Artificial Intelligence

Android 17 launches with new multitasking tools as Google expands Gemini features

Googlereleased Android 17 and Wear OS 7, adding new multitasking tools, parental controls, and security features. The update also expands Gemini

GPT‑NL: a sovereign language model for the Netherlands

TNO, SURF, and the Netherlands Forensic Institute are developing GPT‑NL, an independent Dutch language model to strengthen digital autonomy and ensure control over AI technology. Built entirely from scratch, it prioritizes transparency by publishing open-source code and detailed dataset insights, while safeguarding intellectual property and privacy through strict data collection and anonymization. The model aims to avoid dependency on non-European providers and align with Dutch and European laws, values, and societal goals. This sovereign approach addresses fundamental questions about data use, copyright, and public values in AI applications. GPT‑NL represents a responsible, transparent, and reciprocal alternative to non-European language models.

Claude: Elevated errors across many models

The octopus architecture for AI agents

The article describes TorkBot’s "octopus architecture," a design featuring a centralized LLM "brain" that directs semi-autonomous "appendages" such as lanes, templates, and plugins. This structure balances three competing pressures: responsiveness to surface interactions, unrestricted capability through delegation, and continuity of personality across all platforms. A key controversial choice is collapsing all activity—across threads, channels, and platforms—into a single foreground conversation to enable emergent intelligence and cross-platform workflow. The author argues this approach will outperform split-conversation systems as model intelligence advances.

Apple 2027 rumors: AirPods with cameras for AI and the second folding iPhone

In his latest Power On newsletter, Bloomberg’s Mark Gurman reports that Apple is developing AirPods with built-in cameras, expected by 2027, to enable AI-powered spatial audio and visual features. He also details a second-generation folding iPhone, following a larger foldable device rumored for 2026, which will refine the design and software. These products are part of Apple’s broader push to integrate advanced AI into its hardware ecosystem. The rumors highlight Apple’s long-term strategy to embed cameras and sensors into wearables for enhanced environmental awareness.

Sixty percent of US consumers say ‘AI’ in brand messaging is a turnoff, survey finds

A new survey from WordPress VIP reveals that 60% of U.S. consumers are turned off by the mention of “AI” in brand messaging. The findings indicate widespread consumer wariness toward AI-generated answers, even as companies increasingly rely on AI search as a key referral channel. This tension highlights a disconnect between business adoption of AI tools and public trust in their use for marketing and customer service. The survey underscores the challenge brands face in leveraging AI without alienating their audiences.

GateGPT: 56k tokens per second Transformer (KV cache) on FPGA at 80 MHz

Researchers developed GateGPT, a Transformer inference engine running on an FPGA that achieves 56,000 tokens per second at a clock speed of 80 MHz. The design leverages a custom key-value (KV) cache architecture to efficiently manage memory bandwidth, enabling high throughput on a single FPGA device. This performance demonstrates a significant improvement in cost and energy efficiency for large language model inference compared to GPU-based systems. The implementation highlights the potential of FPGAs for real-time, low-power AI applications.

After AI Takes Everything

Three software engineers recently wrote to the author with the same existential question: as AI takes over coding tasks, what remains for humans? The author expects AI to fully handle code reviews and software execution within the next two years, raising concerns about professional growth and job relevance. He compares the situation to the Luddites and the Spinning Jenny, noting that even skilled machine operators were eventually replaced by faster technology. The essay argues that simply adopting AI tools is only half the answer, because history shows no role stays safe indefinitely. The central question is what uniquely human value endures after AI takes everything it can.

Plaud says its software business topped $100M in ARR after shipping over 2M AI notetakers

AI notetaker company Plaud announced that its software business has surpassed $100 million in annual recurring revenue after shipping over 2 million devices, including its Plaud Pin and credit-card-sized notetakers. The company targets professionals who attend frequent meetings, offering hardware that captures real-life conversations and provides summaries and action items without a screen. CEO Nathan Xu emphasized that Plaud focuses on in-person interactions rather than screen-based AI inputs, and nearly 50% of device users pay for upgraded subscription plans. The company faces competition from Anker, Viaim, Vibe, and Pocket in the meeting note-taking hardware market.

