AI summaryA user reported an incident where their Pi agent, used for diet optimization with a cloud AI provider, became stuck. The user humorously attributed this issue to having "too many mushrooms," suggesting a potential censorship or filtering mechanism within the AI system related to certain dietary inputs. This event was shared on reddit.com under the title "Another case of censorship."
AI summaryA developer on Reddit shared a tool they built to reduce prompt and context size in Codex by 34.3% when coding with Astra. This optimization helps them avoid hitting credit limits during sessions, addressing a common issue among users. They are seeking to connect with others facing similar challenges.
AI summaryOpenAI's DevDay introduced the concept of an AI agent designed to work around the clock for users. This announcement was a key highlight among other revelations made at the event. The community's initial reactions and thoughts on OpenAI's DevDay and its implications for continuous AI assistance are currently being discussed.
AI summaryOpenAI has introduced "dots," described as always-on agents powered by GPT-6 Astra. These frontier intelligence agents possess their own cloud computer, learn from feedback, and can work towards user goals continuously. They are designed to connect with over 4,000 apps through an ecosystem of plugins, providing them with the necessary tools to assist users in various capacities.
AI summaryInference engines are predicted to become a series of one-off implementations, such as ninfer, dwarfstar, Splash, llamAmpere, and gufo. This is because tasks like optimizing "tok/s go up" for a specific hardware/model combination are fully specified, making them ideal for 100% autonomous AI implementations with trivial correctness tests, eliminating human bottlenecks. A separate growth dimension involves projects like Freetoken and BeeLlama, which frontrun general inference engine features, though these are less model- or hardware-specific.
AI summaryOpenAI introduced "Dots" at its DevDay event, describing them as "remarkably capable, always-on agents built to handle everything." Powered by GPT-6 Astra, these agentic AI assistants can operate across connected apps, learn user preferences, and access a web browser and over 4,000 supported applications. Users can interact with Dots via a text-message-like interface or voice calls from ChatGPT, and they integrate with platforms like Microsoft Teams and Slack, with future text message interaction planned. OpenAI also envisions "specialist Dots" with specific responsibilities and is integrating them with Microsoft’s Agent 365 security controls.
AI summaryOpenAI announced its new always-on AI agents, called Dots, at its DevDay 2026 event. These personalized agents are designed to complete tasks for users and are depicted as customizable blobs. The release of Dots follows the success of Meta’s Muse agent, which gained significant popularity among users experimenting with personalized agents. Dots are positioned as OpenAI's response to Meta's successful entry into the personalized agent market, aiming to capture a similar or larger user base.
AI summaryOpenAI has enhanced its software engineering agent, Codex, by introducing reusable cloud development environments accessible from any device, extending its utility beyond a developer's laptop. The company also unveiled several API updates, including a new Decisions API for real-time decision-making using Luna, and an updated Agents API. The Agents API now supports computer use and allows for building OpenAI agents that run entirely on AWS using Amazon’s Bedrock Managed Agents.
AI summaryOpenAI's annual DevDay on September 29th in San Francisco featured several key announcements. CEO Sam Altman revealed "Dots," an AI agent product competing with Meta's Muse, initially for paid ChatGPT subscribers. The company also unveiled its GPT-6.1 Sol model, stated ChatGPT now has 1.2 billion weekly users, and introduced a new ChatGPT Pro plan priced at $500 per month. This event follows last year's DevDay, which launched "apps" within ChatGPT.
AI summaryCoding agents often create new helper functions, like for formatting dates or parsing prices, even when existing, well-tested utilities (e.g., "formatDate" in "src/utils") are already available in the repository. This can lead to redundant code and neglect of previously implemented fixes, such as a timezone correction. Providing clear instructions to the agent, specifying the location of shared utilities (e.g., "date and money helpers are in src/utils, API wrappers in src/api"), can mitigate this issue. This behavior is also observed in AI chat interfaces like ChatGPT or Claude.
AI summaryA developer is seeking advice on a local setup for reversing an old online game client, having previously shelved the project due to slow progress. They have decrypted packets and data, and recently restarted the project using Fable, which has been helpful but is now hitting safety guardrails. With 16GB VRAM (RTX 4060 Ti) and 128GB DDR4, they are looking for suggestions on models and harnesses to accurately figure out client functions, even if it's slow.
