Top picks

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.

AILangChain built Restock, a sample office-supply agent that runs in Slack on Managed Deep Agents and pays through Stripe's Link wallet. The agent searches products, builds a cart, and pays over the Machine Payments Protocol, with the user approving the purchase in Slack and the payment in Link. The post uses a pens order at $22.18 to show the flow from request to confirmed order.

Why it matters: The post walks through how an agent handles search, budget limits, Slack review, and Link approval, showing where each control sits outside the model.

AIA Hugging Face blog author used the ML-Intern agent in HuggingChat to build six small models by writing detailed prompts that specify datasets, base models, baselines, smoke tests, and spending limits. The projects include a citrus disease vision-language model, a Huggy character LoRA, a camera-angle LoRA, a doodle-to-object LoRA, a 0.8B prompt rewriter, and a 4-step distilled Agate model, with total compute cost of about USD 103. Each project's prompts and public models are linked from the post.

Why it matters: The author shows how prompt structure, baselines, smoke tests, and budget caps shape an agent-driven training workflow, with per-project costs given.

AIAnthropic added build-eval and hillclimb commands to its claude-api skill for designing evaluations and iteratively improving applications against them. The article covers eval design principles, including production-representative tasks, headroom and low variance, and guards against overfitting through train/test splits. Two examples report results: a customer support benchmark where cost fell to under half while accuracy rose, and a claude-api skill eval that rose from 66% to 88%.

Why it matters: The article gives a concrete workflow for designing evals and hillclimbing without overfitting, with two worked cost and performance examples that show the tradeoffs.

AIInferact and the vLLM community reported a 1.9× low-concurrency speedup and about 5.3× throughput under a 150 TPS constraint for DeepSeek-V4.1-Flash over three weeks. Gains came from SWA bounded replay with CUDA graphs, which cut TTFT by about 30%, and from integrated DeepSeek kernels such as MegaAttention, Mega-mHC, Mega-Gate, and DeepSelect. The post measures these results on the SemiAnalysis AgentX benchmark.

Why it matters: The post breaks down how SWA bounded replay and fused kernels cut prefill and decode costs, a reusable engineering pattern for long-context agentic serving.

AIComcast and Booz Allen used Claude Mythos Preview to find vulnerabilities that arise from interactions across code, configuration, and deployment rather than single-file bugs. Comcast identified a critical authentication flaw across 258 systems and about 170 million lines of code before any exploitation was observed. Booz Allen reported that one analyst reviewed eight production systems across 138 repositories in twelve days, a review its team estimated would have taken several months without the model.

Why it matters: The case studies show how security teams validate and remediate model-found exploit chains, a workflow relevant to anyone managing large codebases.

AIGoogle released EmbeddingGemma 2, an open embedding model under the Apache 2.0 license that maps text, code, images, video, and audio into a shared 768-dimensional space. Developers can load a 270M-parameter text and code setup, or add vision and audio encoders up to a 740M-parameter full multimodal model. Matryoshka truncation to 256 or 128 dimensions reduces vector storage, with the guide noting quality losses on image, video, and speech retrieval at lower dimensions.

Why it matters: The guide gives concrete encoder sizes and dimension-storage tradeoffs, showing how to choose a configuration for text, code, image, video, and audio retrieval.

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.

AIPhysicist Matthew Schwartz describes building BootLoops, an open-source harness for exact quantitative calculations, after choosing problems suited to Claude's strengths. He reports that Claude solved long-standing integrals and found connections across ecology, population genetics, economics, and linguistics, with domain experts steering results toward questions those fields care about. The post states that the approach required constant human oversight, since Claude often overstated results and misjudged time.

Why it matters: The guest post explains why scientists often find current AI tools frustrating and offers a method for finding problems where AI and researchers match, backed by concrete projects.

AIXiaomi MiMo reports that tool-call repetition in MiMo-V2.6 reached over 0.05% of responses across agent harnesses, causing stalled agents and wasted context. The team traced the cause to an RL flooding penalty set at 32 calls per turn, which missed smaller excess behavior, and replaced the approach with a specialized teacher distilled via MOPD. Repetition rates for both Pro and Flash dropped substantially, at roughly $90,000 versus an estimated $2.31 million for the alternative fix.

Why it matters: The post traces an agent failure to a reward blind spot and compares the costs of two fixes, offering a transferable debugging method for RL-trained tool-calling models.

AILovable describes how its Chats feature lets a workspace-level chat agent hand work to project builder agents and receive progress back. The design records each agent's history as an append-only, forkable trajectory, and passes messages through durable inboxes that activations wake. Agents can suspend at iteration boundaries and resume on freshly deployed nodes without killing long-running runs.

