Summary Qwen's open-source benchmark runs deterministic 365-day agent-store simulations with negotiation, fraud, and shocks.
The details
The benchmark evaluates 18 frontier models across seven dimensions, including year-end assets, fraud avoidance, operational efficiency, and long-horizon learning.
Agents start with ¥100,000, operate multiple stores for 365 days, and source inventory, negotiate, price, fulfill orders, process returns, and manage cash flow.
The environment has 6,886 products, 576 suppliers, and 152 fraudsters.
Demand, returns, and supplier negotiations are deterministic for reproducibility.
GPT-5.6 Sol grew the opening stake to 1,431,425 and led earnings.
Qwen3.8-Max-Preview led open-weight models and horizon learning.
Qwen3.8-Max-Preview progressively reduced negotiated prices over repeated orders.
Research, Safety & Infrastructure
n8n: RAG for Changing Knowledge
Summary n8n's production guide distinguishes runtime retrieval from weight-level behavior adaptation.
The details
n8n's production guide selects runtime retrieval or weight-level adaptation by the underlying failure mode.
RAG retrieves external documents, vector-store records, or knowledge-graph data at runtime.
RAG injects relevant context into prompts without changing model weights.
Fine-tuning trains an existing model on additional examples to improve learned behavior.
Fine-tuning can improve formatting, style, instruction following, or specialized-task performance.
n8n recommends RAG for current, proprietary, large, traceable knowledge bases.
n8n recommends fine-tuning for inconsistent outputs, specific styles, and specialized tasks.
Strong RAG can require hybrid sparse and dense retrieval plus reranking.
Quality fine-tuning datasets are difficult to create and can degrade results.
Research, Safety & Infrastructure
Hugging Face Open-Sources Watercolor RL Pipeline
Summary A reproducible GRPO recipe trains Qwen3.5 to write JavaScript watercolor paintings using aesthetic preference rewards.
The details
The training target is Qwen/Qwen3.5-35B-A3B with LoRA on all linear layers, bfloat16, gradient checkpointing, and an 8,192-token completion limit.
Each rollout produces roughly 150 lines of JavaScript using only 10 drawing methods from a natural-media extension for p5.js.
Reward combines compilation and anti-cheating checks, longer-code incentive, reference-pool style similarity, and render aesthetic preference.
Qwen3-VL-30B-A3B-Instruct compares each candidate with four randomly selected references in both display orders.
The recipe trains 110 steps across 240 episodes with eight generations each.
Models, APIs & Pricing
OpenAI rolls out GPT-6 Astra for agents
Summary GPT-6 Astra reaches selected organizations first, with API and AWS availability planned in following days.
The details
OpenAI rolls Astra to limited organizations, then ChatGPT Plus, Pro, Business, Enterprise, API, and AWS; API label: gpt-6-astra.
Listed API pricing is $10 per million input tokens and $50 per million output tokens, matching cited Claude Fable 5 and 5.1 rates.
OpenAI reports ARC-AGI 3: 99.9%/$19K with Provider Adapter, versus 62.7%/$26K default harness; adapter preserves opaque reasoning state using compaction.
Results: 100% ExploitBench, 42.4% ExploitGym, 99.2% SRE-Bench within four attempts, and eight-needle 100% at 256K–512K, 96.3% at 512K–1M.
Research, Safety & Infrastructure
PGMQ runs SQS-style queues in Postgres
Summary Postgres queue: visibility-timeout delivery, FIFO groups, routing, message archival; no separate broker.
The details
PGMQ supports PostgreSQL 14 through 18 as an extension or SQL-only installation.
Each queue is a table in the pgmq schema; messages are JSON until explicitly deleted or archived.
Reads set a visibility timeout; unacknowledged messages reappear for another consumer after it expires.
The supplied video describes a 100,000-message, 100-worker test on a 2-CPU container but reports no throughput or latency results.
Models, APIs & Pricing
Sources do not substantiate Google image-editing claim
Summary Video claims Google image editor; links lack product documentation or release details.
The details
The description claims Google Pics can edit individual objects, repair image text, translate designs, and work in Docs and Slides.
Neither supplied primary-source link is a Google page or documents a Google image product, availability, rollout, pricing, or implementation.
One linked page promotes AI Profit Boardroom, a paid community advertised at $59 during a stated flash sale, not the claimed release.
