https://eaglhuang.github.io/AI-learning-notes/daily/ 2026-10-11T12:00:00+08:00 Eagl Huang https://eaglhuang.github.io/AI-learning-notes/daily/2026-10-11/en/ 2026-10-11T12:00:00+08:00

Deno team joins Cloudflare and outlines runtime and Deploy transition: Ryan Dahl announced that the entire Deno team is joining Cloudflare, focusing future development on a shared platform rather than a separate runtime and hosting service. The team will combine celld work with Workers and Durable Objects, aiming to make this programming model the default for servers on Cloudflare or users’ own infrastructure. celld builds scaling into the programming model instead of requiring each application to assemble its infrastructure. The Deno runtime will receive another year of support, with monthly bug fixes and security updates; the team will then end development, while Deno remains open source for others to continue. Deno Deploy will operate for six more months before closing, with migration support for paying customers moving to Cloudflare Workers. JSR will continue, moving its infrastructure to Cloudflare; rusty_v8 support will continue, with work toward integration into workerd. Dahl says AI particularly needs better abstractions: Durable Objects combine inexpensive serverless execution, persistent state, WebSockets and a high-level JavaScript interface useful for agent harnesses, explaining celld’s focus. Summary sources: https://deno.com/blog/cloudflare Open WebUI 0.12.0 adds realtime voice, shared replies and skill history: Open WebUI released 0.12.0 with a Realtime voice-call mode administrators can enable. An OpenAI realtime voice model handles small talk and passes substantive questions and tasks to the selected chat model with its history and tools. Tool approvals and questions still require answers in chat. Calls require Allow Call permission and end after one hour. Realtime calls can use rigged VRM avatars with idle, listening and speaking animations and up to sixteen named gestures; mouth movement follows the spoken response. Shared chats and folders gain an Allow replies mode so recipients can continue the same conversation with live messages and typing indicators. People can edit, rate, regenerate or stop only their own messages. With context compaction and native function calling enabled, long tool runs can compact between tool rounds while keeping calls paired with results and including tool names and arguments in the summary. Skills can include scripts, references and templates, with a saved version history that supports comparison and restoration, plus ZIP or JSON export. The release also adds authenticator-app two-factor sign-in that administrators can require for all users, with ten single-use recovery codes and separately configurable bypasses for OAuth and trusted-header sign-ins. Summary sources: https://github.com/open-webui/open-webui/releases/tag/v0.12.0 Claude Code 2.1.296 adds subagent compaction settings and fixes permission checks: Anthropic released Claude Code 2.1.296 with autoCompactWindow for subagents, allowing them to compact earlier than the main conversation. An environment variable can select one model for all workflow agents while other subagents retain their own models. The Read tool gains allow_large to read an entire text file beyond the usual size limits when the full file is needed and context capacity permits. The Claude apps gateway gains a code field in managed policies to apply CLI settings to Claude Desktop's Code tab. The release also fixes a regression introduced in 2.1.295 that ignored saved gateway sign-ins under specific managed settings. Fixes cover missed secret redaction in shared transcripts and debug logs, and Bash permission checks that automatically approved some commands assigning and then using BASH_ARGV0. Those commands now require approval. Edit and NotebookEdit now refuse edits to files that are not valid UTF-8, addressing a bug that could replace every non-ASCII character. In VS Code, Claude in Chrome now asks before browser actions in every session, including sessions connected with @browser. Allowing a site for that session stops repeat prompts. The default limit for upfront MCP tool descriptions and server instructions rises from 2,048 to 4,096 characters. Summary sources: https://github.com/anthropics/claude-code/releases/tag/v2.1.296 Pydantic AI 2.55.0 adds persistent conversations and unified prompt caching: Pydantic AI released 2.55.0 with a Conversation object for carrying and storing a conversation between runs, accepted through the conversation argument at every entry point. A unified cache setting and Caching capability support prompt caching across providers. With Anthropic automatic caching, cache=True also caches instructions and tool definitions; the Coder harness enables prompt caching by default. The release adds an OpenAIDecisionsModel backend with image input and records prompt-cache diagnostics in provider_details. Diagnostics are enabled by default for OpenAI Responses and are opt-in for Anthropic. Under durable execution, re-execution