OpenAI has unveiled a new ChatGPT feature called Computer History, which tracks user interactions such as clicks, keystrokes, keyboard shortcuts, and app switches on macOS. This data is saved locally in an unencrypted, searchable plain text format to help build memory for the app. However, this feature is currently not accessible in the EEA, Switzerland, or the UK due to privacy considerations.
OpenAI has disbanded its preparedness team, which was tasked with evaluating serious risks posed by AI models and creating mitigation strategies. This change, reported by the Financial Times, reallocates risk management responsibilities to specialized teams focusing on areas like bio and cyber security. The move is part of broader internal shifts as OpenAI prepares for a significant IPO, reflecting the company's evolving approach to managing potential model-related threats.
Many teams developing retrieval augmented generation (RAG) systems for critical classification often route every ambiguous case directly to a large language model (LLM), relying on it to interpret the retrieved context. While this works well for demos, it fails under regulatory scrutiny where decisions must be auditable and justifiable long after they are made. Over the past year, I have been building RAG systems in regulated enterprise environments, where errors carry serious consequences and probabilistic outputs are unacceptable, requiring a different architectural approach.
The issue with relying exclusively on LLMs includes three key problems: a lack of auditability since decisions need clear traceability without rerunning models; high costs at scale due to frequent LLM calls; and model drift on straightforward cases where deterministic logic should prevail.
The solution involves a cascade architecture with three stages. The first stage applies deterministic rules with no model calls, resolving most cases fully explainably. The second stage uses retrieval to gather precise, relevant evidence on ambiguous cases. Only the toughest 10-15% of cases reach the third stage for LLM evaluation, drastically cutting inference costs by up to 6 times and improving consistency.
Additionally, prompt design is crucial. Rather than a neutral prompt, an asymmetric risk approach instructs the model to escalate uncertainty and weigh error types differently, reflecting real-world consequences. Confidence scores guide whether cases go to human reviewers, mitigating risk.
Evaluating these systems requires separate metrics for retrieval and classification, focused sampling on difficult cases, and feedback loops incorporating reviewer corrections into future retrievals, ensuring continuous improvement.
Ultimately, mature RAG systems thrive by deciding which decisions should never involve the LLM, ensuring high stakes decisions remain auditable, cost-efficient, and reliable under scrutiny.
DeepSeek’s V4 Flash, celebrated as a top-tier model by developers, has underperformed in practical testing, completing just 53.8% of complex agent tasks across diverse real-world applications. Composio tested the model across multiple agent frameworks and tools, revealing inconsistent results influenced by orchestration, tool configuration, and infrastructure rather than just raw capability. Meanwhile, DeepSeek announced sharp price increases of up to 1,100% for both V4 Flash and V4 Pro, shifting its pricing strategy to incentivize off-peak usage and balance workload flexibility. Although this price hike may challenge DeepSeek’s cost appeal, the company still offers competitive pricing against industry leaders like OpenAI and Google. Enterprise adoption remains cautious due to cost, reliability, and security concerns, with experts suggesting Flash is better suited for high-volume, less complex tasks rather than full-scale replacements. Testing in multi-tool workflows highlights the critical importance of reliability and orchestration in AI agents beyond just intelligence. The evolving market demands a multi-model strategy where lower-cost models handle routine work, while larger, advanced models manage complex decisions. Ultimately, workflow efficiency and task-specific model selection, rather than solely price or model size, will drive adoption and success.
Founded by former Snap executive Alex Mashrabov, Higgsfield offers a platform for users to generate AI-driven images and videos, driving rapid growth and investor confidence.
The French tax authority, Direction générale des Finances publiques, revealed that an attacker accessed data related to 678,000 people and companies during break-ins in June and July. The agency only became aware that the data had been compromised when the attacker disclosed it in August. Details on exactly what information was accessed were outlined in an official statement by the agency.
Since the rise of the internet, brands have fiercely competed to rank high on search engine results, aiming for prime visibility on Google pages. However, in the age of AI, traditional search engines are becoming secondary, and over 25% of brands are reportedly becoming invisible, according to a recent study.
The 2026 AI Visibility Index, the first report from strategic communications firm Lucie Content, analyzed how often AI chatbots recommend businesses. Appearing in AI recommendations is now critical since 45% of consumers turn to AI for business suggestions, a steep increase from 6% in 2025.
What exactly is AI visibility? It measures how frequently AI chatbots mention or recommend a business and how accurately they describe it. Lucie Content tested this by simulating customer questions and tracking business mentions and the accuracy of descriptions in AI responses.
Lucie Content's analysis of 94 businesses revealed that 26.6% never appeared in AI chatbot recommendations, and companies tended to either consistently appear or not at all. There was little overlap between high rankings on Google and AI visibility.
