#Deployment/Engineering

AISemiAnalysis tested offloading DeepSeek-V4.1-Flash's Engram embedding table from HBM to host DRAM and to local SSD. On B200 configurations, DRAM delivered more total tokens per dollar and higher P90 interactivity than SSD at every measured point. The report concludes SSD offloading is likely not worth the tradeoff for production serving in its unoptimized setup.

AIGoogle's Envisioning Studio, with Google Labs, co-developed custom Google Flow tools with designers Jane Wade and Sergio Hudson ahead of New York Fashion Week. Wade's Styling Suite let her style runway looks on digital models before producing physical samples, while Hudson's Runway Visualization helped him stage his show within a tight budget. The source says the tools are built with natural language and no coding experience.

AILiquid AI released LFM2.5-VL-3B-DSpark, a speculative-decoding draft model for its LFM2.5-VL-3B vision-language model. The source reports decoding up to 2.66× faster on a single H100 with SGLang, up to 3.13× on Apple M5 Max with MLX-VLM, and up to 2.14× on Apple M3 Ultra with llama.cpp, with output unchanged under greedy decoding.

AIHuawei unveiled the Atlas 960E superpod, which links up to 4,096 NPUs using near-packaged optics (NPO) and claims eight exaflops at FP8 precision and up to one petabyte of high-bandwidth memory. Its Hi-ONE engine provides 7.2 terabits per second of transmission capacity, and Huawei says 5,500 engines replace 48,000 conventional 800G optical modules, cutting power consumption by more than 550 kilowatts. Huawei has proposed its NPO implementation agreement to the Optical Internetworking Forum, though it has not yet become a standard.

AIInferact, with Red Hat AI, NVIDIA, and Huawei, co-led an optimization effort that makes Kimi K3 on vLLM 2.2–2.8× faster. The work spans scheduling, KDA state handling, and custom MoE kernels. The vLLM project's background post cites those throughput gains on a B300 benchmark against v0.27.1 and links a technical deep dive.

AIHamel Husain and Shreya Shankar's FAQ explains AI evals as tests of whether an AI system does what users and the business want. It recommends starting with error analysis on at least 30 traces, then turning recurring failures into binary code-based checks or LLM judges validated against human labels.

AIAnthropic is partnering with Accenture, through its specialist AI business Faculty, on independent evaluation of frontier models, including red-teaming, alignment assessments, and safeguard testing. Anthropic and Accenture each expect to invest at least $1 billion in this capacity over five years. The source says embedded evaluators would have employee-comparable access, but standards for access and reporting, and a settled funding system, do not yet exist.

Why it matters: The source ties a new evaluation arrangement to an unresolved question of who funds and sets standards for independent AI evaluators, which is useful context for governance debates.

AIGoogle is restricting access to its Gemini 2.5 models to users who have actively used them in the past, to keep performance reliable. The models are not deprecated and remain available through the API until further notice. For new projects, Google recommends its latest models, 3.5 Flash-Lite or 3.8 Flash.

AIIan Johnson (@enjalot) describes his job shifting from UX to AX, meaning agent experience. His work now includes researching agent transcripts, building more ergonomic agent APIs, and empathizing with agents, all to build software for people.

AIA global fintech scaled its AI coding agent traffic by running the GLM-5.2 model on Together AI's Dedicated Model Inference, after capacity planning failed to keep pace with unpredictable engineering-hour bursts. The customer gained self-service endpoint provisioning, a metrics API for diagnosing queuing, and live configuration changes that shipped with zero downtime. The setup runs dozens of B200 GPUs at 256K context across multiple replicas.

AIvLLM now supports NVIDIA hardware video decoding through PyNvVideoCodec, moving video decoding off the CPU so multi-GPU video captioning can scale to 8 GPUs. In benchmarks on 8xH100 GPUs, GPU-based decoding provides more than double the throughput of the CPU-based decoder for Qwen/Qwen3-VL-8B-Instruct with 8 single-GPU vLLM replicas. The functionality is included in standard CUDA vLLM releases, and PyNvVideoCodec==2.0.4 is required for custom installations.

AIxAI released Grok Voice Transcribe 2.0, a speech-to-text model that is twice as accurate as Grok Voice Transcribe 1.0 at the same price, and ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard. Batch transcription costs $0.10 per hour of audio and streaming $0.20 per hour, with diarization, timestamps, and key terms included. Existing Speech-to-Text API integrations gain the improvement with no code changes, and developers must pin grok-voice-transcribe-1.0 to stay on the older model during the transition.

AIOpenAI has released ChatGPT for Word, letting users access ChatGPT and Codex directly inside Microsoft Word. The post says ChatGPT for Excel and PowerPoint has been growing rapidly, and Word completes that set. The quoted ChatGPT post adds that the tool can draft from notes, rewrite paragraphs, proofread, suggest edits, and flag formatting issues.