preserves default run_id and conversation_id values. Planning records plan-store calls so recovery does not repeat writes; Memory records tool calls so forked DBOS workflows do not repeat writes. Each failed FallbackModel attempt records its model, timing, error and usage, while RunUsage counts rejected responses. On gpt-realtime and GPT-Live, hang_up() can end a WebRTC call from the server. The retain_audio_max_seconds setting bounds retained audio. All packages now require Python 3.11 or newer; Python 3.10 installations resolve to 2.54.0 or earlier. Relevant extras raise the minimum openai package version to 3.26.0 for Decisions support. Summary sources: https://github.com/pydantic/pydantic-ai/releases/tag/v2.55.0 CrewAI 1.15.27 adds per-run evaluation cost and time tracking: CrewAI released 1.15.27 with per-run cost and time tracking in crewai eval --models, without repeating progress lines. Evaluations now record why a run stopped before reaching a verdict. The release adds DeepInfra as an OpenAI-compatible provider and XPU to the OpenCLIP provider's device options. It fixes stopSequences being sent to OpenAI GPT-6, GPT-5.6 and gpt-oss. The GitHub loader preserves source attribution, and SQLite flow initialization contention is fixed. Shorter source-limited summary. Summary sources: https://github.com/crewAIInc/crewAI/releases/tag/1.15.27 ttok 1.0 changes its default tokenizer; Willison notes missing GPT-6 confirmation: Simon Willison released ttok 1.0, a tool for counting and truncating text by token count. After upgrading to 0.4 and feeding it a file, he noticed it still defaulted to the GPT-4 tokenizer, prompting a new default and a 1.0 release. The release switches to o200k_base; ttok --list-models also shows how tiktoken handles model-name prefixes such as gpt-5*. Willison cautions that OpenAI has not officially confirmed that GPT-6 and the GPT-5 family share a tokenizer. He cites William Liu’s September 26 experiment, in which seven GPT models each reported 44,794 input tokens across the same 31 fixtures, matching on every fixture. Liu explicitly limits these observations to raw-text count endpoints, without generation or billed-usage validation; equal counts do not establish that the underlying tokenizers are identical. Shorter source-limited summary. Summary sources: https://simonwillison.net/2026/Oct/9/ttok/ https://github.com/simonw/ttok/releases/tag/1.0 https://github.com/williamliu-ai/token-count-compare/commit/8d8a2178538bb37beb548fc37378135e4ff4b0c8

https://eaglhuang.github.io/AI-learning-notes/daily/2026-10-08/en/ 2026-10-08T12:00:00+08:00

GPT-6 brings interactive answers to ChatGPT: OpenAI is rolling out GPT-6 with Intelligent UI in ChatGPT's Chat tab. Answers can combine text, charts, buttons, forms and interactive tools, appearing progressively while the model continues thinking. Rollout starts October 7 for Plus, Pro, Business and Enterprise, and October 8 for Free and Go, subject to enterprise settings. Plus, Pro, Business and Enterprise use GPT-6 Sol; Free and Go use GPT-6 Luna. Work and Codex models are unchanged. Official help limits Intelligent UI to Instant through Extra High reasoning: Pro effort still uses GPT-6 Astra and does not support it. It is rolling out on the web and supported updated apps; older macOS and Windows desktop apps are unsupported. On the web, turning off Layout and visuals reduces visual elements but does not guarantee text-only answers. Intelligent UI has no separate quota; existing model, tool and plan allowances apply. Summary sources: https://openai.com/index/gpt-6-for-everyone/ https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt https://help.openai.com/en/articles/6825453-chatgpt-release-notes ChatGPT previews College Planner for teens: OpenAI previewed College Planner for ChatGPT for Teens. Initially aimed at US grades 10–12 applying to four-year colleges, it will collect requirements, deadlines, tasks and financial-aid steps in one plan; broader coverage is planned. It also helps track aid deadlines and explore scholarships and fee waivers. The announcement also describes quizzes from notes and multi-photo capture that combines pages into a PDF on iOS, with Android forthcoming. OpenAI is also supporting an independent student advisory council at Boston Children’s Hospital. The flashcard help page explains that students can build cards from a topic or uploaded notes, flip them, mark answers they know or need to practice, and shuffle the deck. Decks save automatically in Library and can be reopened or recalled in chat. Students can request additions, simpler definitions or corrections in the conversation. The guidance recommends checking cards against original notes or study guides, especially before an exam. The council’s own page describes two tracks: high-school students aged 14–18 develop individual community digital-wellbeing projects, while students aged 16–20 test emerging AI tools together and coauthor industry recommendations. Participants learn from researchers and clinicians from October 2026 through May 