Large language model (LLM) agents for web interfaces have advanced rapidly, yet open-source systems still lag behind proprietary agents. Bridging this gap is key to enabling customizable, efficient, and privacy-preserving agents. Two challenges hinder progress: the reproducibility issues in RL and LLM agent training, where results often depend on sensitive factors like seeds and decoding parameters, and the focus of prior work on single-step tasks, overlooking the complexities of web-based, multi-step decision-making. We address these gaps by providing a statistically driven study of training LLM agents for web tasks. Our two-stage pipeline combines imitation learning from a Llama 3.3 70B teacher with on-policy fine-tuning via Group Relative Policy Optimization (GRPO) on a Llama 3.1 8B student. Through 240 configuration sweeps and rigorous bootstrapping, we chart the first compute allocation curve for open-source LLM web agents. Our findings show that dedicating one-third of compute to teacher traces and the rest to RL improves MiniWoB++ success by 6 points and closes 60\% of the gap to GPT-4o on WorkArena, while cutting GPU costs by 45\%. We introduce a principled hyperparameter sensitivity analysis, offering actionable guidelines for robust and cost-effective agent training.

Anthropic's critics argue it's rapidly accumulating power. The company says that's what responsible AI development looks like. Anthropic has spent the last five years warning the world about how advanced artificial intelligence could enable mass destruction, destabilize society, and cause a litany of other grave harms. But simultaneously, it has become one of the most powerful forces pushing AI capabilities forward. The company is now among the top developers and distributors of cutting-edge AI models and courts customers like the US military.

Training-free source selection for LLM families with shared vocabularies arises in scientific string domains such as SMILES, protein, and genomic sequences, where candidate corpora share a tokenizer but differ in prediction targets. This creates an activation-dark regime: representation-similarity metrics can be uninformative without assumptions about label-conditioned error geometry, while classical update-geometry metrics are computationally prohibitive at vocabulary scale. We show that, in a shared-output head setting, representation metrics (e.g., CKA) are non-identifiable for transfer; models can share identical representations yet have orthogonal head updates. The key identity is that head Fisher alignment is exactly a cosine between kernel mean embeddings in the joint activation-error space, exposing activation, error, and coupling factors rather than requiring a materialized Fisher matrix. FisherSketch estimates this cosine directly in a single streaming pass, making K=128,256 head Fisher alignment practical with a 16 KB task signature (m=4096) and a 192 KB per-task streaming state, small enough to store next to a model hash, but encoding transfer-relevant update structure. Beyond source selection, the same signatures and marginals provide a diagnostic instrument for studying whether LLM task similarity is driven by activations, errors, or their coupling; shared-parameter and internal-layer validations, together with Llama-3.1-8B verbalizer-shift experiments, show that FisherSketch remains informative when activation similarity cannot distinguish tasks.

Amazon-owned MGM Studios' decision to drop the OpenAI movie is just part of AI and film industries becoming increasingly intertwined. On, we take a look at where this is all headed. This week on Uncanny Valley, our hosts discuss Amazon's controversial decision to drop Luca Guadagnino's film about OpenAI's Sam Altman--which reportedly did not paint him in a favorable light. Alongside Google DeepMind's $75 million brand new partnership with indie film studio A24, how much of a dent is AI actually having in the films we see? They also dive into the recent upheaval of workers--from electricians to software engineers--against data centers. Plus: Meta's program to track employees' data gets paused after a massive leak, and Anthropic is now getting along with the government thanks to CEO Dario Amodei no longer being in the room. Write to us at [email protected] . You can always listen to this week's podcast through the audio player on this page, but if you want to subscribe for free to get every episode, here's how: If you're on an iPhone or iPad, open the app called Podcasts, or just tap this link . Before we start, two quick things. If you've been enjoying listening to the show, we would appreciate it if you took a second to rate it in your podcast app of choice. It really helps us reach more people. And second, if you have any questions related to tech, privacy, or politics that you would like me, Zoë, and Leah to take on, now is the time to submit them to [email protected] . It doesn't matter how big or how small, we want to hear from you and get you answers. We're discussing Amazon's MGM Studios' sudden decision to drop the OpenAI biographical movie just as they were wrapping up production.

