Duckity is an abuse-protection service for app APIs that combines rate limiting, network signals, and VDF-backed computational challenges without prompting users with visual or auditory puzzles. It processes only client IPs and requests, stores and processes data in Frankfurt, Germany, and offers SDKs for web and native platforms, with a free tier of 10,000 challenges per month and a Pro tier at $49/month for 100,000 challenges.
This interactive page applies data envelopment analysis (following Zhou et al., 2022) to compare AI models' API cost per 1,000 benchmark tasks against their benchmark scores, identifying cheaper models or blends that match a given score. It notes that savings are benchmark estimates excluding latency, quotas, switching and other deployment costs, and that users should test alternatives on their own tasks. The single-slide creation task uses the free disposal hull variant and compares single models only, since its preference ratings cannot be averaged.
Anthropic released Message Threads as a public beta on its Claude platform, which transmits only what has changed since the last message instead of resending the full conversation history. According to Anthropic, this cuts time to first token by about 35 percent for requests between 500,000 and one million tokens, with coding agents, computer-use agents, and document- or image-heavy workflows benefiting most. Threads stay encrypted in memory between requests and expire after five minutes of inactivity, token usage is unchanged, and the feature works on the Claude platform and through AWS, Microsoft Foundry, and Google Cloud.
Zingg describes how adding a simple steward approval flag to its incremental entity resolution and MDM platform raised harder questions about what a human decision locks in, how long it holds, and who may reopen it when underlying data changes. The post lays out design principles (members stay put, no automatic merges, growth allowed), walks through concrete edge cases such as a father-and-son household split, and argues that the system being corrected should not control when a human decision expires. It closes with general guidance for any system where agents and people share decisions over time.
Angle Health, an AI-native health benefits technology company, piloted ElevenAgents voice agents for provider-service workflows, answering calls in eight languages and working directly with its core administration system. During the pilot, the agents handled about 80% of total provider call volume, with more than 90% resolved end to end without transferring to a person, and the provider team gained back nearly 9,000 minutes a day. Angle Health plans to expand the pilot to its commercial, brokerage and account management segments.
Tests show that GPT-6 Luna and Claude Opus 5.5 each matched the live price for only 2 of 15 products from memory, and built-in web search improved exact matches but still produced wrong prices with source links on every run. The article argues that LLM grounding requires live data, a clear status when a check fails, and code rather than the model writing each price into the answer, and it builds a Python agent using Apify's E-commerce Scraping Tool to fetch prices from Amazon, Walmart, Best Buy, Target and Shopify stores.
Heimstaden is using OpenAI's ChatGPT Enterprise and Codex to transform property management, cut costs, and empower employees across Europe.
Evolv is a local-first, single Go binary AI agent workspace that rewrites its own UI, tools, skills, prompts and plugins live while running, with changes hot-reloaded over WebSocket without refresh or restart. It ships with snapshot/rollback time travel, a Dream mode for idle self-evolution, a 5–100 agent swarm mode, and support for Ollama, the free Kilo Gateway, or BYOK providers. The agent can also edit its own Go kernel and any user project folder, with evolved components security-linted before mounting.
LEANOX.AI is an AI gateway that sits between an application and existing model providers, caching only content likely to be reused and continuously optimizing so more requests qualify for cached pricing. Integration requires pointing the SDK base URL at LEANOX.AI and sending a LEANOX.AI key, with provider accounts and billing unchanged. The post claims up to 90% lower cost for repeated tokens, with highest savings for Anthropic and Google Gemini and moderate, model-dependent results for OpenAI.
This Show HN post introduces an AI object counter that counts items in uploaded JPG, PNG or WebP photos, marking each detection so users can review the total before trusting it. It offers industry-specific counters (bottles, pills, bricks, logs, coins, livestock and more), a tally sheet with CSV export, and Premium options including bulk audits, PDF reports and shareable team links. The page stresses that counts are drafts based on visible pixels only, so hidden or stacked items may be missed and totals should be reviewed or recounted by hand when reported or paid for.
PyTorchKR and XCENA co-hosted KVCache Meetup Korea on Sep 30, 2026, with Tensormesh sponsoring, featuring talks from XCENA, SK hynix, FuriosaAI, and LMCache on LLM inference and KV Cache optimization. Sessions covered CXL memory expansion and the Maru KV Cache storage engine, GPU and HBM utilization analysis with H100 GPUs and vLLM, request scheduling and KV Cache management in Furiosa-LLM on RNGD, and LMCache's MP Mode for unified KV Cache management. A survey of 72 responses cited the program (66.7%) and venue and facilities (41.7%) as positive points, with requests for better audio, screen visibility, interpretation, and real production examples.