2027. Regular meetings are virtual, participation is free, and applications for that school year have closed. Summary sources: https://openai.com/index/teens-learn-and-plan/ https://help.openai.com/en/articles/20001533-flashcards-in-chatgpt https://digitalwellnesslab.org/student-advisory-council/ Google Playground turns prompts into playable games: Google introduced Playground, an experimental browser platform where prompts control game rules, physics, characters and environments. Creations can stay private, be shared by link or enter a screened public gallery; selected genres support multiplayer. Launch is for US adults, with creation access rolling out by Google AI membership tier. Unity Spark integration remains in testing, with a closed beta still forthcoming. Official help separates playing from creating: the catalog is free to play; free creation tokens refresh weekly, with access via a waitlist, while subscriptions unlock immediate creation access. Creation Studio can export a ZIP containing HTML, JavaScript, CSS and generated assets. Published games remain editable, with changes kept as drafts until republished and the existing public version remaining playable. Credits are deducted only after successful compilation into a playable game; failed or timed-out generation automatically refunds them. Additional terms state that unlisted games remain accessible to link holders, and public games may be discovered on other Google products. Google does not claim ownership of users’ games but retains rights to its supplied features, services and assets. Creators must secure required permission for people’s images and third-party intellectual property and must not remove AI-origin watermarks, labels or metadata. Summary sources: https://blog.google/innovation-and-ai/technology/ai/playground-experimental-gaming-platform/ https://playground.google/helpcenter https://playground.google/terms SynthID Detector opens globally in English: Google expanded SynthID Detector from a media-professional trial to global public access in English. It checks images, video and audio for AI generation by Google and partners including OpenAI, NVIDIA and Kakao, with Apple forthcoming. The expansion complements existing verification in Search, Gemini and Chrome. The technical overview explains that SynthID embeds imperceptible watermarks when content is generated. Image and video marks are designed to survive cropping, filters, frame-rate changes and lossy compression while preserving quality. Audio marks cover content such as Lyria output and NotebookLM podcasts and are designed to withstand added noise, MP3 compression and speed changes. Identification relies on embedded marks, rather than universal detection of every AI-generated item. The original 2025 announcement described uploading content, scanning for watermarks and highlighting likely marked portions, such as audio segments or image areas. Applications initially opened to journalists, media professionals and researchers. Google’s Gemini help cautions that repeated alterations can prevent watermark detection, while simple or abstract content and small edits can be inconclusive. SynthID verification inside Gemini currently recognizes Google AI content only, unlike the partner coverage announced for Detector. Failure to find a Google watermark does not rule out generation by another AI system. Summary sources: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/ https://deepmind.google/models/synthid/ https://blog.google/innovation-and-ai/products/google-synthid-ai-content-detector/ https://support.google.com/gemini/answer/16722517 Liquid AI releases two open-weight d1 decision models: Liquid AI released open-weight d1-3B and d1-omni-600M on Hugging Face. These decision models return structured judgments without generating output tokens, so the announcement reports end-to-end latency. d1-3B supports text and vision; d1-omni-600M is an experimental checkpoint handling text with images or audio. The release includes llama.cpp support and vendor measurements across hardware and workloads. Reported inference latency is for d1-3B; d1-omni-600M remains an early research release. The model cards describe a state plus named questions, returning yes/no probabilities, named choices or ratings. d1-3B is based on LFM2.5-VL-3B, with about 3.12 billion parameters and 32,768-token context; it is not a chat model and does not write text. The approximately 587-million-parameter omni checkpoint shares its trunk across modalities. A request may carry images or one audio clip, not both. Audio training used English-speaker/assistant requests and clips are cut at 30 seconds. With images, state and question text is limited to 896 tokens. Both cards list the LFM 1.0 license. Summary sources: https://www.liquid.ai/blog/d1-open https://huggingface.co/LiquidAI/d1-3B https://huggingface.co/LiquidAI/d1-omni-600M Copilot local sandboxing reaches general availability with MXC: GitHub made local sandboxing generally available in Copilot CLI, the Copilot app and VS Code Agent Host sessions. Microsoft