Customizing Vertical Voice Agents with ASR and LLM Fine-Tuning https://t.co/dBSfkUiKeX
Customizing Vertical Voice Agents with ASR and LLM Fine-Tuning https://t.co/dBSfkUiKeX
All news · Today5 items
All AI news
All collected and published items, ordered by time. Picks are a selected subset.
Tutorials
Tutorials, hands-on experience, usage tips, prompts and tool how-tos, and in-depth technical explainers.
NVIDIA AI · X17 AWS Machine Learning BlogPicks63 Natural language security queries at Orca Security using Amazon Bedrock
Orca Security partnered with AWS to build Discovery AI on Amazon Bedrock, translating natural language questions into validated Sonar queries across 900+ cloud infrastructure and application models. The architecture combines example retrieval and reranking, relevant schema selection, agentic generation, and validation, with domain experts continuously updating the example bank. Compared with its initial baseline, Orca reports 90 percent lower costs, 93 percent fewer input tokens, and 75 percent lower latency, with response times falling from approximately 60 seconds to 15 seconds.
Why it matters: The architecture offers a transferable approach to querying large schemas by retrieving validated examples, selecting relevant schema fragments, and feeding validation errors back into generation.
Anthropic · XSame news68 Anthropic explains building in-house source control and migrating its monorepo from git
Anthropic’s Engineering blog shares how it built a new version control system and fully migrated its monorepo from git in five months. In May, it merged around 2,500 code changes per day; in September, it merged more than 30,000 per day. That growth created bottlenecks in developer infrastructure built for human rather than agent scale: https://www.anthropic.com/engineering/redesigning-source-control
The same story is shown in the picks as 《Why Anthropic built Scion to replace git for its monorepo》
AWS Machine Learning BlogPicks63 Mixture-of-Experts kernel engineering on Amazon SageMaker AI
AWS explains three fused MoE optimizations used in Amazon SageMaker AI, reporting up to 10 percent higher throughput and up to 9 percent faster Time to First Token (TTFT). On H200, swapping matrix operands improved throughput by 2–4 percent, while routing FP8 tensor-quantized models to Triton improved throughput by 10 percent. Eliminating unnecessary Triton JIT recompilation improved TTFT by 9 percent at 8-way tensor parallelism; the post identifies availability in vLLM and the AWS LMI v28 container.
Why it matters: The optimization examples offer a transferable approach to MoE serving performance by examining matrix layout, backend selection, and compiler specialization as separate sources of throughput and latency overhead.
AWS Machine Learning BlogPicks75 Build long-horizon agents with Grok 4.7 on Amazon Bedrock
AWS demonstrates how to build and evaluate a tool-calling agent using Grok 4.7 on Amazon Bedrock, comparing four reasoning effort levels across three synthetic tasks. The agent answered 67 of 72 runs correctly; xhigh achieved 18 of 18, while low achieved 15 of 18 using about an eighth of the output tokens. The guide covers task design, automated grading, production monitoring, and operational controls, noting that these small-scale results are indicative rather than statistically conclusive.
Why it matters: Tasks with provably unique answers and repeated runs provide a reusable way to compare reasoning effort, while separate correctness and token measurements make the trade-off concrete.
AWS Machine Learning Blog56 How ShipStation Global built an AI IT help desk on Amazon Bedrock AgentCore
ShipStation Global built Deckhand, a Slack IT help desk agent on Amazon Bedrock AgentCore, to answer questions, analyze error screenshots, and create or update Jira Service Management tickets. The implementation guide explains its Strands agent, knowledge retrieval, MCP tools, and extraction of resolved Slack threads using DSPy’s Recursive Language Model. In its first three months, 812 of 1,221 agent conversations were handled in chat, while 409 became tickets; positive feedback rose from 82 percent in April to 96 percent in June.
Simon Willison40 Simon Willison Quotes Geoffrey Huntley on Enforcing Agent Use of a Diagnostics Library
Simon Willison quotes Geoffrey Huntley describing a pre-commit hook that checks the git diff for direct calls to observability modules that bypass his diagnostics library. When it finds one, the hook fails the commit and injects a prompt telling the agent to use the library. Huntley says the approach makes the agent “ten out of ten.”
宝玉 · X56 Downloading WeChat Channels Videos With Codex and REA
The author describes using Codex + REA to download a video playing in WeChat Channels with a simple prompt. One conversation round was insufficient; Codex configured a local proxy server, required several password entries to install certificates, and wrote a script to decrypt the downloaded video. The author presents this as a case study, notes that more mature download solutions likely exist, and suggests packaging the workflow as a Skill if it works reliably.