AIOpenAI launched 47 community-built plugins for legal work in ChatGPT, created by legal experts from LegalQuants, Skills.law, and LECG rather than white-labeled by OpenAI. The plugins are already live in the ChatGPT plugin store, alongside 26 partner-built plugins from companies including Thomson Reuters, Harvey, Legora, and iManage.

AIMeta is rolling out Muse for Mac today, a personal agent that can complete tasks directly on the user's computer with explicit permission. Examples include organizing the downloads folder, finding lost files, and summarizing messages and notes, with more capabilities coming soon.

AIOpenAI has introduced Astra for Law, a new offering powered by GPT-6 Astra with tools, settings, and context for lawyers and legal technology firms. The offering includes data privacy as a core feature, plus 26 partner-built plugins and 47 community plugins for legal work in ChatGPT.

Why it matters: The offering targets legal work specifically, with partner and community plugins and data privacy as a core feature, which shows how frontier models are packaged for professional sectors.

AIGoogle Research is sharing an experiment that lets educators generate custom, guided STEM simulations tailored to their curriculum using generative UI. It is releasing a sample library of over 30 English interactives for physics, chemistry, biology, and math, all AI-generated and reviewed by teachers. Schools using Google Workspace for Education can sign up through the Google for Education Pilot Program to give feedback.

AIGoogle teamed up with Speakeasy to ship new Google GenAI SDKs for its Interactions, Agents, and Webhooks APIs. Speakeasy has also open-sourced its OpenAPI generator suite used to generate the SDKs.

AIGoogle AI Studio announced that Speakeasy is open sourcing its OpenAPI client generation suite under AGPLv3, following a May 2026 vendor shutdown that disrupted Google's SDK pipeline. The suite covers SDK generation for 7 languages, an agent-native CLI generator, and a documentation MCP server generator. Google says its pipeline now serves six targets with roughly one engineer maintaining it.

AIAnthropic's Claude Code now runs projects from a single conversation, where Claude directs parallel threads that keep working after the user closes their laptop. The feature is in beta for select Pro and Max users in cloud sessions, with wider availability for all Claude users promised soon.

Why it matters: The quoted launch replaces scattered sessions with one coordinator that runs parallel threads in the background, a workflow change worth weighing for complex projects.

AIAnthropic's Claude Code is rolling out Projects on desktop and web, in beta for select users. A project splits work into threads, runs them as parallel cloud sessions, passes context between them, and keeps running after the user leaves. The author says Claude keeps context across tasks and can give an aggregated status update on request.

Why it matters: The post explains how Projects shifts work from managing single sessions to coordinating many parallel tasks, a change that affects how Claude Code users plan and track their work.

AIGoogle AI Studio released antigravity-preview-09-2026, an updated harness for Gemini managed agents, now live in the Interactions API and AI Studio and running on Gemini 3.8 Flash. The release adds a Files API for moving data into and out of the agent's sandbox and a Credentials API that stores secrets encrypted so the model never sees them.

Why it matters: The post shows what changed in the agent harness and how the new Files and Credentials APIs keep secrets out of the model's context, useful for developers building agents.

AISkillsync, a Y Combinator W26 company, launched a tool that converts AI chat sessions between coding agents, including messages, reasoning and tool calls. The conversion engine txcript is open source, and the local-first app runs on Mac with a CLI and MCP support. Only sessions shared into team workspaces leave the user's machine.

AIUnsloth Desktop now supports multi-user accounts, alongside a revamped Docker image and custom Jupyter Notebook with custom themes, titles, and expandable cells. The update adds RDNA1+2 support, ARM64 Windows CUDA support, faster GRPO, and FP8/INT8 image diffusion support for 2x faster inference.

AISierra has become AIUC-1 certified after an independent audit by Schellman and testing by the Artificial Intelligence Underwriting Company (AIUC), a new standard for AI agents that tests resistance to manipulation and unauthorized access. Schellman found that Sierra met all applicable AIUC-1 requirements, and the technical evaluations recur at least quarterly with a full audit each year. The certification complements Sierra's existing SOC 2 Type II, ISO 27001, and ISO 42001 attestations.

AIAwni Hannun says Apple already offers MLX, a tool for efficiently running AI models on Apple hardware. The post is a short announcement with no benchmarks, prices, or version details.

AIReplicate invites users to try Pruna AI's p-video-2-pro model at the linked Replicate page. The post offers no further details on capabilities, pricing, or performance.

AIGoogle DeepMind has made its AlphaGenome Atlas freely accessible to researchers worldwide. The database is intended to help them decode the genetic causes of disease.