Execution Containers maps policies to native controls on Windows, macOS and Linux, restricting tool access to files, networks, credentials and other capabilities. Enterprise policies can require boundaries developers cannot weaken. Tool isolation is separate from model choice. Local sandboxing is included with Copilot at no additional charge, while containment options differ by operating system. Microsoft distinguishes process containers on Windows 11, macOS and Linux; session and WSL containers on Windows 11 only; and experimental MicroVMs on Windows 11/Linux. Policies cover processes, files, networks and UI resources, outside the agent’s control. Windows process containers can report activity: Learning mode blocks and records ungranted access, while Permissive mode allows and records it. Intune process-container management and Entra/Agent 365 extensions are still forthcoming. Summary sources: https://github.blog/changelog/2026-10-07-local-sandboxing-for-github-copilot-now-generally-available/ https://blogs.windows.com/windowsdeveloper/2026/10/07/microsoft-execution-containers-policy-driven-containment-for-ai-agents/ Copilot CLI can discover running Ollama models: Copilot CLI 1.0.94-0 adds /model discovery for running Ollama models. Users review the provider and endpoint, then explicitly add a model with or without switching. Ollama and models must already be installed; tool calling and streaming are required. Local selection neither enables offline mode nor disables GitHub telemetry. The CLI reference supports OpenAI-compatible endpoints, Azure OpenAI and Anthropic, configured through environment variables for the endpoint and model. Unauthenticated local Ollama needs no API key. Separately setting COPILOT_OFFLINE=true prevents GitHub-server connections; a remote provider still receives prompts and code over the network. The reference recommends at least 128k-token context; missing tool calling or streaming produces an error. Separately, the app documentation marks custom providers as public preview. GitHub sign-in is required, but using only an external provider needs no Copilot plan. Providers are added under Model providers in settings; their models appear alongside GitHub-hosted options. Credentials stay in the system credential store and are not shown in the UI. Summary sources: https://github.blog/changelog/2026-10-07-discover-local-models-in-github-copilot-cli/ https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/use-byok-models https://docs.github.com/en/copilot/how-tos/github-copilot-app/use-byok-models GitHub expands AI secret detection: GitHub’s fine-tuned classifier detects credentials from code context. Existing AI password alerts upgrade automatically and remain included with GHSP/GHAS. AI push protection is in private preview; Copilot /security-review classifier checks are coming to private preview. Both will consume AI Credits. The latter stay off by default; running a review does not enable them. The alert reference places AI findings with generic patterns such as private keys in the Generic list, which can include more false positives or test secrets. Its 5,000-alert repository cap includes closed alerts. AI-detected secrets show only their first location, and generic alerts are excluded from security-overview summary views. Billing documentation distinguishes product, SKU and bundled AI-credit budgets. Ordinary budget notifications alone do not halt overspending; an available stop-at-limit control must be enabled. User-level budgets always enforce a hard stop. Usage can count toward multiple applicable budgets, and exhaustion of any budget with stopping enabled can block further use. The application card also warns that AI secret detection can miss credentials, including secrets in test code. It describes analysis of repository-committed text without direct user interaction with the model. The system verifies that returned secret strings exist in the input before presenting AI-marked alerts. Summary sources: https://github.blog/changelog/2026-10-07-purpose-built-model-for-leaked-secret-detection/ https://docs.github.com/en/code-security/concepts/secret-security/about-alerts https://docs.github.com/en/billing/how-tos/set-up-budgets https://docs.github.com/en/enterprise-cloud@latest/code-security/responsible-use/security-and-quality-ai-features NVIDIA describes governed agent validation with AI-factory digital twins: NVIDIA describes validating supported configurations and software integrations in DSX Air before hardware arrives. Agents return evidence and recommendations; production changes remain subject to policy and human gates. The example’s models run on Brev GPUs outside the simulation. Performance and power estimates still require suitable models or measurements. The topology reference supports a drag-and-drop network editor or JSON/DOT imports defining nodes, attributes and links. Existing topologies export as JSON; the API also imports and exports JSON. A simulation can include a zero-touch provisioning script accessible to nodes over its enabled out-of-band