All AI news

AIInferact says vLLM maintainers used megakernel optimization to reach 700 tokens per second per user on TPUv7 running Kimi K3. SemiAnalysis, which shared the work, reports this is 56% better performance than Nvidia's GB200 NVL72. Inferact links a full technical breakdown of the TPU megakernel work on its blog.

AIGitHub rebuilt the pull request view in the GitHub Copilot app to keep review fast on very large diffs, testing it on an open source pull request with 2,200 files, over a million changed lines, and more than 400 inline review comments. The core difficulty is that review comment heights can only be measured at render time, which breaks the fixed-geometry virtualization used for code-only diffs. GitHub split the document height into a deterministic code domain and a separately measured domain for comment blocks.

AIWrite-up and megakernel: 2/2

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.

AICursor announced improvements to the token efficiency of its agent harness, linking to a blog post with the details. The post itself offers no figures or specifics beyond that headline claim.

AIKarl's AI Watts (@aiwarts) has open-sourced Awesome Seedance on GitHub, a collection of AI video prompts. It includes 463 cases traced to original posts, 264 cross-model retest runs with public verdicts, 25 copy-ready templates, and 60 installable AI-video Skills.

AIThe author compiled 436 AI videos made with Seedance into a prompt collection and a Skill, now available on GitHub as awesome-seedance. The post asks GitHub users to star the repository to help it reach a thousand stars in a day, and notes the X post has about 400,000 reads.

AIAnthropic's Cat Wu points users to ways to try Opus 5.5 and Claude Tag in Slack. The main post carries no further details on access, features, or availability.

AITinker says an open LLM can serve a Jev-like interface that takes discrete options and returns fast probabilities, since next-token prediction is already a probabilistic classifier. A post by @ekzhang1 reports that a $5, 10-minute supervised fine-tuning run on Tinker improved Qwen3.6-35B-A3B's handling of Jev-style prompts, with +8% on GPQA Diamond and +12% on MMLU-Pro.

AIThe Together AI blog shows how to fine-tune a Qwen3.5 4B base model into a classification model using about 38,000 examples sampled from six Hugging Face datasets, at a training cost of roughly $17.0. The tutorial covers cloning the tev1 repository, normalizing data with provided scripts, launching a Together AI fine-tuning job that takes about 25 minutes, and deploying the result to a dedicated H100 endpoint.

AISwitching effort levels mid-session on Opus 5.5 does not break the prompt cache, according to Lydia Hallie of Anthropic. This holds when running Claude Code v2.1.280 or later.

AIA prompt shared by Alex Albert asks Claude to recreate San Francisco's Market Street as it stood on April 17, 1906, before the earthquake, using Blender. It requires building a source file from Sanborn fire insurance maps, the Miles Brothers film, period photos, and USGS topography, with reusable Blender Python generators for facades, street lamps, and vehicles, ending in a 10-second video up the street.

AIOpenBMB thanked a developer for testing MiniCPM5-2B as a worker in a multi-agent workflow handling invoice matching, short payments, duplicate references, and disputes through tool calls. The background post says GPT-6 Astra coordinated the MiniCPM5-2B workers, verifying 32 synthetic invoices in 67.8 seconds with 232 executed tool calls. The demo does not move money.

AISuno Studio includes a compressor plugin that reduces dynamic range by turning down peaks and applying makeup gain to raise quieter parts. Its controls include threshold, ratio, attack, and release, and it is especially useful for shaping uploaded or recorded audio. Studio is browser-based and comes free with every Suno Premier subscription.

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.

AISpaceXAI says its combined support team handled a 175% rise in tickets without hiring, crediting Grok Bot, which it says would otherwise have required about 200 additional staff. The company reports resolving tickets for $0.20 to $0.30 each, versus the $1 to $4 per resolution it attributes to traditional AI support tools. Grok Bot is also reported to resolve 99% of refund requests without human intervention.

AIMike Knoop notes that logprobs used to be widely exposed by LLM inference APIs and sees the market maturing so that parts of the LLM stack can be packaged in new, useful ways. He links this to Bryan Helmig's post on prompting with max_tokens: 1 plus logprobs for fast, parallel judgments, which Helmig says has a lot more depth than he expected.

AIThe article explains how Mixture of Experts models change inference by making prefill, midfill, decode attention, and decode experts distinct workloads. It describes how KV cache state, expert routing, and parallelism choices shape compute, memory, and network demands across an inference cluster.

AIMax Woolf reports that prompting agents to iteratively optimize code can make it faster than current state-of-the-art libraries. He says the blog post includes his prompts and benchmark results.

AILMSYS Org shared a blog post about NVFP4 KV cache, a topic linked from its 2026-09-16 article. The post itself contains only a link, so no further technical details, figures, or results can be confirmed from this source.

AIDatabricks CEO Ali Ghodsi praises how Superhuman scales its AI infrastructure, linking to a Superhuman blog post on scaling GEC inference. The post itself is only a link with no details provided here, so the specific techniques and results cannot be confirmed from this source.

AIvLLM achieved 5000 total token throughput per GPU in high-throughput PD serving of Qwen3.8-2.4T on a GB300 NVL72 cluster under an 8K/1K workload. The low-latency scenario reached 180 generated tokens per user, with both results shown on the Pareto frontier. The post also provides srt-slurm recipes and explains the tuning process used to create them.

AISGLang and RLinf have published a blog post about integrating Cosmos3 for reinforcement learning. The main post provides only a link to the full blog, so no further technical details can be confirmed from this source.

AIA commenter praises a practical setup where a small local model runs a log-debugging pipeline entirely on CPU through OpenClaw and Ollama. They argue that workflow design can make a small model far more useful than its raw size suggests.

AIOpenBMB's 2B-parameter MiniCPM5 model runs as a local news desk on an older i5-9400F PC with 16GB RAM and no cloud API. The developer built a system that collects official sources hourly and sends a 24-hour Telegram recap with a lead story and links.

AIJasper Lu's new blog post walks through training a search agent with GRPO, showing how small reward function changes teach a model to avoid sloppy tool calls, prune unnecessary documents, and balance persistence against token efficiency. The post makes every rollout browsable and releases the code as open source, with the full process from learning rate sweeps to reward shaping documented.

AILMSYS Org published a blog post introducing an SGLang SSD expert pack, with the full details available on its website. The post itself gives no further technical specifics, so the summary is limited to the announcement.

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.

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.