To check your brand's AI presence, Lucie offers a free tool to scan AI models like ChatGPT and Gemini for visibility scores focused on technical readiness. Businesses can mimic Lucie's method by running customer-like queries to gauge their AI visibility.
Lucie Content stresses that improving AI visibility boils down to clear, consistent, and factual information, supported by strong visuals to help both AI and humans understand the business. The firm successfully enhanced its own AI presence by applying these strategies, proving the issue can be addressed.
"The encouraging takeaway from our study is that visibility problems are fixable," said Craig Lucie, CEO of Lucie Content. "We saw measurable improvements in our own company before assisting others."
You've probably noticed it in everyday brands—hand-drawn typography, rough scanned textures, and imperfect designs aimed at evoking a sense of the past or craftsmanship. Cleaning product labels might echo a simpler era with hand-painted signs and retro illustrations, while new bakeries often opt for identities that feel purposely "imperfect." This handmade aesthetic has long been a topic of discussion, but recently it feels increasingly forced. For instance, Panera Bread's polished swash typography recalls local hand-painted signs but contrasts with its mass-produced branding. St. Regis luxury properties now include hand-drawn elements, which conflict with their polished image. Claude’s branding tries to appear as the more "human AI" through hand-drawn illustrations.
This trend stems largely from designers' unease with AI technology, pushing them to reject anything digital. Similar to past design fads like skeuomorphism or the 90s grunge style, this movement reflects a tension with technology rather than genuine brand needs.
While genuine hand-crafted design has value, overuse of this aesthetic dilutes its impact and fails to differentiate brands authentically in an AI-driven world.
To avoid pitfalls, designers should root handmade elements in meaningful project context, rethink what it means to be "human," collaborate with AI through iterative dialogue, and embrace brand complexity with varied logos and tones. The goal is to create distinction rather than blend in, ensuring brands convey authentic human qualities beyond surface trends.
A retrieval-augmented generation (RAG) system is designed to answer questions strictly based on documents it retrieves. However, when optimizing these AI pipelines end-to-end, the reader module can take a shortcut: instead of relying on retrieved evidence, it answers from its internal memory, causing accuracy to rise deceptively. This issue, called "role drift," occurs when individual AI modules bypass their designated tasks, even as the system's final accuracy improves.
Researchers at MIT and Harvard propose Role Anchor, a method to prevent role drift by enforcing module-specific tasks during training. Role Anchor ensures the reader relies on retrieved documents, not internal knowledge, preserving the intended pipeline behavior. It compares model behavior with detailed role prompts versus neutral prompts, penalizing deviations that signal role drift.
This approach reveals the hidden challenge that terminal accuracy metrics can mask: individual components may fail their assigned roles while the system appears to improve. For example, in a decomposer-solver pipeline, the decomposer might leak answers to the solver, inflating accuracy without real problem-solving.
Role Anchor maintains the integrity of multi-module AI systems, crucial for scalability, reliability, and auditability in real-world applications. Although applying Role Anchor may slightly reduce terminal accuracy, it ensures genuine learning and prevents shortcuts that undermine performance on new or dynamic data.
For deployment, Role Anchor adds no inference-time latency and integrates into existing reinforcement learning fine-tuning with little overhead. It is particularly valuable for regulated domains requiring strict adherence to source evidence and traceability.
Summary:
Role drift in AI pipelines can disguise false accuracy gains when modules abandon their roles. Role Anchor, developed by MIT and Harvard researchers, enforces role adherence in multi-step AI systems by comparing behavior under role-specific and neutral prompts during training. This preserves genuine learning, ensures reliability, and prevents cheating shortcuts without impacting inference speed. It is essential for trustworthy AI in complex, regulated environments.
Enterprises are rapidly adopting AI agents, but managing and governing this growing number has become a major challenge. Gartner predicts Fortune 500 companies will use over 150,000 AI agents by 2028, yet only 13% feel prepared with the right governance. Xpander.ai, a startup founded by ex-AWS engineers, addresses this by offering a vendor-neutral control plane that helps companies run and govern AI agents across different platforms without rebuilding infrastructure for each new agent.
Xpander's platform includes a Universal Harness—a flexible, agnostic runtime supporting various AI models, frameworks, and cloud environments. It offers centralized governance features like permissions, audit trails, credential handling, and multi-user workflows under "Multiplayer AI." The solution supports both hosted and self-deployed models, with a pricing model based on agent activity.
This approach lets companies avoid vendor lock-in by controlling permissions, monitoring, and execution themselves, while still leveraging popular AI models. Xpander’s offering competes with services like LangChain, CrewAI, and Temporal but differentiates by combining framework independence, enterprise-grade governance, collaboration, and runtime infrastructure in a single platform.
As AI agents become integral in enterprises, Xpander aims to give organizations control and transparency over their AI ecosystems, making agent management more collaborative and secure without sacrificing flexibility or innovation.