management network. Simulation management saves all-node checkpoints by default when stopping and can restart from a chosen state. Reverting a topology that already has checkpoints deletes them all before re-enabling editing. History records timestamps, actors, categories and descriptions for state transitions, configuration changes and errors. Stopping releases compute resources; favorited checkpoints are protected from automatic deletion. The linked video-search implementation first collects the scenario, events and objects from a person, then incorporates organizational-document retrieval. Its reporting tool combines video analysis and retrieved context into structured reports with timestamps and citations, optionally letting users confirm or edit the prompt beforehand. The RAG Blueprint handles embedding, reranking and vector search for document retrieval. Summary sources: https://developer.nvidia.com/blog/validate-ai-factory-changes-with-digital-twins-and-ai-agents/ https://docs.nvidia.com/dsx-air/custom-topology/ https://docs.nvidia.com/dsx-air/simulation-management/ https://developer.nvidia.com/blog/integrating-context-aware-video-ai-agents-into-enterprise-workflows/

https://eaglhuang.github.io/AI-learning-notes/daily/2026-10-07/en/ 2026-10-07T12:00:00+08:00

OpenAI reports Ironclad agent results: OpenAI and Ironclad evaluated 11 contracting workflows with 8–50 criteria each. Tasks cover nondisclosure agreements, procurement approvals and jurisdiction-dependent clauses. GPT-6 Astra (Max) averaged 55.0%, versus GPT-5.6 Sol (High) at 41.6%; estimated attempt times were 19.2 and 37 minutes. OpenAI says timings were simulated from assumed speeds, not measured customer savings. Shorter source-limited summary. Summary sources: https://openai.com/index/advancing-computer-use-with-ironclad/ Simon Willison documents Codex-assisted Datasette tracing: On October 6, Simon Willison said he asked Codex to work out how to run Parseable and receive Datasette traces, then wrote the working instructions himself. His local example uses Parseable 3.2.4 with Datasette 1.0a41, local storage, an exporter endpoint, a service name and a stream. After browsing Datasette, Parseable displays requests and their SQL-query traces; the notes also give alternate-port settings. Parseable defaults to ports 8000, 8001 and 8002; changing them also requires updating Datasette’s exporter endpoint. Datasette’s changelog dates OpenTelemetry support to September 24, covering HTTP requests, database queries and startup, with metrics including query latency and write queues. Official documentation requires launching through opentelemetry-instrument: direct Datasette startup produces no telemetry, and the default batch processor exports every five seconds. The documentation adds that omitting OTEL_SERVICE_NAME files traces under unknown_service, and shutdown flushes the final batch. Summary sources: https://simonwillison.net/2026/Oct/6/datasette-parseable-opentelemetry/ https://til.simonwillison.net/datasette/datasette-parseable-opentelemetry https://docs.datasette.io/en/latest/changelog.html#opentelemetry https://docs.datasette.io/en/latest/internals.html#turning-tracing-on Atlassian expands its OpenAI partnership: On October 6, Atlassian and OpenAI expanded their partnership to power agents in Rovo and Atlassian’s platform with OpenAI models and Teamwork Graph context connecting people, projects, documents and decisions. OpenAI reports that over 3,000 Atlassian developers use Codex in terminals, editors and code review. The agreement extends a collaboration begun in 2023 and expands Atlassian’s access to OpenAI frontier models. Atlassian lists MCP-based enterprise context as available now; autonomous work pickup, test execution, session synchronization and multi-agent orchestration remain design plans with human checkpoints. The announcement also lists DX as a currently available tool for observing developer impact. It describes recording requirements, agent progress and decisions in Jira. The existing plugin announcement describes selecting Jira work, Confluence content or people in prompts, plus a pinned Home for resuming work. Home gathers assigned Jira work, recent Looms, projects and Bitbucket pull requests awaiting review; selecting an item opens a conversation with a prefilled prompt. That page says access remains subject to the user’s Atlassian permissions and organizational controls. Summary sources: https://openai.com/index/atlassian-partnership/ https://www.atlassian.com/blog/company-news/atlassian-openai-strategic-partnership https://community.atlassian.com/forums/discussion/3288784/bring-your-atlassian-work-into-chatgpt-and-codex Anthropic expands cyber verification: Anthropic divided CVP into Defense, Red Team and Specialized Access. Defense covers incident response and vulnerability analysis; Specialized Access is reserved for a limited set of organizations authorized to test high-risk safety systems. Its Opus 5.5 evaluation ran five attempts on ten scenarios: Defense blocked 46 trials; Red Team blocked none and