OpenAI will initially only release ChatGPT 5.6 to government-approved customers OpenAI will initially only release ChatGPT 5.6 to government-approved customers You may not be able to use the new ChatGPT 5.6 as soon as it's finished. According to a report in, OpenAI plans to stagger the release of its new AI model, and the first users will only be parties that are approved by the federal government. The publication's sources said that, according to a staff memo from CEO Sam Altman, federal leaders will be approving access customer by customer during this preview period, hopefully followed a couple of weeks later by a more general release of the 5.6 model. We've made clear to the US government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases, Altman reportedly told employees in the memo. Several agencies appear to be involved in directing the change in course from OpenAI.

When you purchase through links in our articles, we may earn a small commission. Stop asking ChatGPT to rewrite your drafts. Turn ChatGPT into a writing coach rather than a ghostwriter with this prompt. So you've drafted a cover letter for a job application, but it just doesn't feel right. Maybe your writing is a little flat (it's hard to write about yourself, after all) or the letter just comes across as unfocused.

GPT-5.5 Instant is now more capable at processing complex questions. OpenAI has updated GPT-5.5 Instant, the model you interact with the most when you use ChatGPT, to be better at understanding context and adapting to queries as you alter them to add more conditions or clarifications. The company updated ChatGPT's default model to GPT-5.5 Instant in May. Back then, it said that the model produced 52.5 percent fewer hallucinated statements during testing and 37.3 percent fewer factual errors. Now, the model has been upgraded to be more capable when it comes to identifying the underlying goal of a task or a question and carrying context over across multiple back-and-forths as you talk to it.

Hyperparameter selection is a critical step in the deployment of modern artificial intelligence systems, given the need to tune degrees of freedom such as inference-time parameters, implementation-level settings, and thresholds driving decision rules. Despite its practical importance, hyperparameter selection is typically performed using best-effort empirical methods such as grid search or Bayesian optimization, which provide no formal statistical guarantees on reliability or safety. This monograph presents a unified statistical framework for reliable hyperparameter selection, centered on the learn-then-test (LTT) paradigm, which formulates the problem as multiple hypothesis testing over a candidate set of hyperparameters. The framework enables the selection of hyperparameters that provably satisfy application-specific reliability requirements -- such as bounds on average risk, quantile risk, or information-theoretic constraints -- with explicit, finite-sample control of error probabilities. The supporting statistical machinery, namely p-values, e-values, and concentration inequalities, is developed from first principles in a dedicated appendix.

OpenAI and Broadcom have unveiled the design for Jalapeño, their first jointly-made chip. The pair of companies announced plans to collaborate on a making a custom AI accelerator in October 2025. In its blog post today, OpenAI called Jalapeño its first Intelligence Processor: an accelerator architected around OpenAI's vision for the future of LLM inference. In other words, the processor is designed to run its large language models. The AI company claims that so far, Jalapeño is offering performance per watt substantially better than current state-of-the-art in chip technology.

Despite the rapid advancement of LLM-based agents, the reliable evaluation of their safety and security remains a significant challenge. Existing rule-based or LLM-based evaluators often miss dangers in agents' step-by-step actions, overlook subtle meanings, fail to see how small issues compound, and get confused by unclear safety or security rules. To overcome this evaluation crisis, we introduce AgentAuditor, a universal, training-free, memory-augmented reasoning framework that empowers LLM evaluators to emulate human expert evaluators. AgentAuditor constructs an experiential memory by having an LLM adaptively extract structured semantic features (e.g., scenario, risk, behavior) and generate associated chain-of-thought reasoning traces for past interactions. A multi-stage, context-aware retrieval-augmented generation process then dynamically retrieves the most relevant reasoning experiences to guide the LLM evaluator's assessment of new cases. Moreover, we developed ASSEBench, the first benchmark designed to check how well LLM-based evaluators can spot both safety risks and security threats. ASSEBench comprises 2293 meticulously annotated interaction records, covering 15 risk types across 29 application scenarios. A key feature of ASSEBench is its nuanced approach to ambiguous risk situations, employing Strict and Lenient judgment standards. Experiments demonstrate that AgentAuditor not only consistently improves the evaluation performance of LLMs across all benchmarks but also sets a new state-of-the-art in LLM-as-a-judge for agent safety and security, achieving human-level accuracy.