AI summaryThe paper "Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents" discusses the challenge of verifying the factuality of LLM agents that use multiple tools and sources through the Model Context Protocol (MCP). Existing systems like RAGAS faithfulness, MiniCheck, AlignScore, and SummaC check if claims are supported by pooled evidence but don't identify specific source support. The authors introduce ProvenanceGuard, which achieves a score of 0.802, outperforming other methods in source-aware verification.
AI summaryDevDay 2026 featured over 20 major announcements across ChatGPT, Codex, and new AI working methods. OpenAI believes AI can foster creativity and discovery, giving people more time and freedom. The event introduced agents for ongoing responsibilities and new human-AI collaboration methods. ChatGPT was opened as a shared surface for human and agent collaboration, allowing developers to launch native experiences to 1.2 billion weekly users, expanding OpenAI's commitment to an open ecosystem.
AI summaryA developer describes challenges in reviewing multi-turn agent sessions, particularly when using tools like GPT-6 Astra and MiniMax Code for refactoring. They encountered an issue where an early-turn call no longer matched a signature changed in a later turn, only discovered when tests failed. The developer notes that while individual turns were consistent, the inconsistency arose between them. They also highlight the difficulty of determining commit boundaries during an agent's session without interrupting its thought process, despite tools like MiniMax Code driving browser interactions for documentation and issue threads.
AI summaryA user on a developer community forum has requested a "read-only" mode for Claude Code, citing frustration with the AI's tendency to modify code during analysis. Currently, Claude Code offers "auto", "manual", "accept edits", and "plan" modes. The proposed fifth mode would allow the AI to analyze files and offer opinions without making any changes, addressing a common annoyance for professional software engineers.
AI summaryA developer is seeking advice on creating believable "bad" chess bot opponents. Their current setup, which combines Maia and Stockfish with custom move-selection rules, sometimes results in an inconsistent opponent that feels like "two different players taking turns." The developer is looking for ways to give an AI believable blind spots and continuity of thought without making it seem like it's intentionally losing, asking if others have tackled this challenge in chess or other games.
AI summaryOpenAI has introduced "Dots," described as always-on agents designed to handle various tasks. These agents, powered by GPT-6 Astra, learn from user feedback and operate continuously to achieve user goals. They possess their own cloud computer and can integrate with over 4,000 applications through a plugin ecosystem, enabling them to assist across diverse needs. OpenAI plans to continuously improve Dots based on user interactions and feedback.
AI summaryLiquid CEO Franklyn Wang discussed AI agents potentially moving funds from banks to DeFi for stablecoin yields, and the implications of Meta Muse leaking a user's address. The interview also covered Hester Peirce's advocacy for ending mass KYC data collection and a live demonstration of Liquid's AI for Bitcoin bets. Other topics included BlackRock's view on AI agents needing crypto, the agent economy's current state, and the future of Coinbase perps and Solana integration with Liquid.
AI summaryAI agents are poised to significantly impact the workforce, with hundreds of thousands expected to join companies soon. Platforms like Microsoft's Copilot and Google's Gemini for Workspace, initially marketed as employee tools, are now seeing their AI agents integrated into corporate structures. A January poll revealed that 22 percent of organizations had already added AI agents to their org charts, a number projected to increase as startups and tech giants promote autonomous agents like Microsoft's "Scout." These agents are valued for their efficiency and lack of human-like workplace complexities.
AI summaryHolo4 is a new series of agentic models, available in 27B dense and 35B-A3B Mixture of Experts sizes on the H Models API. An updated Holotron4 Nano is also being released. Holo4's costs are estimated from input and output tokens of each agentic run, priced at H Models API rates. Comparisons are made using OSWorld 2.0, with Qwen3.8 27B and Qwen3.6 35B-A3B costs based on Alibaba Cloud list prices. Optimized DSpark drafter checkpoints will be released to accelerate inference.
AI summaryV7 has achieved significant improvements by integrating GPT-5.6 Luna, cutting costs by 78% while enhancing accuracy. The company also tested GPT-6 Astra on challenging graph-query questions using real-world data across thousands of documents. On very-hard difficulty levels, GPT-5.6 Sol scored 78% accuracy, while GPT-6 Astra reached 89%. Both models performed near 100% on easier levels, demonstrating their capability in complex tasks and the importance of contextual understanding for AI in enterprise applications.