— Inference chips and open-source models debuted on the same day, showing that competition in AI infrastructure has shifted from parameters to silicon.
OpenAI released the first benchmarks for its in-house inference chip Jalapeño, claiming higher throughput and energy efficiency than Nvidia Blackwell; Qwen 3.8-Flash-Next is set to launch today, with the community already preparing local deployment plans. Anthropic unified shared memory between Claude chat and Cowork, and X sent a cease-and-desist letter to open-source project Nitter. On the open-source front, IBM released the Granite 4.2 reasoning model family, while agent infrastructure papers such as Prime Agent and SchemaRouter emerged in clusters.
OpenAI unveiled the first benchmark results for its Jalapeño inference chip at Hot Chips. In SemiAnalysis InferenceX tests, it outperformed currently available Nvidia Blackwell systems in per-user token count and per-kilowatt throughput. The chip was co-designed with Broadcom, going from team formation to tape-out in about 16 months. Why it matters: If mass production delivers on these results, inference costs and latency could drop significantly, directly impacting LLM service pricing and large-scale deployment; a potential paradigm shift for GPU-dependent backend architectures.
The community generally acknowledges that hardware advances will lower inference costs, but some believe it needs validation under production workloads and carries hype risk.
Qwen 3.8-Flash-Next (125B parameters, a6B active) is scheduled for open release on ModelScope at 2026-08-26 15:00 UTC, touting lightning-fast inference. Unsloth has confirmed Day 0 quantization and fine-tuning support, and the community has begun preparing disk space and local deployment plans. Why it matters: At 125B scale with low active parameters, if it can run locally, it would significantly lower the barrier to high-quality reasoning models for individual developers and edge deployment.
The comment section generally looks forward to the release, focusing on local performance and hardware requirements, though some find the model naming confusing and are dissatisfied with OpenRouter integration.
Anthropic merged the memory systems of Claude chat and Cowork, allowing projects, preferences, and context learned in one scenario to be reused across scenarios, with users able to read, edit, or delete memory content on any topic. Previously, memory was often fragmented due to architectural limitations. Why it matters: The pain point of repeated briefings in agent workflows is directly eliminated, and improved cross-session state maintenance will enhance the usability of long-cycle automated tasks.
X sent a cease-and-desist to the Nitter project, demanding the shutdown of instances and code repositories, citing alleged scraping. The Nitter repository was archived as read-only on August 25, and all instances are experiencing rate-limited errors. Nitter previously provided an ad-free, tracking-free privacy-friendly frontend by scraping public tweets. Why it matters: This is another legal blow to third-party frontends following the 2024 API restrictions, further narrowing access to the open web and signaling clear risk to developers relying on public data.
Most lament Nitter's shutdown, believing X's forced login harms the open web, though some see it as X's legal right.
IBM released Granite 4.2, its first dense decoder-only reasoning LLM family, with 3B, 8B, and 30B sizes, all pretrained from scratch on ~15T tokens with context windows extended to 512K. The 8B and 30B models underwent agentic RL training in real sandbox environments, supporting thinking/non-thinking toggles and native tool calling, all under Apache 2.0 license. Why it matters: Open-source reasoning models at 30B scale offering 512K context and tool calling provide a commercially viable alternative for local deployment and further development.
ウェブは全体像、購読者にはあなた専用を——興味に合わせた AI 精選、プライベート RSS の統合、コミュニティの見解付きで毎朝配信。ずっと無料。
45 号配信 · 毎日150件超から読む価値ある30件に厳選
Apodex 1.1 improves verifiable sustained work on complex tasks through environment scaling and agentic coordination extensions.
🤖Apodex 1.1 improves sustained, verifiable progress on complex real-world tasks by scaling executable environments and training agents to coordinate long-horizon work with state maintenance and recovery.
Prime Agent is an open-source long-horizon agent harness that extends LLM long-horizon capabilities with recursive subagents and persistent REPL, with 18k+ stars on GitHub.
🤖Prime Agent is an open-source harness that uses recursive subagents, persistent computation, and agent-to-agent coordination to extend language models' long-horizon capabilities across coding and reasoning tasks.
