake to US government' OpenAI is reportedly in early stage talks to give a 5% stake in the ChatGPT developer to the US government as artificial intelligence companies attempt to smooth relations with Donald Trump's administration. The OpenAI chief executive, Sam Altman, has argued that giving the US public a financial stake in the company is the best way to share the benefits of AI, according to the Financial Times, which cited two unnamed people familiar with the discussions. The proposal would also involve other US AI companies giving a similar stake to the government, the FT reported, although it is not clear yet whether companies such as Anthropic, Google and Meta would agree to the plan.

Watermarking promises a statistical trace of large language model (LLM) use, but real documents, after editing or paraphrasing, rarely arrive as purely human-written or purely machine-generated. This motivates a quantitative question beyond detection: what proportion of a document is generated from a pre-specified watermarked LLM? We study this watermark proportion estimation problem under the Gumbel--max watermarking mechanism, treating the next-token prediction (NTP) distributions as unknown and arbitrary nuisance parameters subject to a non-degeneracy condition. We compare two observation regimes: in the full observation regime, the estimator observes the pseudorandom vector and the selected token at each position; under the more popular setting of pivotal reduction, it observes only a scalar pivot, which follows a one-dimensional Uniform--Beta mixture distribution. Under pivotal reduction, we develop a Laguerre-polynomial estimator and establish a matching information-theoretic lower bound for the sample complexity. For full observation, we introduce an event-counting estimator and show a matching lower bound, yielding a substantially smaller sample complexity. As our results imply, although reducing to pivotal statistics is an elegant and widely used procedure, it is not always sample-efficient for estimating the proportion of watermarks.

Transformer-based large models have demonstrated remarkable generalization abilities across different tasks by leveraging a context-aware attention module for in-context learning. With richer context, transformers adapt more effectively to the current use case without any parameter updates. However, the quadratic computational and memory complexity with respect to context length significantly slows data processing in softmax transformers. Linear transformers were proposed to address this issue by reducing the complexity to linear dependence on context length, but the design and understanding of the feature mapping in linear attention, from a theoretical viewpoint, remain unclear. In this paper, we investigate the approximation and generalization abilities of linear transformers under a two-staged sampling process from domain generalization. We show that linear transformers perform in-context learning as learning a mapping from context distributions to response functions. A dimension-independent convergence rate is obtained for our generalization analysis, which also exhibits the tradeoff between the regularities of data distributions and latent features. Guided by our theoretical framework, we propose a new perspective on activation and loss design for linearizing pretrained softmax large language models.