#Tutorial/How-to

AIFractalyze optimized Qwen3-Omni on vLLM-Omni for a single RTX 5090, using AWQ-4bit at batch size 1 with text prompts. In its tests, time to first audio dropped from 213ms to 23ms compared with stock vLLM-Omni. vLLM hopes the optimizations will be contributed upstream to benefit more users.

AIThe PyTorch Accelerator Integration Working Group released updates on its H1 2026 progress toward standardizing how new hardware connects to the framework. Key workstreams include the Cross-Repository CI Relay (CRCR), which automatically reports downstream backend test results to a shared dashboard, and refactored test suites that decouple PyTorch's 600,000-plus tests from specific accelerators.

AIA native word-level transcription model outputs structured, timestamped arrays of word, spacing, and audio_event tokens directly from audio input, without a secondary forced-alignment pass. Audio events such as laughter or applause are tagged separately, which the source says helps with captioning, searchable archives, and highlight identification. The source notes Scribe's word-level transcription supports up to 5 independently transcribed channels.

AIReliable AI agent systems need deterministic policy checks, not just better prompts or stronger models, because a model's proposed action can succeed at the API level while still updating the wrong account. The article recommends separating the model's proposal from a policy service that checks actions before execution and records an audit trail. It also advises treating agent context as untrusted input, using narrow capabilities instead of broad tokens, and building in stopping rules and idempotent recovery.

AIBrian's Grok Bot method treats each bot like a new hire: define the role, test it on text first, run three trials, escalate based on evidence, and add a second agent only after a bottleneck appears. Each bot's role is defined by five fields: a real name with a short label, a one-line job tied to an outcome, what it owns, its inputs, and what it may do freely versus what it must ask before doing. The post frames an Agent Team as the final result of this process, starting with one coordinator and three specialists.

AIThe /break-ui Skill, added to the Skills For Designers and Engineers repo with 43K stars and 1.9M installs, plays the most annoying real user to stress UI components with worst-case but realistic data. It targets bugs manual testing misses, such as "1 members" pluralization errors, zero-value "0 seconds ago" rendering, cross-timezone date shifts, and emoji or CJK names breaking initials logic. The skill reports issues before fixing them, and only changes the data, never the component.

AIUber uses a control plane and data plane gateway to automatically convert its internal APIs into MCP tools, with 800+ MCP servers and 5,000+ tools hosted. The design includes an AutoCrawler that generates tool descriptions with an LLM, a default-disabled discover-not-expose security model, and techniques such as Omni MCP, Response Projection, and Code Mode to limit context bloat.

AIThe article argues that improving an AI setup can sometimes mean giving the AI fewer rules to follow, based on the author's experience across a million words of failed drafts. The source text provided is mostly paywall and subscription material, so no further specific figures, products, or benchmarks can be verified.

AIOpenRouter's shell and bash tools, along with those from OpenAI and Anthropic, run an agent's commands in provider-managed sandboxes during the same API request, so developers don't provision or patch containers. OpenRouter's tools are in beta, with sandbox time billed at $0.0001 per second and a 30-second minimum for a new or sleeping container. The article compares the four providers and notes that self-run sandboxes remain better for custom base images, GPU work, or multi-hour sessions.

AIPerplexity's Computer can build custom vertical AI apps, such as one that guesses an image's location using 3D and satellite views. The example app was built with the Perplexity SDK for web search, local place lookups, and visual clue extraction, and it uses Cesium for the 3D globe and satellite imagery.

AIKling AI shares behind-the-scenes details of its short film "The Beat," which has passed 5 million impressions across social platforms. The post says the film used Kling 4.0 features including a 30-second continuous shot, Omni Reference supporting up to 15 multi-modal references, Multi-Keyframe control for up to 10 keyframes, and 10-bit HDR output.

AILangChain says its coding agent costs fell significantly for a second straight month after adopting three steps. The steps are cost visibility through LangSmith tracing, per-user cost caps via its LLM gateway, and harness optimization such as model routing in its open-source OpenSWE cloud agent harness.

AIPerplexity's Computer can now generate inline visualizations when users ask it to "Visualize" a topic, producing interactive widgets and animations within the thread. The feature is best used on Standard or High effort, and an example given is an inline 3D cutaway of a jet engine.