KVBoost enables chunk-level KV cache reuse without requiring shared prefixes, reducing prefill latency for shared content at arbitrary positions.
Quantization-Aware Healing lets 4-bit compressed models outperform their full-precision originals on 7/9 benchmarks, overturning the conventional wisdom that quantization means quality loss.
Bookshelf is a self-hosted ebook library running on object storage, supporting browser reading and Kobo OPDS catalogs.
The comment section generally appreciates the diversity of self-hosted ebook libraries and recommends Calibre alternatives; however, some believe Bookshelf relies too heavily on object storage and npm dependencies, needing stronger OPDS and end-to-end encryption.
LatticeDB positions itself as the SQLite of graph databases, a single-file embedded database with native vector and BM25 full-text indexing.
The comment section generally appreciates LatticeDB's local file and SQLite-style design, focusing on data scale, concurrent writes, and comparisons with existing tools, though some believe its practical application scenarios need further exploration.
EchoWM is an open-source omnimodal world model that jointly generates 720p video, ambient sound, music, and speech while responding to continuous 6-DoF navigation.
🤖EchoWM is an omnimodal world model that generates synchronized high-resolution video, sound, music, and speech while following continuous 6-DoF navigation trajectories across first- and third-person views.
SenseTime open-sources SenseNova U1.5 Lite, an 8B image generation model supporting 4K direct output, long instruction following, and native image editing.
ChatGPT free accounts now get access to the task scheduling tool, while paid accounts can configure Gmail, Slack, and GitHub event triggers for prompts.
Paradigm, together with UBTech and over ten embodied intelligence vendors, launches the PhanthyMotus co-building plan, with the community contributing over 600k lines of code.
基于官方 DeepSeek Harness 打造的 Electron 桌面端，深度适配 macOS 和 Windows，提供最佳的，开箱即用的体验。
⚡️A native, local-first alternative to Logitech Options+, written in Rust 🦀 — remap buttons, DPI, and SmartShift over HID++. No account, no telemetry.
29 editorial diagram types for Claude Code. Self-contained HTML + SVG. No shadows, no Mermaid-slop.
Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop and mobile.
Strip multi-vendor AI provenance marks: Unicode text hygiene, statistical rewrite hooks, and C2PA/metadata from PNG/JPEG/SVG/PDF/DOCX/HTML/MD
Qwen3.8-Flash-Next (~125B-A6B + 51B n-gram) memory estimate: Ideal 4-bit quant ≈ 82 GB (58 GB main weights + 24 GB n-gram tables) Real-world quants likely land in the 80–90 GB range. The big n-gram table is sparsely accessed → excellent candidate for system RAM offload. This architecture could be surprisingly local-friendly once the weights drop.
It worked pretty well, digging through and analyzing some local files. Claude code regularly stops at some point and fails to continue while DSH worked for 2 h, recognized that it could benefit from reading more context and ... bummer: It left the project folder (although DSH was set up correctly) and started to walk through my other files which I did never allow! I know, it is only a preview. But be warned. Do not expect it to follow simple rules.
We have been working on some performance optimisations for Qwen3.8 and other models. The main new feature that we introduced is adaptive speculation for Llama.cpp What is it? MTP and DFlash work well to speed up inference work, especially for dense models. However, different content types need different settings. Llama.cpp only supports a single value. This fork introduces adaptive speculation. You set the minimum and maximum and the engine will adjust the number of tokens that are suggested aut
OpenAI CFO Sarah Friar explains how advances across chips, compute, models, and products compound to deliver more useful intelligence at greater scale and lower cost.