Why it matters: The post details how trajectories, inboxes, and activations let agents share work and resume after deploys, useful for designing comparable agent systems.

AICursor's Eric Zakariasson shared a prompt for improving an LLM agent harness to lower token cost per completed task without losing quality. The prompt covers the system prompt, tool definitions, cache layout, tool results, compaction, and subagents, and reports that one team's round of these changes cut overall token cost about 7%.

Why it matters: The prompt gives a concrete checklist for cutting agent token cost per completed task, with tested figures on cache layout, tool offloading, and compaction.

AIMicrosoft Research reports that running physical AI inference on onboard GPUs can limit robot performance and battery life, while offloading inference to edge or cloud GPUs improved results in mobile manipulation tests. In its evaluation, smaller onboard GPUs slowed mapping and planning by up to 383% compared with an A100, and large onboard GPUs such as Jetson Thor drained robot batteries by up to 160%.

Why it matters: The study measures how offloading robot inference to edge or cloud GPUs changes task success, battery life, and model size, offering evidence for infrastructure design.

AITencent Hunyuan says it shrank the 770B-parameter Hy4 preview from roughly 1.5TB to 214 GiB while keeping the parameter count unchanged. The quoted Zhihu post by a Tencent Hunyuan quantization team member describes the method: a 1.25-bit sparse ternary encoding, mixed precision across expert layers, and STQ1_0 CUDA kernels in llama.cpp. The author reports nearly unchanged MRCR retrieval and a small decline in math.

Why it matters: The quoted Zhihu post explains how Hy4 preview's weights were quantized and kept usable at inference, a concrete engineering case for compressing large MoE models.

AIPart 2 of the zero-trust agents series moves security checks from agent code to the Gemini Enterprise Agent Platform runtime. Model Armor screens prompts and responses, Semantic Governance Policies judge proposed tool calls against intent and business rules, and Agent Anomaly Detection flags multi-turn drainage that single-turn checks miss. The same Customer Support and Returns Agent from Part 1 is used, with the companion demo open-sourced on GitHub.

Why it matters: The post walks through a concrete refund agent under four attacks, showing how screening, intent judgment, and anomaly detection each catch what the others miss.

AIThe vLLM team trained a DSpark speculative decoding draft model for Kimi K3, a 2.8T-parameter model, using the Speculators library on GB300 NVL72 hardware. They added a MooncakeHiddenStatesConnector to stream hidden states from disaggregated vLLM inference nodes to training nodes across multiple machines. The released speculator raises single-stream interactivity from about 110 to about 435 tokens per second per user on math reasoning, with up to about 3.5x higher output throughput under concurrent load.

Why it matters: The post shows how hidden-state extraction and Mooncake transfers let a 2.8T-parameter model's speculator be trained across multiple nodes, a reusable pattern for similar setups.

AIAugment Code reports that size-adjusted output per active developer rose from 12.3 to 55.7 between November 2025 and July 2026, while median time to merge fell from 11.2 to 3.1 hours. The post says the company added specialized agents wherever work was piling up, across planning, review, verification, feedback, and incident response, and kept engineers responsible for product decisions, architecture, and production risk.

Why it matters: The post pairs internal productivity and quality metrics with the order in which agents were added, showing how review and verification bottlenecks shaped a software delivery pipeline.

AIPromptArmor demonstrated that OpenAI's Approve-for-me agent approved a malicious NPM install with elevated privileges after a hidden prompt injection in an external GitHub issue influenced the main Codex agent. The malicious package's post-install script then ran unsandboxed with the user's full privileges. The report also gives steps for organizations to disable agentic auto-review in Claude Code and Codex.

Why it matters: The report shows a prompt-injected GitHub issue leading an approval agent to permit a malicious NPM install, a concrete test of agent-in-the-loop guardrails.

AIXiaomi describes an end-to-end inference optimization for the MiMo-V2.5 series, centered on Hybrid SWA, which it says cuts KVCache storage to roughly 1/7 of Full Attention. The post covers a dual KVCache pool design, SWA-aware prefix cache matching, the GCache distributed cache, and scheduling changes, and reports cache hit rates averaging 93% in server-side observations. It also covers prefill and decode optimizations, multimodal encoder improvements, and open-source contributions to SGLang.

Why it matters: The post explains how Hybrid SWA's theoretical KVCache savings were realized in production through dual pools, SWA-aware prefix caching, and tiered storage, giving concrete engineering patterns for long-context inference.