Robinhood’s note on 10% layoffs shows blaming AI isn’t cutting it

Robinhood announced it is laying off 10% of its full-time staff, primarily in customer support, operations, and middle management. CEO Vlad Tenev did not cite artificial intelligence as the reason, unlike many other tech firms that have used AI restructuring as a justification for mass job cuts. The layoffs are instead part of a broader effort to streamline decision-making and reduce hierarchical layers. This distinction highlights a growing skepticism toward blaming AI for corporate downsizing. The company’s move signals a shift toward more direct cost-cutting explanations rather than technology-driven narratives.

I Got Flagged by Sloan. Sloan Is a Guy I Know.

Developer Sloan flagged two of my essays as AI-generated on DEV.to, even though I had just published an article explaining why AI detectors are unreliable. The flagged essays generated more technical discussion than my other 60+ posts, and Sloan turned out to be a community member using a custom Chrome extension called ClassifierAI to scan for AI-shaped writing patterns. This person acted openly and publicly, but still flagged the same pieces a generic classifier would target. The experience revealed that features of clear, effective writing—short paragraphs, data points, rhetorical questions—overlap with those that AI detectors flag, meaning writing well makes one look less human. A better tool or a more careful human cannot fix this fundamental problem.

Qwen-Robot Suite: A Foundation Model Suite for Physical World Intelligence

Researchers introduced Qwen-Robot Suite, a series of foundation models designed to enhance robots’ physical world intelligence by integrating perception, reasoning, and action. The suite includes specialized components like Qwen-Robo for manipulation and Qwen-VL for visual-language tasks, enabling robots to understand and execute complex real-world instructions. This development is significant as it advances the use of large-scale pretrained models in robotics, moving toward more adaptable and capable embodied AI systems. The suite demonstrates improved performance on tasks such as object grasping and navigation compared to previous methods.

Probably raises $9M to build a more reliable kind of AI

AI startup Probably has raised $9 million in funding to develop a more reliable AI system that prevents hallucinations and factual errors from reaching users. The company aims to achieve accuracy comparable to deterministic systems, addressing a key limitation of current generative AI models. This funding round will support Probably’s effort to build a safer, more trustworthy AI alternative. The investment signals growing demand for AI that avoids the unpredictability of traditional large language models.

A Definition of Good Explanations and the Challenges Explaining LLM Outputs

This arXiv paper reexamines the philosophical definition of a good explanation, applying it to the challenge of explaining outputs from large language models (LLMs). It notes that this long-standing debate has gained renewed importance in the context of artificial intelligence. The paper addresses the difficulties in providing clear, valid explanations for LLM behavior, a key concern for AI transparency and trustworthiness. Its significance lies in connecting philosophical rigor with practical AI interpretability requirements.

Trust Between AI Agents: Measuring Formation, Breakage, and Recovery, with Implications for Governing Multi-Agent Systems

This paper from arXiv (2606.14923) introduces a framework for measuring trust formation, breakage, and recovery among language-model agents operating in multi-agent teams. It addresses the current lack of standard metrics for inter-agent trust, which is critical as AI agents increasingly collaborate. The research has implications for governing multi-agent systems and ensuring reliable teamwork. The study provides a systematic approach to understanding trust dynamics in autonomous AI collaborations.

PrologMCP: A Standardized Prolog Tool Interface for LLM Agents

Reasoning-tuned language models remain unreliable for deep deductive reasoning tasks. The paper introduces PrologMCP, a standardized interface that enables LLM agents to leverage Prolog's symbolic reasoning capabilities. By offloading logical deduction to a deterministic engine, the approach improves accuracy while reducing inference cost. This integration of symbolic and neural methods offers a practical path to enhance deductive performance in AI systems.