World Action Models (WAMs) improve planning by incorporating future world evolution into action generation, yet existing methods allocate a fixed imagination budget to every scene. We propose RISE (Refining Imagination through SElective Rollout), a system-level adaptive imagination framework that makes sequential Roll/Stop decisions according to the expected planning benefit of continued rollout. At each step, a Latent Evaluator estimates the risk revealed by the current prefix and how much plan
Policy optimization (PO) for Large Language Models faces a stability--exploration trade-off, currently mediated by an action-side Policy-KL regularizer. This puts practitioners in a double bind: keeping Policy-KL constrains response behavior and consumes the action-side exploration budget, while dropping it leaves the optimization without an explicit drift control. We argue for an alternative that breaks the dilemma by moving regularization to the input side. As training progresses, the distribu
Found that you can actually run a 35B Qwen model on a Pi with very impressive intelligence and stability. Built connectors for car ODB to read all about car internals, and manufacturer's cloud service for stuff like changing AC or opening/ locking doors. Gave it info such as the full car manual. And then hooked it up with my other agents in our discussion room! So now it can answer car questions such as "when should I add oil and what kind of oil?" and help you fully offline, and when online tal
Semantic vision encoders have become a central visual interface for multimodal understanding and semantic conditioning in image generation. However, their final tokens discard fine-grained visual details, leading to poor pixel reconstruction and limiting their use in reconstruction-sensitive tasks such as image generation and editing. In this work, we ask whether understanding, generation, and editing can be modeled in a single visual representation space built from a pretrained semantic ViT. We
Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather th
Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox,
Many people find AI-based chatbots helpful in keeping up with news, but a study by Pattie Maes and her colleagues at the MIT Media Lab points to a big problem with this strategy. Participants who evaluated paired news headlines and images over the course of four weeks were initially 21% percent more accurate at telling…
Elon Musk wants to harness solar energy for AI.
Grand Theft Auto fans, eager for news about one of the most anticipated video games of all time, appear especially vulnerable to this new cyberattack.
arXiv:2608.21417v1 Announce Type: new Abstract: Flexible manufacturing requires industrial robots to be reprogrammed rapidly as product variants change. This paper presents a language-model-based workflow that generates, validates, and iteratively corrects ABB RAPID robot programs from natural language task descriptions. A dual-stream retrieval-augmented generation (RAG) pipeline grounds code generation in verified technical documentation and production templates, reducing domain-specific errors
arXiv:2608.21382v1 Announce Type: new Abstract: Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whether a language model's answer is read from generated text or from per-option likelihoods. Work on this harness sensitivity reports it as aggregate score variance, leaving unexamined which items the variance falls on and whether they are the items that separate one model from the next. We treat the e
arXiv:2608.21366v1 Announce Type: new Abstract: Driven by massive amounts of web-scale data, generative AI (GenAI) has achieved remarkable progress, enabling various applications in diverse sectors. The advances of GenAI have actuated practitioners to use AI-synthesized data for training next-generation AI models. Undeniably, using synthetic data has alleviated the increasing stringent demand for data supply. Unfortunately, it also introduces a new critical issue: in a self-consuming cycle betwe
Solid-state transformers could also benefit EV charging and, someday, households.
下一代Agent天然会走向云端和更大规模的计算资源
An industry legend starts covering the inevitable!
OpenAI banned Russia-origin accounts using AI to promote a fake Israel-based think tank and a “sovereignty” index praising Russia and criticizing the West.
An interactive world model must follow the user's actions, remember the places it has shown, and stream in real time. The tension is structural: control wants a short horizon, memory wants an unbounded one. ReWorld separates the two during training and bounds them at inference. Mixed per-head attention windows confine most heads to the recent past while a small set of global heads attends over the entire history, and random head routing keeps either capability from binding to particular heads; r
Open-ended real-world interaction admits multiple valid behaviors: an agent may answer directly, ask for clarification, provide progress updates, or confirm before acting. This flexibility breaks a core assumption behind group-based RL: rollouts compared within a group are no longer guaranteed to be behaviorally comparable. As a result, reward-model preferences over interaction style can distort relative advantages and steer optimization toward reward-preferred behaviors rather than context-appr
It's a macOS menu bar app that reads the text of your focused window every few seconds through the Accessibility API. No screenshots, no video, or OCR. It writes plain markdown, one file per day, into a folder you choose. Point Claude Code (or anything with file access) at the folder and ask "what did I work on Tuesday?" or have it build memory about your projects. The folder contains an AGENTS.md explaining the format to the model. Thought it might be useful for someone so chucked it up on GitH
Existing image editing frameworks predominantly follow the training paradigm of text-to-image diffusion models. However, extending this paradigm to image editing highlights two inherent discrepancies, specifically, the insufficient attention to edit concept granularity and the training inefficiency caused by sparse supervision signals. To address these issues, we establish a comprehensive hierarchical taxonomy featuring over 1,000 fine-grained edit concepts and build ConceptEdit-12M, a massive d
Folks have been daisy-chaining Macs for AI—this refresh keeps that in mind.