Apify's Scraping Toolbox offers 15 production-grade data APIs, including Google Maps, LinkedIn, Instagram, TikTok, YouTube, RAG web crawling, review monitoring and AI-visibility tools, priced per event from $0.002 per result plus run fees where applicable. Claude, GPT, Cursor and n8n can discover and call the tools through the Apify MCP server, and agents without an Apify account can pay per call in USDC via x402 on Base.
Transformers introduces Model Integration Contracts (MICs), which pair a shrunk fixture model built from a real checkpoint's configuration with a runnable usage path and human-reviewed baselines, run on cheap CPU runners in PRs. Replaying frozen MICs over five months of commits showed a small UMT5 fixture flagged a decoder-attends-to-future-tokens regression from July 30 to September 28, 2026, which a forward pass alone would have missed. The tmic CLI and agents help draft contracts, and the MIC CI job currently reports failures without blocking PRs.
OpenAI introduced Pando, a Rust-written Git-compatible server backend now serving all development in its monorepo, which keeps the standard Git client and protocol while reporting median push latency down nearly 80% and nearly 1 PB of Git data per day. Pando uses per-branch causal consistency instead of repository-wide ordering, stores each pushed pack as an immutable blob with ref versions in a metadata store, and serves fetches via packfile-uri download links through a regional cache called Express with hit rates above 99%. It has served production traffic since April 2026, is not open source, and all performance figures are self-reported by OpenAI.
Agents in Amp can now collaborate with users and teammates in Artifacts, documents that belong to a single thread and that the agent, the user, and teammates can all edit. The agent writes an artifact instead of a long reply when it has something to read, such as a plan, report, or call notes, and when it returns to the artifact it reads the user's edits first and works them in rather than overwriting them. Users can open Views → Artifacts in any thread or ask Amp to write one.
agent-desktop is a CLI and C library for macOS, Windows, and Linux that reads a desktop app's accessibility tree as JSON, giving each actionable element a ref that an agent acts on and then re-reads the UI to verify. It contains no model and no GUI, prints one JSON envelope per command, and covers desktop apps only, not web pages. It is limited to what apps expose through their accessibility APIs (AX API, UI Automation, AT-SPI).
OpenRun, an open source self-hosted GitOps platform for deploying web apps and internal tools declaratively, argues that AI coding tools have removed most of the speed advantage of low-code platforms' visual editors, since generating a small app from a prompt now takes about as long as assembling it on a canvas. The post contends that low-code downsides remain—proprietary app formats, editor-only editing, limited testing and refactoring, per-user pricing and security features gated to higher tiers—while code-first apps benefit from reviewable pull requests, easier maintenance and no lock-in to a tool or model. It also notes that code-first internal tools still need a platform for hosting, SSO, access control, audit logs, database credentials and deployment, which is the gap OpenRun targets.
PhoneUse lets AI agents like Claude Code, Codex, Cursor and Windsurf control phones via MCP, API or CLI, using either cloud Android devices or real retail iPhones tapped by a robot stylus with an overhead camera. It offers 12 tools including phone.acquire, screen.read, tap, type, swipe, app.open and sms.read, with real phones providing genuine carrier SIMs for SMS and 2FA. Pricing starts at $20/mo for cloud phones ($0.024/min) and $149/mo for real iPhones.
Claude Code v2.1.317 adds a warning to /schedule when a new cloud scheduled task receives connectors you did not request, a confirmation before a web or shell tool call sends your account email address to the network, and onFailure: "block" for Agent SDK hook callbacks. It also adds anthropic.plan, anthropic.seat_type, anthropic.provider and anthropic.custom_base_url to OTel metrics, events and spans, plus a cost report endpoint at GET /v1/organizations/analytics/cost on the Claude apps gateway admin API. The release includes numerous fixes across MCP servers, auto mode, self-hosted runners, OpenTelemetry exports, VSCode, cloud sessions and Remote Control.
LLM 0.37 adds gpt-6.1-sol and gpt-6.1-terra, a new `llm openai endpoint add` command for configuring OpenAI-compatible endpoints, and switches the default model to gpt-6-luna at half the price of the previous default, gpt-5.6-luna. Embeddings support now allows different binary content to have different content-types, needed for models like Gemini Embedding 2 that can embed text, audio, images, and video. The release includes fixes from 12 authors, almost all AI-assisted, which the author says he treats as bug reports with attached examples and found less burdensome to review than expected.
