#Tutorial/How-to

AIAi2 says its old scheduler made every scheduled workload eventually run at HIGH priority. Researchers kept idle jobs running to reserve GPUs for experiments, because the incentives rewarded holding capacity even with no active work.

AIAi2 has published a blog post on the design decisions behind its scheduler, along with rollout lessons and next priorities. The post links to The text provided does not include further details of the scheduler's features or results.

AITessl Blog describes an agentic code review workflow built for teams whose coding agents produce pull requests faster than humans can review them. The workflow runs review against a written standard in the repository, applies four parallel perspectives covering correctness, security and privacy, scale and resilience, and maintainability, then records each finding, verdict, and response. Tessl Code Review, which the post says is free to start, runs these perspectives as skills, and the team's memory of past decisions is fed back into the standard.

AIAi2's AI Infrastructure team replaced its priority-based scheduler for GPU clusters with GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The team says the change moved debates over how much GPU time each research project deserves from case-by-case operational decisions into a transparent budgeting process. The clusters range from 88 to 1024 GPUs across NVIDIA H100, B200, and B300 hardware, and serve about 150 internal researchers.

AIPostman describes the architecture behind Agent Mode, its AI agent for API testing, documentation, discovery, and implementation. The post covers limiting tools per task, using schema-based queries, building purpose-shaped context handlers, and running on Amazon Bedrock with cross-Region inference and prompt caching. Postman reports that tool-selection errors rose once the visible toolset exceeded about 40 tools.

Why it matters: The post shows concrete patterns for tool scoping, context handling, and Bedrock routing and caching, which apply to any team moving an agent past a prototype.

AIAi2's AI Infrastructure team replaced its priority-based GPU cluster scheduler with a system using GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The team says the change turns decisions about how much GPU time each research project receives into a transparent administrative budgeting process. Its clusters, which range from 88 to 1024 GPUs including H100, B200, and B300 units, serve about 150 researchers facing demand two to three times available capacity.

AIBaseten explains how to decide which layers of a model can run in 4-bit NVFP4 without losing needed information. The post compares architecture-based heuristics, isolated-layer sensitivity scoring, and SaturationQuant, which accounts for other quantized layers. It also covers calibration with representative data and block-level scales of 16 values.

Why it matters: The post explains how to choose which layers run at NVFP4 precision using heuristics, sensitivity scoring, and saturation-aware scoring, with clear calibration steps.

AIGuizang says he turned his workflow into a Grok bot that users can install with one click. The bot runs on Grok's cloud virtual machine to collect content, write code, and render a daily morning AI news video without using a local computer.

AIThe Claude Blog published a guide to building scheduled agent automations with Claude Managed Agents (beta) that reads custom sources such as Slack and GitHub and posts a daily brief. The guide covers scoped vault credentials, per-source bookmarks so no window is lost or repeated, and confirming each Slack post before updating records. It also covers read-only access, a per-run spending cap, and a reference implementation with a Claude Code setup command.

AISimon Willison says he built a Newsletters index for his blog almost entirely by voice, using the ChatGPT desktop app's Codex voice mode while cooking dinner. The feature imports weekly Substack posts via RSS and undocumented API, monthly newsletters from a GitHub archive repository, and a private sponsors-only newsletter. He says he switched back to typing for review and fixes before deploying the pull request.

AIClément Delangue of Hugging Face says he loves the building-in-public approach, responding to Matth Lapeyre's post about Microduck's battery testing. Lapeyre reports the new custom board ran over 3 hours on a 2600 mAh battery, with average current dropping from about 1.13 A to 0.82 A and peak CPU temperature falling from 114°C to 60°C without throttling.

AISimon Willison says he built a new feature for his blog entirely by voice using Codex Desktop while cooking dinner. The post links to a write-up on his site about the voice-driven workflow.

AIMerve Noyan, of Hugging Face, says a new guide on speculative decoding is in progress, followed by one on quantization. She points readers to the Llama App docs page comparing prefill and decode.

AIGoogle's Playground experiment lets users describe a game in ordinary language and have generative AI build a playable browser-based result that can be revised through further prompts. TechRadar's reviewer built a dragon platformer, Mystic Dragon Glide, from a couple of sentences and a satirical puzzle RPG, Red Tape Hero, from a longer prompt. Playground produced working controls, objectives and music, but the reviewer found the results impressive as prototypes rather than games they would want to play for dozens of hours.

AIMarkTechPost publishes a hands-on tutorial implementing RRSI (Regularized Recursive Self-Improvement), a method that lets an LLM agent revise its own harness around a frozen model. The full loop drafts edits with Claude Opus on Vertex AI and scores them in Docker benchmarks, but the edit-selection rules are plain Python that the tutorial runs in a simulated environment with a calibrated noise band.

AISimon Willison released ttok 1.0, changing its default tokenizer from GPT-4 to a newer GPT-5/GPT-6 tokenizer after finding the old default was outdated. OpenAI has not confirmed that GPT-6 shares the GPT-5 tokenizer, but William Liu's experiment found all seven GPT models tested reported 44,794 tokens on a 31-fixture corpus.

AIHong Kong's Centre for Artificial Intelligence and Robotics (CAIR), Chinese Academy of Sciences, unveiled CARES 4.0, a multimodal clinical AI agent that carries out tasks rather than only answering questions about images or video. Built on the Harness agent framework and CAIR's own CT, MRI, ultrasound, endoscopy and EEG foundation models, it has been validated at several top-tier hospitals. CAIR says the system gives suggestions with reasoning paths and sources, and that the doctor remains the final decision-maker.

AIThe University of Michigan's EECS 498 course Applied Agentic Software Engineering teaches a coding agent across three phases, from applying and analyzing agents to building one. Its Elephant-Goldfish Model packages a design-first workflow into five Skills, with human handoffs between each step, and the course materials are public on GitHub.