The other linked page promotes the free AI Money Lab community and lists training resources, prompts, tools, and n8n workflows.
Tools & Apps
12 items
Products and workflows worth trying, with limitations and direct links.
Coding Agents & Developer Tools
Hugging Face launches local coding-agent memory
Summary funes indexes agent traces locally and can sync private Hugging Face datasets across machines.
The details
funes is a single binary that incrementally indexes completed agent turns.
Its default inference backend requires no ML runtime.
funes uses a deterministic pipeline and stores traces in a local Lance dataset.
Results return original trace text with agent, timestamp, session, and turn provenance.
A command opens surrounding context.
Bound memory can sync to a private-by-default Hugging Face dataset.
Indexing redacts credentials.
Chunks are rescanned before publishing.
Automation & Agent Systems
llm-gemini 0.34 adds Gemini 3.8 Flash
Summary The LLM plugin exposes Gemini 3.8 Flash with three thinking levels and fixes async model-version recording.
The details
llm-gemini 0.34 adds support for Gemini 3.8 Flash with low, medium, and high thinking levels.
The release fixes async responses failing to record the resolved model version.
Google also released Gemini 3.8 Flash Cyber, but the source says it is available only to trusted defenders.
Author's HTML/JavaScript experiment: Gemini 3.8 Flash completed a task in 13 seconds for 1.8 cents; anecdotal, not a benchmark.
Coding Agents & Developer Tools
shadcn releases cn for faster Tailwind class merging
Summary Zero-dependency cn replaces clsx plus tailwind-merge with compatible class joining and conflict resolution.
The details
cn is framework-agnostic and runs in browsers, Node, Bun, Deno, edge runtimes, and React, Vue, Svelte, Solid, Astro, or server-template projects.
It replaces clsx plus tailwind-merge; shadcn projects can migrate with `npx shadcn@latest migrate cn`.
The repository reports 356,000 differential tests for tailwind-merge output parity and best-of-five isolated benchmarks against clsx plus tailwind-merge.
cn supports Tailwind CSS v4 and v4 prefixes; its CLI requires Node 20+, while Tailwind v3 users are directed to tailwind-merge v2.
Coding Agents & Developer Tools
Together AI Open-Sources Open Customer Insights
Summary A private MCP-accessible workspace makes calls, tickets, Slack, and company context searchable.
The details
The repository provides grounded chat.
It provides semantic and keyword search.
It provides company timelines.
It provides customer-status views.
It provides competitor tracking.
It provides source retrieval across calls, tickets, and Slack.
The stack uses Next.js and Bun.
Convex provides hybrid vector search and scheduled ingestion.
Clerk handles authentication.
Together AI provides chat and embeddings.
AI SDK enables tool calling.
Optional integrations support Gong call transcripts, Pylon tickets, Slack, and company enrichment.
Jobs safely skip when credentials are absent.
The deterministic demo seed creates 12 fictional companies and 36 calls.
It creates 48 tickets and 84 searchable chunks.
It creates 24 Slack mentions and 24 timeline events.
The repository ships no hosted deployment, customer data, or credentials.
Summary Agent Builder compiles plain-English workflows into reusable automations that recover from browser failures.
The details
Agent Builder builds and tests automations from plain-English workflow descriptions, then compiles them into reusable code.
Agents run on schedules or triggers across APIs, applications, and logged-in websites; Airtop says broken runs self-heal.
Airtop claims up to 100x runtime efficiency over uncompiled LLM agents; it supplied no benchmark methodology, pricing, or limits.
The platform brings browser automation to agents built with Claude, n8n, Make, or Codex.
Airtop says it is SOC 2 Type II certified and HIPAA compliant.
Products & Launches
Swiftlet streams Qwen MoE weights from storage on Apple devices
Summary The open Swift and Metal runtime runs 35B and 80B Qwen MoE models with low RAM by loading routed experts on demand.
The details
On an M4 Max with 64 GB memory, Swiftlet runs the 8-bit Qwen3-Next-80B-A3B, whose weights total 78.8 GiB, in 6.9 GiB RAM at about 4.8 tok/s.
The 4-bit 397B model, with 207.6 GiB of weights, runs in 12.6 GiB at about 1.4 tok/s; its approximately 17B active parameters drive slower decoding.
A base M1 with 16 GB decodes 4-bit Qwen3.6-35B-A3B at about 2.45 tok/s; an M4 Max reaches about 19.5 tok/s.