Alexandr Wang · X19 art cannot be created or destroyed, only remixed (pure code recreation of *the video* using muse wonder)
art cannot be created or destroyed, only remixed (pure code recreation of *the video* using muse wonder)
vLLM · X58 ReplaySSM reports decoding speedups and production use, with support already in vLLM
vLLM praises ReplaySSM work by @zwljohnny, @tri_dao, and @NVIDIA, saying Mamba2 standard decode and FlashInfer-backed MTP support are already in vLLM. Caching SSM inputs and updating the checkpoint only when needed reduces recurrent-state memory traffic. In the quoted post, @zwljohnny reports production use in GLM-5.3-Flash and Kimi-K3, availability in TensorRT-LLM, vLLM, FlashInfer, and SGLang, speedups of up to 1.47x for standard decoding and 2.1x for speculative decoding, and low-precision SSM state whose rounding error never propagates through the recurrence.
Anthropic · X60 Anthropic reports progress in reducing the effectiveness of distillation attacks against Claude
Anthropic says it has made significant progress over the last few months in reducing the effectiveness of distillation attacks against Claude. The company explains how in this post: https://www.anthropic.com/research/distillation-defenses
AWS Machine Learning BlogPicks67 Guide to accessing OpenAI GPT-6 models on Amazon Bedrock with global cross-Region inference
AWS explains how to invoke OpenAI GPT-6 Sol and GPT-6 Luna through Amazon Bedrock global cross-Region inference from Thailand, Malaysia, Singapore, Indonesia, and Taiwan, and GPT-6 Astra from Singapore. The guide covers Converse, OpenAI-compatible APIs, coding clients, IAM permissions, SCP exceptions, quotas, and monitoring. Billing and quota tracking remain in the source Region, while requests run in supported destination Regions; each output token consumes 10 tokens of quota.
Why it matters: The guide provides a transferable approach to configuring cross-Region inference permissions by scoping SCP exceptions to inference profiles while preserving Region controls for other services.
Simon Willison64 Simon Willison reviews Python 3.15 and adopts Python 3.11 features
Simon Willison reviews Python 3.15 and explains how Python 3.10’s end of support lets his open source projects adopt Python 3.11 features. Highlights include explicit lazy imports that address startup costs in his LLM CLI plugins, unpacking in comprehensions, and the Tachyon sampling profiler. He provides uv upgrade commands and identifies newly usable features such as tomllib, asyncio TaskGroup, and SQLite improvements, including removal of a Datasette compatibility workaround.
GitHub Blog · AI & ML49 How to Teach GitHub Copilot About Your Codebase
GitHub’s beginner guide explains four ways to tailor Copilot to a project: custom instructions for repository conventions, agent skills for repeatable procedures, custom agents for broader specialized tasks, and MCP servers for connecting to external tools and data. It recommends starting with a convention you often explain and adding it to `.github/copilot-instructions.md`.
Perplexity · XPicks67 How Perplexity built the Projects filesystem as a Git forge for agents
This is how Perplexity built the filesystem behind Perplexity Projects as a Git forge for agents. Each agent works in its own sandbox and commits to a shared repository where people and other agents can build on its work across sessions. Read more: https://t.co/jcrTptb2IW https://t.co/dXnhroO4mg
Why it matters: The shared repository provides a concrete example of how work from separate agent sandboxes can remain available for people and other agents across sessions.
AWS Machine Learning BlogPicks66 Run multi-node inference workloads using Ray LLM Deep Learning Containers
AWS shows how to deploy Qwen3.5-9B across two GPU nodes on Amazon EKS using Ray LLM Deep Learning Containers, with separate prefill and decode deployments behind one OpenAI-compatible endpoint. The walkthrough configures the application in YAML and uses NIXL over EFA to transfer the KV cache between nodes. It covers cluster provisioning, model invocation, transport verification, and resource cleanup.
Why it matters: The guide provides a transferable deployment pattern for separating prefill and decode, including how to verify the transport backend and avoid a silent fallback to TCP-speed communication.
AWS Machine Learning BlogPicks61 Deploy decision models with llama.cpp on Amazon SageMaker AI
AWS explains how to deploy GGUF models with its llama.cpp Deep Learning Container on Amazon SageMaker AI, then add a SemIf adapter that returns probabilities over declared options without generating answer tokens. The guide includes CPU and GPU deployment code. Across 231 JevBench tasks using Qwen3.5-4B Q4_K_M, the GPU configuration delivered 4.9 times faster p50 and 32.2 times faster p95 latency than the CPU configuration, with similar accuracy.
Why it matters: The deployment pattern shows how applications can score predefined choices without parsing generated prose, while keeping authorization, thresholds, and fallback rules under application control.