AIPixVerse says users can describe a character to their AI agent and use the PixVerse plugin to generate a full design sheet. The post provides no further details on features, pricing, or availability.

AISebastian Raschka released round six of his Reasoning from Scratch series, introducing Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO) with an implementation. The video covers accuracy and format rewards, DeepSeek-R1 training, and GRPO versus PPO, then walks through a training loop and evaluates checkpoints on MATH-500.

AIReplit chat now generates interactive charts when users ask Replit Agent to visualize data. Users can also choose GPT-6.1 Sol from OpenAI or Claude Sonnet 5.5 from Anthropic when building with Agent, or stay in auto mode. Jev is available through Replit AI Integrations for classifying content, routing requests, and scoring leads without managing API keys.

AIvLLM changed its KV cache layout to block-major BLHNC, cutting transfer descriptors about 10x and halving mean KV transfer time on NVLink. The original slowdown came from fragmented KV layout that split one 200K-token request into 32K tiny copies, making NVLink slower than InfiniBand.

AIPrime Intellect reports that DEP8 provides about 5x the prefix-cache capacity of TEP8 on the same GPUs. The post argues that fast KV retrieval alone does not ensure fast first tokens, since cached KV often sat ready while requests waited to join a batch. Halving the prefill budget reduced median queue wait time and time to first token (TTFT).

AIPrime Intellect says long-context agent serving depends on retaining history, scheduling new work, and moving cached state efficiently. It optimized three paths separately: prefill topology and scheduling, compressed KV with a fused attention kernel, and a transfer-friendly cache layout.

AIVercel has added Jev to the AI SDK for Python, and the team tested it in two experiments: detecting whether typed text is Python or English, and writing Python one decision at a time. The main post is a short endorsement praising a writeup about Jev and Python.

AIBaseten tested the MetaInfer skills-only approach by having Claude Code build an inference engine, VibeQwen, for Qwen-3.6-35B-A3B in NVFP4 on a single B200. On single-stream text, VibeQwen decoded 90% faster than a tuned vLLM 0.25.1 deployment (1,792 vs. 943 TPS) and cut time to first token from 28 ms to 12 ms, with a 71% throughput gain at concurrency 32. The author notes this was an outcome-focused run that allowed some numerically different outputs as long as accuracy stayed at or above the BF16 baseline.

Why it matters: The post tests a skills-only inference engine method on a real model and states the speed and accuracy constraints used, helping readers judge how far such automated optimization can be trusted.

AIAnthropic's ClaudeDevs describes a "You should know" mod that spins off a sideagent to observe Claude's output. The mechanism is detailed in a newer blog post on getting started with Claude Code mods.

AIThe vLLM team integrated Helion, a PyTorch-native kernel DSL, into vLLM's linear backend, using per-shape autotuning to select among Standard GEMM, Split-K, and Swap-AB variants. On NVIDIA Hopper GPUs, the Helion backend outperformed the default CUTLASS and DeepGEMM backends across the evaluated models, with more than 10% throughput gains for some workloads. The work focuses on FP8 and INT8 quantized GEMM.

AIThe Linux Foundation Education has launched a PyTorch Certified Associate (PTCA) Certification Pathway that combines four self-paced learning modules with the PTCA exam. The pathway includes 15–17 hours of self-paced learning and hands-on labs covering tensors, data handling, model development, and performance optimization. The source recommends additional hands-on practice before taking the exam.

AITypeSafe AI shows how to build long-document search by combining its Jev with PageIndex, with no vector database or embeddings required. The project is open source on GitHub under VectifyAI's jev-doc-search repository.

AINVIDIA reports that fine-tuning Nemotron 3.5 ASR reduced word error rate on Najdi and Hijazi Saudi Arabic from 55% to 30%. The post says a new tutorial shows how to adapt the model for other dialects and languages.

AITogether AI's CPO and product team outlined how they use AI to make the whole team more productive, not just individuals. The approach includes a shared context repo readable by any AI harness, cutting half a day of research to about 5 minutes, and evals that test their product the way agents actually use it.

AIAirbnb CTO Ahmad Al-Dahle, who joined from Meta in January, says 60% of the company's code is now AI-authored and pull-request throughput per engineer is up about 1.6x. Roughly half of Airbnb's support tickets are now resolved purely by AI, which the company tested with synthetic data before production. Airbnb's internal context graph Everest helped speed up the grocery delivery and airport pickup services, which took eight to nine months and about six weeks to build, respectively.