Qwen3.6-35B-A3B also runs on iPhone 17 in about 2.5 GB RAM at about 1 tok/s. The repository says long-prompt prefill is expensive and current optimization work targets kernel dispatch, not I/O.
Automation & Agent Systems
Claude Fable 5.1 broadens copyright refusals
Summary Anthropic’s published consumer-app prompt blocks lyrics and recognizable copyrighted visual output.
The details
Fable 5.1 declines whole or partial lyrics, poems, books, and articles.
It also declines choruses, hooks, and user-supplied lines.
Verified pre-1929 works are allowed.
It prohibits specific artwork, covers, posters, logos, app icon sets, and product designs.
It prohibits known characters and code-generated reproductions through SVG, Canvas, CSS, HTML, plotting, and ASCII.
Filtering persists across rephrased requests.
Published core prompts omit feature- and tool-specific context blocks.
These include end_conversation, memory, past-chat, web-search, artifact, and file-creation instructions.
Summary A community-linked CLI is presented for Claude Code or Codex task dispatch.
The details
The supplied description says Grok Bots run on an online computer.
It says Grok Bots can share files, browser state, and logins.
It says Grok Bots can be scheduled.
It says Grok Bots can be coordinated in group chat.
It claims ScriptedAlchemy/grok-bot-cli can create bots from the command line.
It claims ScriptedAlchemy/grok-bot-cli can assign tasks from the command line.
It claims ScriptedAlchemy/grok-bot-cli can run bot groups from the command line.
Suggested use cases include role-based application testing.
Suggested use cases include authenticated support investigation.
Suggested use cases include competitor monitoring.
Suggested use cases include dashboard monitoring.
AI Labs Pro's primary page describes a private builder community.
Its primary page describes weekly implementation guides and resource kits.
AI Labs Pro's primary page does not provide Grok Bot technical documentation or CLI specifications.
Products & Launches
Any Human Ever Samples Lives From History
Summary Site randomly samples birth years, places, and lives from 100 billion people.
The details
Any Human Ever randomly draws one life step by step: a birth year, place, and life sampled from real data.
The site covers more than 100 billion people estimated to have ever lived.
It visualizes years before present logarithmically; brighter clusters represent larger populations.
It provides full sources for each generated story, but supplied material does not specify datasets or sampling methodology.
Products & Launches
Cerebras documents unpruned public model serving
Summary Cerebras says its public API serves original architectures with storage-only selective weight quantization.
The details
Cerebras says public-endpoint models are original, unpruned versions, not REAP-pruned variants.
Cerebras uses selective weight-only quantization for storage; sensitive layers remain full precision and dequantize on the fly.
Activations, attention, and KV cache remain full precision and unquantized.
Public models are available on free-trial and pay-as-you-go tiers, subject to tier limits.
Products & Launches
ICANN approves .name third-level domain shutdown
Summary Neil Fraser says Verisign's approved plan removes .name's third-level namespace, disrupting domains, email, APIs, IoT services.
The details
Fraser says Verisign proposed eliminating the entire .name third-level hierarchy on 15 April 2026, and ICANN approved it on 28 July 2026.
He says the termination is expected in February, even for domains registered and paid through 2040.
Fraser estimates 22,000 people will lose domains.
According to Fraser, a newly vacant second-level domain's registrant could recreate a former third-level address, receive email, or control linked services.
Automation & Agent Systems
Audacity 4.0 rebuilds editing UI on Qt
Summary Audacity 4 adds direct multi-clip editing, flexible workspaces, native high-DPI rendering, and a new project format.
The details
Clips can now be directly multi-selected, grouped, moved between mono and stereo tracks, and moved over other clips by replacing overlapping audio.
The Qt-based interface adds native high-DPI rendering, dockable or floating panels, saved workspace layouts, and Modern, Classic, and Music workspaces.
Audacity 4 introduces a new project format.
Audacity 3 projects can be converted without altering originals.
Converted projects cannot be saved in the earlier format.
Macro Manager, scripting pipe, VAMP, and LADSPA plugin hosting are unavailable in 4.0.
Supported plugin formats are VST3, Nyquist, LV2 on Linux, and Audio Units on macOS.
Repos
3 items
Relevant open-source projects, with the adoption signal separated from the headline.
Open Source Radar+1,601 stars today
Matt Pocock releases composable engineering agent skills