宝玉 · X51 BaoCut Adds Claude Design and Claude Slides Alongside Video Editing
Baoyu integrated reverse-engineered Claude Design and Claude Slides into BaoCut, an open-source Agent originally focused on video editing. He reports substantially improved design performance when GPT-6.1 Sol uses the Claude Design Skill, possibly because Claude Design defaults to Claude Design System Skill; standalone Skills would not support editing and annotation, so he integrated them into BaoCut. Users can download or run BaoCut locally, use a local agent CLI such as codex, and select the design skill during a session. https://github.com/JimLiu/baocut/commit/7aa049e8 https://baocut.com
宝玉 · X61 Baoyu uses REA and Computer Use to investigate CapCut subtitle templates
Baoyu describes using REA with Computer Use to investigate how CapCut renders dynamic subtitles, without knowing where their formats were stored. He had it operate CapCut, download Templates, and monitor network URLs and disk save locations. After downloading several dozen through clicks, Codex found clicking too slow and constructed URLs to download directly, ultimately downloading 2291 Templates.
宝玉 · X62 Using REA to Trace How Codex Attaches Browser Annotations
Baoyu describes using the open-source REA (Reverse Engineer Anything) MCP service to understand how Codex and the ChatGPT app attach browser annotations to the composer, without decompiling Codex itself. After REA traced the Electron app's code path, Codex built a test page that could open a comment box and prefill input via document.oai.annotation.request(), an OpenAI API the author had not known about. Cursor and Claude Code did not support a similar API.
The Verge · AI22 Testing Local AI With Hermes Agent and Qwen on an M5 Ultra Mac Studio
A Verge writer began running local LLMs on an M5 Ultra Mac Studio with 256GB of unified memory, installing the open-source Hermes Agent desktop app and the 125-billion-parameter Qwen 3.8 Flash Next model, about 105GB in size. Early tasks included a daily briefing via Telegram, reorganizing a 400-plus-game Steam library, and local data analysis; the writer is now working on automating laptop benchmark tests.
Hugging Face Blog34 How Ai2 Rebuilt Cluster Scheduling with GPU Time Budgets
Ai2 replaced its original priority scheduler with GPU time budgets, hierarchical fair allocation, and time-slice contracts, turning debates over how much GPU time each research project should get from case-by-case operational decisions into a transparent budget allocation process. Its cluster manages thousands of NVIDIA H100, B200, and B300 GPUs, with scales ranging from 88 to 1,024 GPUs, and demand from roughly 150 internal researchers has long been 2-3 times the available compute.
AWS Machine Learning Blog40 How Postman runs Agent Mode on Amazon Bedrock for 40 million developers
Postman runs Agent Mode on Amazon Bedrock, providing 40 million developers with an AI-native way of working that spans API testing, documentation, discovery, and implementation.
Simon Willison34 Simon Willison builds a Newsletters page for his blog using Codex voice mode
Simon Willison used the voice conversation mode in the Codex tab of the ChatGPT desktop app to develop a Newsletters page for his blog almost entirely by voice during a half hour spent cooking, with GPT-6 Astra High as the underlying model.
Hugging Face BlogPicks73 Hugging Face author trains six models in two days with ML-intern, spending about $103 in total
The author used ML-intern in HuggingChat, describing the requirements in prompts, to build six models in two days and publish them to the Hub, at a total compute cost of about $103. One project distilled the 9B prompt rewriter for Qwen-Image 2.1 into a 0.8B version that runs on a CPU, with a 99.7% valid-output rate and about one-quarter of the teacher model's token usage. The author published all seven prompts and recommends asking for baseline comparisons, smoke tests, and a budget cap.
Why it matters: The author built six models in two days with ML-intern and published all prompts and cost details, offering an approach others can adapt to their own training workflows.
NVIDIA Blog59 Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents
NVIDIA explains how its developers use frontier AI agents, including GPT-6 Astra, with Omniverse libraries to build and refine simulations through natural-language instructions, human review and iterative testing. Examples span warehouse scaling, cable routing, robot asset preparation, autonomous-driving workflows and sensor-based digital twin refinement. One warehouse simulation sustained 30 frames per second with 80 conveyors, 40 arms and 800 boxes during a two-minute test; a robot CAD example used estimated physical properties that were not calibrated against the real robot.
AWS Machine Learning Blog22 How to isolate GPU clusters across teams and share them fairly with Amazon SageMaker HyperPod
Amazon SageMaker HyperPod published a multi-tenant reference architecture that lets multiple teams share the same HyperPod EKS cluster.
NVIDIA AI · X18 How H-Company Optimizes VLM Serving for Computer Use Agents With NVIDIA Dynamo https://t.co/gixR8bJLTj
How H-Company Optimizes VLM Serving for Computer Use Agents With NVIDIA Dynamo https://t.co/gixR8bJLTj
雷锋网58 Zhijian Dynamics' Feng Zongbao: Achieving robot-built robots with 600 trajectories under a 0.5-millimeter gap
At the IROS 2026 venue, Feng Zongbao, head of reinforcement learning at Zhijian Dynamics, presented a solution that enables a robot to autonomously tend two CNC machine tools, completing grasping, loading, insertion, removal, and placement under the condition of only a 0.5-millimeter gap between the workpiece and the chuck, with training using only 600 real robot trajectories.