AIArchitectural Decision Records (ADRs) give coding agents durable project context, helping them distinguish intentional decisions from implementation details. Agents can over-apply accepted but obsolete ADRs, so the author recommends explicit AGENTS.md instructions treating accepted ADRs as binding, prompting agents to flag conflicts, and keeping each ADR current rather than recording amendment logs.

AIDongxi NLP (@dongxi_nlp) says it followed a suggestion from @karpathy to write the ASD-STE100 controlled English standard into a system prompt. The post is brief and gives no benchmark results, prompt text, or test details.

AIServiceNow CoreAI introduced AutoSynthData, which uses a target model's failures and a stronger teacher's successes to generate and validate new agent training tasks. In EnterpriseOps Gym experiments, the Hybrid domain produced 2,000 samples and raised Gemma-4-26B-A4B-it mean Pass@1 by 7.2 percentage points, while the ITSM domain produced 1,994 samples and raised it from 18.77% to 27.18%.

Why it matters: The post shows how failure analysis, teacher demonstrations, and verifier checks combine into a repeatable pipeline for generating targeted agent training data.

AIEvery's senior editor describes moving from single-thread chatbot prompting to delegating complex projects to teams of coordinating subagents, using skills, orchestrator threads, context packets, MCPs, and computer use. He says a subagent workflow verified employee equity costs across multiple grants, strike prices, and vesting schedules, and returned a draft Slack message for approval. The shift was prompted by a June tweet in which Codex placed a colleague at Level 5 of the "Eight Levels of AI Adoption" framework.

AICreative director Johnson Sheng tested Kling 4.0 for commercial video production, focusing on stability during fast camera moves and dynamic action. He reports stable motion in whip pan and handheld push-in shots, a 30-second single-take fight scene, and consistent props and characters across scene changes. The post also covers performance and emotion control through prompt adjustments and multilingual generation. Kling states Kling 4.0 is in closed beta with an official launch planned for October, supporting up to 4K resolution and 10-bit HDR output.

AIClaude Code can now be modded to change how it behaves, customize its UI, and add custom features. Mods are written in a few lines of TypeScript or generated by Claude, and ship inside plugins installed with /plugin in the CLI or desktop app.

AIAndrej Karpathy says people will spend more time understanding language model outputs and suggests better formats than plain text. He recommends asking for ASD-STE100 controlled-language explanations, diagrams, interactive HTML pages, or custom explainer videos, and he is most bullish on the video format.

AIThe article compares LangChain/LangGraph and CrewAI workflow orchestration with OpenRouter's native model and provider routing. It says OpenRouter's models parameter provides an ordered, error-driven fallback list, while LangGraph and CrewAI handle state, memory, and delegation. Frameworks can also run on OpenRouter as the model layer underneath.

AITool definitions and tool-call responses differ between OpenAI, Anthropic, and Google, so a tool that works on one model may fail on another. The article compares six agent frameworks, including LangChain, CrewAI, and the OpenAI Agents SDK, by where each performs schema translation. It also describes OpenRouter's API-layer normalization, which accepts an OpenAI-style tools array and returns a standard tool_calls response for tool-capable models.

AIFireworks reports that numerical mismatch between training and rollout engines can destabilize reinforcement learning, with a GLM 5.2 experiment showing collapsing reward without alignment and stable reward with it over 25 steps. The post notes that MoE models add further alignment challenges, as Qwen3.5-MoE differences in expert output combination caused disagreement even when one implementation used higher precision. Fireworks says it co-develops its trainer and rollout engine to keep frontier RL training aligned across numerics, kernels, and MoEs.

AIClaude Code now supports mods that change its behavior, customize the UI, or add features, written in a few lines of TypeScript or generated by Claude. Mods ship inside plugins and are installed with /plugin in the CLI or desktop app, and the author says mods can be shared so others can try them.

AIAnthropic's Thariq describes Claude Code mods, which can read conversation scope such as turn count and token usage. Mods run in process, so they can spawn subagents, parse their results with structured output, and modify the UI, which hooks cannot do. Mods ship inside plugins and are installed with /plugin in the CLI or desktop app.