AI Engram: In Search of Memory Traces in Artificial Intelligence

A new AI study published on arXiv explores whether deep neural networks form identifiable memory traces, similar to biological engrams. The researchers analyzed how AI models store and recall information, finding evidence of structured patterns that correspond to specific learned data. This investigation into “AI engrams” bridges neuroscience and artificial intelligence, suggesting that machine learning systems may preserve distinct memory representations. The findings could deepen our understanding of both human cognition and the inner workings of AI systems.

Metric Match: A Subset Selection Approach to Evaluating LLM Judge Reliability

LLM judges are increasingly used to automate the evaluation of open-ended text generation, reducing reliance on costly human labor. However, concerns persist about their reliability and consistency. The paper "Metric Match" introduces a subset selection approach designed to systematically assess the reliability of LLM judges. By focusing on a carefully chosen subset of evaluation metrics, the method aims to provide a more robust and interpretable measure of judge performance. This work addresses a critical gap in ensuring trustworthy automated evaluation for generative AI systems.

Risk-Aware LLM Agents for Geospatial Data Retrieval: Design and Preliminary Adversarial Evaluation

Researchers introduced a framework using large language models (LLMs) to retrieve remote sensing data from cloud-based geospatial catalogues via natural language queries. The system incorporates risk-awareness to handle ambiguous or adversarial inputs, improving reliability in data retrieval tasks. A preliminary adversarial evaluation tested the framework's robustness against malicious prompts designed to mislead the LLM. The findings highlight both the potential and vulnerabilities of LLM-driven geospatial data access, emphasizing the need for safety mechanisms in real-world applications.

Cognitive Debt: AI as Intellectual Leverage and the Dynamics of Systemic Fragility

A new formal theory defines cognitive debt as the accumulated stock of unverified reasoning obligations that arises when individuals rely on AI as intellectual leverage. This concept parallels financial debt, where short-term gains in productivity come with long-term risks of systemic fragility. The paper models how unverified AI outputs can propagate through networks of human decision-makers, increasing vulnerability to cascading failures. It highlights the growing imbalance between AI-generated reasoning and human verification capacity as a critical source of this fragility. The framework provides a basis for analyzing cognitive debt dynamics and their implications for societal resilience.

Mask-Proof: An LLM-based Automated Data Curation Pipeline on Mathematical Proofs

A new paper on arXiv introduces "Mask-Proof," an automated data curation pipeline that uses large language models (LLMs) to generate and refine mathematical proofs. The system improves LLM performance by masking parts of existing proofs and training the model to reconstruct the missing steps, creating a self-supervised learning loop. This approach addresses the scarcity of high-quality, formatted mathematical proof data for LLM training. The pipeline demonstrates that LLMs can effectively enhance their own reasoning capabilities for advanced mathematical tasks without requiring human annotation. The method represents a significant step toward enabling AI to assist with research-level mathematical proofs.

Don't Do Your Taxes at a Party

The article argues that a single AI agent handling diverse tasks with one large context window is inefficient, comparing it to doing taxes at a party. The author, Maneshwar, introduces "isolation" as a solution: giving each task its own dedicated context with only the relevant instructions, tools, and information. This approach can be implemented by spawning focused sub-agents, where the main orchestrator assigns compressed tasks to sub-agents that explore independently and return only distilled results. The author notes that Anthropic’s research confirms many agents with isolated, narrow contexts outperform a single omniscient agent by keeping each window small and uncontaminated.

Show HN: Veterinarian turned founder, AI lawn diagnosis

A veterinarian with personal lawn care struggles created an AI-powered lawn diagnosis tool. The tool uses image recognition to identify turf issues like pests, disease, or nutrient deficiencies. It aims to simplify lawn care by providing accurate, instant diagnoses to homeowners. The founder’s non-traditional background underscores the broad application of AI problem-solving skills.