AWS Machine Learning Blog34 How Amazon Quick and Amazon Bedrock use real-time ACL checks to redesign RAG access control
Amazon Quick and Amazon Bedrock Knowledge Bases have added real-time ACL checks at query time, alongside pre-retrieval ACL filtering. They verify permissions directly with authoritative data sources, avoiding reliance on ACL data that may be stale or incorrectly mapped.
AWS Machine Learning Blog36 How Cornerstone OnDemand cut database diagnosis time by 78% with Amazon Bedrock
Cornerstone OnDemand built Orion AI, a multi-agent system based on Amazon Bedrock and the open-source agent orchestration framework Strands Agents. It reduced database diagnosis time from 45 minutes to 10 minutes, a 78% reduction. It also compressed manual lifecycle operations from more than 10 steps to one interaction, a 70% reduction, and cut median redundant alerts by 65%. A three-person team delivered it in six months.
AWS Machine Learning Blog22 AWS publishes a six-week AI builder program to bridge the gap between AI knowledge and practical skills
AWS shared a structured six-week program in which business staff without engineering backgrounds build working AI prototypes using Amazon Bedrock AgentCore, the Strands Agents SDK, and Amazon Nova models, with about four hours of work per week.
AWS Machine Learning Blog38 How to automate fixes after AWS DevOps Agent investigations with AWS Lambda Durable Functions
AWS demonstrated an automated remediation workflow. Amazon EventBridge receives an investigation-complete event from AWS DevOps Agent, then AWS Lambda Durable Functions calls Amazon Bedrock to analyze the root cause and propose fixes from a pre-approved allowlist of tools.
AWS Machine Learning Blog22 How to make the business case for agentic automation beyond hours saved
AWS proposed the Agentic Value Model to measure value that traditional ROI models miss in agentic automation, across four dimensions: hours saved, exception handling, decision quality, and resilience to change.
Boris Cherny · X54 Boris Cherny explains how he prompts Claude with goals, effort, and verification
Boris Cherny recommends talking to Claude the way you would a coworker. For most tasks, he says there is no need for overly scaffolded or prescriptive prompts; give Claude a goal, and it will figure it out. He says prompts mattered a lot in the Sonnet 3.5 days. Now, he emphasizes communicating what you want it to do, how much effort it should spend, and how it should verify that it did the right thing.
AWS Machine Learning BlogPicks60 How to build a context-aware AI assistant on AgentCore and OpenClaw
An AWS blog post explains how to run the open-source agent system OpenClaw on the Amazon Bedrock AgentCore runtime to build Sprout, a personal assistant with long-term memory.
Why it matters: The source provides reusable methods for container wrapping, namespaces, and prompt-cache ordering, helping readers build an agent with long-term memory.
Simon Willison22 Processing OpenTelemetry traces with Parseable and Datasette
After Datasette 1.0a41 added OpenTelemetry support, the author used Codex to get Parseable running and receive traces from Datasette.
Boris Cherny · X24 I asked Opus 5.5 to make an interactive website companion for the latest @AcquiredFM Home Depot episode. It came out pretty nice! Water color illustrations all drawn by Claude, too: https://claude.ai/artifact/RdE6uE7WozcP2MLfx3BDky Episode here: https://www.acquired.fm/episodes/home-depot
I asked Opus 5.5 to make an interactive website companion for the latest @AcquiredFM Home Depot episode. It came out pretty nice! Water color illustrations all drawn by Claude, too: https://claude.ai/artifact/RdE6uE7WozcP2MLfx3BDky Episode here: https://www.acquired.fm/episodes/home-depot
AWS Machine Learning Blog36 Build a voice-based travel assistant with Amazon Bedrock AgentCore, Managed Knowledge Base, and Nova Sonic
Amazon Bedrock AgentCore, Amazon Nova 2.5 Sonic, and Amazon Bedrock Knowledge Bases can be used to build an airline voice travel assistant. Travelers can use voice to check itineraries, change seats, update meal preferences, and transfer to a human customer-service agent.
AWS Machine Learning Blog22 Management and governance best practices for Amazon SageMaker HyperPod
Amazon SageMaker HyperPod can connect clusters to projects through SageMaker Unified Studio, letting team members launch training and fine-tuning workloads in workspaces while retaining governance controls at the infrastructure level. The article proposes a four-layer control model—organization, project, cluster, and workload—and explains how to design identity, capacity, and observability policies across those layers. Infrastructure teams retain control of cluster operations while ML teams use approved compute in project contexts.