completed 34. Red Team currently admits organizations only and permits testing solely on authorized systems. Separate official security requirements mandate an actionable security contact, prohibit shared sign-ins and require enabled user profiles so requests are attributable to named users or workloads. Anthropic may request continuing evidence of compliance with those requirements. Suspected breaches or misuse involving CVP access must be reported to Anthropic within 72 hours, or within 24 hours for a security incident. Shorter source-limited summary. Summary sources: https://www.anthropic.com/news/cyber-verification-program https://support.claude.com/en/articles/17202708-cyber-verification-program-security-requirements Google launches EmbeddingGemma 2: Google released Apache-2.0 EmbeddingGemma 2 with 740M parameters, a 270M text-only configuration and optional vision and audio encoders. The announcement lists 768-dimensional output, reducible to 512, 256 or 128 dimensions. Google reports that quantized weights on Pixel 11 Pro use as little as roughly 191MB active RAM for text-only and 567MB for the full model. The official product page describes a shared embedding space for text, images, video and audio, supporting cross-modal search and RAG retrieval with an 8K-token context. Its examples include finding images with text and video moments with voice memos; it lists audio inputs up to 5.5 minutes. Shorter source-limited summary. Summary sources: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ https://deepmind.google/models/gemma/embeddinggemma/ Mistral Large 4 enters public preview: Mistral’s October 6 changelog announces Large 4 in public preview. It describes a general-purpose multimodal mixture-of-experts model with a one-million-token context, while saying open weights are still forthcoming. Shorter source-limited summary. Summary sources: https://docs.mistral.ai/resources/changelogs OpenAI shares mathematical results and research details: OpenAI released internal-model mathematical results, including many Lean formalizations, ten reasoning summaries, attempted-problem counts and compute estimates. The company estimates average compute per result equivalent to roughly three hours of ChatGPT Pro thinking and plans further formalizations. The announcement also provides paper-revision and citation procedures. In its same-day statement, independent advisory group AGMAI said its role neither judges the results’ impact nor endorses OpenAI’s production process. The group considers release the beginning of human understanding and incorporation into mathematical knowledge, with assessment belonging to the mathematical community. AGMAI’s September 29 recommendations call for model, prompt, reasoning-summary, time and compute-cost disclosure, explicit formalization status, and explanations of problem selection and failed attempts at comparable problems. The group also advocates lab funding for human understanding while keeping those activities community-led. Shorter source-limited summary. Summary sources: https://openai.com/index/sharing-ai-progress-in-mathematics/ https://agmai.org/ https://agmai.org/general-sep29/

https://eaglhuang.github.io/AI-learning-notes/daily/2026-10-06/en/ 2026-10-06T12:00:00+08:00

ReviewBench: An open benchmark for AI code review: GitHub launched ReviewBench, a reproducible benchmark for AI code reviewers. It models real pull-request distributions and combines multiple sources of reference findings with engineer validation. Severity, precision, and recall help separate missed issues from noisy comments. For teams adopting review agents, it offers an evaluation starting point, but an offline result still needs validation against the team's own repositories. CodeMender updates (v0.12.0): Google released CodeMender v0.12.0 with optional hidden-file scanning and a larger default discovery limit, from 500 KB to 2 MiB. Fixes are constrained from making unrelated configuration changes, high-confidence findings require viable exploit evidence, and token accounting now covers the whole session. These changes address scan coverage, repair boundaries, and cost visibility; hidden-file scanning is opt-in and no quantified false-positive improvement was supplied. Secret scanning adds detectors for Lovable, Supabase, and more: GitHub expanded secret scanning to cover Lovable API keys, Pydantic observability and AI Gateway credentials, and two Supabase token types. Lovable also joined the issuer-notification program for public repositories. The update is relevant when moving AI-built prototypes toward production, but additional detectors do not guarantee complete secret coverage or automatic remediation, and partner notifications apply only within their supported scope. October 5, 2026: Cortex Agent code execution tool (General availability): Snowflake made the Cortex Agent code execution tool generally available. Once enabled in an agent specification, it can execute generated scripts for data processing, calculations, visualizations, and Python-based skills. Default sandbox access