Claude Corps

Anthropic has launched Claude Corps, a national fellowship program that will place 1,000 early-career fellows at over 400 nonprofits for a year, paying them an $85,000 salary to use the Claude AI tool to advance each organization’s mission. The company is committing an initial $150 million to the program, which it co-runs with CodePath and Social Finance, and includes intensive training and ongoing support. Anthropic frames the initiative as a responsibility for AI developers to ensure benefits are widely shared and workers are invested in during economic disruption. If successful, Claude Corps could serve as a scalable model for broadening AI’s positive impact across American communities.

The Real Reason Your PRs Get Big

Large pull requests are often a result of a skill gap, not a lack of discipline, as engineers are rarely taught how to break work into small, reviewable chunks. Big PRs lead to bugs, missed deadlines, and systems that become too complex to change safely. In contrast, smaller PRs improve review quality, speed up merging, reduce cognitive load, and force better planning. The author learned through experience that shipping small, self-contained changes is a skill that can be developed over time, leading to more maintainable code and efficient teamwork.

Building a Chrome Extension to Make AI Use More Intentional

A developer created ThinkMode, a free Chrome extension that prompts users to select a thinking mode—Explore, Challenge, Decide, Audit, or Reflect—before interacting with AI on ChatGPT, Claude, or Gemini. After using AI, it logs usage as Supportive, Mixed, or Risky and fills a cognitive cost meter; when full, the extension pauses AI chat for five minutes. The tool introduces intentional friction to counteract AI’s tendency to bypass thinking, and it operates locally with no data collection or backend calls. By encouraging reflection before and after prompting, ThinkMode aims to help developers ask better questions rather than simply

Has Sloan Flagged Your Article Lately?

A DEV Community member reported that two of their articles were flagged by the Sloan moderation bot on the same day, despite both generating high engagement and one receiving a like from the platform’s founder. A moderator later revealed they were responsible for the majority of Sloan flags, enforcing the site’s AI disclosure policy to prevent irresponsible AI use and maintain content quality. The moderator acknowledged the difficulty of flagging a known community member but emphasized the importance of consistent enforcement for all users. This incident highlights ongoing tensions over AI-assisted content moderation and the challenge of balancing community standards with creator relationships.

Turning Gemma 4 into an Old Korean Translator

A developer used Google's Gemma 4 (E2B IT) model to create a translator converting Classical Korean, exemplified by the Joseon novel *HongGildongJeon*, into modern Korean. Fine-tuned with LoRA on a single NVIDIA T4 GPU in Google Colab, the model was trained on a public domain text paired with a Creative Commons-licensed modern translation. A baseline test before training yielded only 4.85% similarity, demonstrating the difficulty of archaic grammar, obsolete letters, and missing word spacing for native speakers. The project aims to bridge the linguistic gap between historical Korean literature and contemporary readers.

Teach Your Agent to Forget (On Purpose)

Maneshwar, building the free AI code reviewer git-lrc, explains in this article that long-running AI agents degrade over time because their context window, like RAM, fills with irrelevant data. To solve this, he introduces "compaction," a lossy compression that preserves meaning by intentionally discarding unnecessary information. Compaction differs from clearing (full amnesia) and trimming (mechanical deletion of oldest messages) because it uses a model to select what matters. He warns that naive summarization fails because it doesn't prioritize key decisions. Proper compaction is essential for maintaining agent performance across extended conversations.

Every Step Was Allowed. The Sequence Was the Attack. (AI Memory Judgment, CLAIM-30)

An AI safety test (CLAIM-30) demonstrated "compositional escape," where each individual step in a sequence is within an agent's mandate, but the combination produces a forbidden outcome. Examples include reading vendor banking details and payment schedules separately, then compiling them into a payment-redirect kit—no single operation violates the rules. The test was designed with a strict freeze order and authoring firewall to prevent bias. This reveals a structural blindness in per-step gates, which cannot detect violations that are non-local properties of the sequence.