With large pretrained models, existing methods have effectively improved instruction-based video editing. However, most of them rely on an in-place editing assumption. They align the edited video with the given source clip frame by frame over a fixed time span. This pattern fails for open-ended streams, e.g., restyling a live game or applying a camera move to an ongoing shot. In such cases, edits must extend to future frames as they arrive, rather than be applied to a static input clip. In this
Privilege escalation attack grants "full control" and freedom from Meta's servers/apps.
arXiv:2608.21363v1 Announce Type: new Abstract: A protocol is presented for recording the governance decisions of automated AI runtimes. When a runtime releases, blocks, defers, redacts, or escalates an individual output, AIREP records that decision as a single signed object that any party can check offline, independent of the runtime that produced it. A record carries the decision as one of a closed set of verbs under a stated policy basis, references its input, output, and evidence by hash rat
arXiv:2608.21374v1 Announce Type: new Abstract: Literature reviews are essential to scientific progress, but rigorously evaluating automatically generated reviews remains difficult because many aspects of research utility depend on expert judgment rather than reference-overlap metrics. We introduce LitReview Arena, a battle-style evaluation platform with a structured protocol tailored to literature review quality: domain experts with AI paper-writing experience compare anonymized drafts, are mat
arXiv:2608.21444v1 Announce Type: new Abstract: Multi-drone systems are increasingly positioned for safety-critical missions such as search and rescue (SAR) and critical infrastructure monitoring. Yet, real-world adoption remains constrained not only by autonomy performance, but by the difficulty of integrating agentic behavior into professional work: operators must understand, trust, and govern automation under uncertainty, time pressure, and accountability. This position paper synthesizes the
User representation learning in real-world industrial scenarios is commonly scaled by increasing user amount, behavioral sequence length and model size. However, existing methods face two challenges: (i) Bottleneck for raw data scaling at billion-scale capacity, as performance exhibit diminishing performance gains with larger-scale raw text user behavioral input, which can be mitigated by tokenization. (ii) Lack of quantitative analysis of how tokenization configurations should scale with data s
arXiv:2608.21418v1 Announce Type: new Abstract: Manufacturing knowledge graphs that integrate data from heterogeneous industrial systems face a trust deficit: consumers cannot determine whether queried data is valid, whether it was valid when a decision was made, where it originated, or how it was acted upon. We argue that four trust capabilities -- SHACL validation, PROV-O provenance, domain-aware bi-temporal versioning, and graph-native decision objects -- compose through shared correlation id
Use the Admin plugin for ChatGPT Work and Codex to analyze workspace usage, manage members and permissions, adjust limits, and act on admin requests.
Recent large language models (LLMs) can operate as coding agents that build complete games from natural language requests. Game development is especially demanding because program logic, visual and audio content, interfaces, interaction and playability must function together in one executable artifact. Measuring this capability therefore requires evaluation of both game product and the development process. Existing benchmarks often assess the game development capabilities of LLMs by evaluating t
At $10k, you could get - 6.2B tokens with Qwen 3.8 Max (Qwen Pro plan) - 5.7B tokens with DeepSeek V4 Pro OpenRouter - 100B tokens with DeepSeek V4 Flash OpenRouter As a firm believer of local inference, unless you need it for data sovereignty, it's much more cost effect to wait for smaller models to keep getting better. In the meantime, find a reasonably priced 24GB - 32GB card for Qwen 3.8 27B, and offload hard tasks to OpenRouter. Qwhen 3.8 35B A3B?