is limited to current-session data. This offers a managed execution option for data agents, but the news is its move from preview to GA, and the default isolation description should not be generalized to every configuration. Announcing Amazon Nova 2.5 Sonic with improved reasoning for voice agents: AWS launched Nova 2.5 Sonic in Bedrock, claiming improvements in reasoning, instruction following, tool use, and latency. The accompanying Strands Bidi Agents GA supports multiple voice-model providers, session and tool-call tracing, and background reconnection. For developers, observability and connection management matter alongside model quality. Performance claims remain vendor-reported, and reconnection does not remove the underlying model's session limit. Our approach to EU text provenance rules: OpenAI opened optional text watermarking for selected API models, with the API default remaining off. Eligible ChatGPT and Codex text output in the EU is scheduled for rollout over coming weeks, while detector access initially requires approval. Developers should distinguish available opt-in controls from future deployment plans, and treat detection as a limited provenance signal rather than proof of authorship or definitive evidence of AI use. ChatGPT plans visual-ad testing and expands measurement tools: OpenAI announced a visual ad format for ChatGPT, with a limited US test planned for later this month. Ads will be labeled and remain separate from generated images. The company also expanded conversion-data and attribution integrations and is developing brand-suitability evaluation pilots. For AI product teams, the update illustrates new commercial touchpoints around generation; attributed conversions should be distinguished from incremental impact. Partner case studies are not a guarantee of results for every advertiser, and the new format is not yet broadly available.

https://eaglhuang.github.io/AI-learning-notes/daily/2026-10-04/en/ 2026-10-04T12:00:00+08:00

Simon Willison argues for default hard spending limits: Willison argues that agents make paid APIs and cloud services easier to deploy, increasing exposure to runaway usage. He advocates default spending limits that stop service at the budget ceiling, with an explicit choice required to remove them. Prime Inference opens its model-serving platform: Prime Intellect launched Prime Inference with serverless endpoints, reserved capacity and an OpenAI-compatible interface. Its engineering account explains cross-datacenter failover, caching and scheduling for long-context agent traffic. BootLoops pairs verifiable calculations with domain judgment: Physicist Matthew Schwartz describes BootLoops, a harness that gives agents checkable computational tools for cross-domain research. His central lesson is that technically correct results can still lack scientific value; domain experts must guide the questions and inspect the conclusions. OpenAI reports reasoning-extraction activity and mitigations: OpenAI reports disrupting coordinated attempts to extract protected model reasoning and strengthening protections across users and models. The company says the activity did not directly access stored user conversations. This September disclosure concerns activity observed in July. GPT-6.1 Sol targets lower-cost agent workflows: OpenAI introduced GPT-6.1 Sol for coding, computer use and professional tasks. Standard API prices are $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens; the company reports stronger performance than GPT-6 Sol. Sonnet 5.5 emphasizes efficiency at unchanged token prices: Anthropic says Sonnet 5.5 generates output over 30% faster and costs up to 30% less per task in its tests. Input and output prices remain $2 and $10 per million tokens; the savings chiefly come from lower token usage. Holo4 brings screen and tool interactions into one model: H Company released Holo4 in 27B dense and 35B-A3B MoE versions for GUI interaction, code execution, MCP and APIs. The team also publishes public-benchmark trajectories so developers can inspect the actions behind its reported scores. Codoku offers renewable program-reasoning tests: Codoku asks agents to fill incomplete programs under global static and dynamic constraints, with fresh puzzles generated on demand. The authors evaluated five frontier models on 300 puzzles and report that even proprietary models solved only about half of the large puzzles. LEGO-Anything separates valid 3D artifacts from accurate scenes: LEGO-Anything has coding agents iteratively write and execute Blender programs from single images. Its benchmark uses 208 images from 104 scenes and separates artifact validity, geometry and appearance, exposing a substantial gap between usable outputs and accurate reconstruction. GitHub reports 24 Android flaws from staged AI audits: GitHub Security Lab describes staged auditing workflows and reports finding and reporting 24 Android-app vulnerabilities. The account explains why models can misjudge impact or exploitability, and why validation and researcher review remain necessary.