OpenAI describes how it built Pando, a Git-compatible backend designed for the scale of development that Codex makes possible.
In distributed AI workloads, a single GPU or fabric event can cascade into NCCL errors on other ranks, scheduler state changes, and thousands of downstream messages, so the reported exception shows where the workload stopped rather than where the failure began. The post explains how to run NVIDIA Attribution on sample workload evidence.
In this podcast episode, Legora CTO Jacob Lauritzen discusses how the legal-tech company builds and tests agents whose traces it cannot read because the data belongs to law firm clients, using simulated test environments modeled on customer work. He covers model choices and the addition of a model picker, an estimate that models already do about 60% of legal work well, and a citation fix that took accuracy from 70% to 99% in one release. The conversation also touches on Tabular Review, running everything on Azure, forward-deployed and legal engineers, and the future of law firm partnerships, the billable hour, and junior lawyer roles.
Jacob Lauritzen, CTO of Legora, discusses building AI agents for lawyers after the company's $550M Series E at an $8.8B valuation, covering how its evals team simulates customer work with three engineers and three lawyers because client data is confidential, and how Legora runs everything on Azure for law-firm approval. He says models already handle about 60% of legal work well, that complex analysis and editing across million-token contracts still need more intelligence, and that Legora has begun post-training open-weight models for narrow tasks like legal research. The company passed $100 million in ARR in April, less than 18 months after launch, and now has more than 1,800 customers.
As distributed AI training scales, failures across hardware and software dependencies increasingly disrupt training progress, making fast, reliable recovery essential. The material notes that a stalled training rank can trigger an NCCL timeout, a GPU node can stop responding to the scheduler, or a straggler can slow the entire job, and that each disruption is manageable on its own.
CloudOpsBench is a benchmark for AI agents on real cloud operations, and its leaderboard ranks models by resolution rate, harness, release date, time, tokens and cost. The top entry, claude-fable-5-1 (max) with claude-code, reaches a 60.6% resolution rate at 38.6h, 97.9M tokens and $404.3, while the lowest listed, kimi-k3.1-pro (high) with terminus-2, reaches 19.1%.
AI-TRANSLATOR is a synchronous speech and text translator whose translation core emits normalized events to an event bus, which fans them out to any number of independent text sinks (console, WebSocket display page, transcript, Telegram) and voice sinks in parallel. Stage 1 (MVP) is implemented and tested, with Linux as the primary platform, macOS and Windows x86_64 supported for building and running, and a Telegram voice-chat sink built on the libntgcalls media engine. Interfaces may still change before the 1.0 release.
In this Arrested DevOps podcast episode, host Matty Stratton talks with Sylvain Kalache, who leads an AI lab at incident management platform Rootly, about how AI-assisted incident response tools (often called "AI SRE") combine long-term memory such as codebases, incident reports and Slack history with live signals to generate and investigate hypotheses. They discuss where these tools excel on simple incidents versus complex ones, the need to keep humans in the loop, and the risk that automating away easy incidents erodes engineers' troubleshooting skills, drawing a parallel to aviation's autopilot and simulator training.
Lyngvi is an MIT-licensed, agent-agnostic sandbox for running autonomous coding agents against your own repositories on macOS and Linux via Podman, driven from a JavaFX desktop app or a CLI that share one Java core. It gives each conversation its own container with no default network route (egress only through an allowlisted Squid proxy), a copied working tree with a fresh .git, and secrets held in a gateway sidecar, with changes returned as host-classified patches (apply, hold, reject) committed only to a working branch. It ships reference adapters for Claude Code and OpenCode, supports local model servers through the proxy, and pins its base images by digest.
Sierra has redesigned Agent Studio around Ghostwriter, its always-on agent for building agents, so users can investigate issues, edit journeys, run simulations, and test fixes within a single conversation. Ghostwriter now follows users through Agent Studio and sees what they are viewing, combining capabilities previously spread across Explorer, Sierra Assistant, and other specialized agents, with each piece of work kept in its own shareable or forkable session. The redesign also brings related agents into a single project, adds line-by-line change comparisons, Preview Chat and Simulations, and release governance with required approvals and controlled rollouts.
Brendan Falk, who co-founded the developer tools startup Fig and sold it to AWS in 2023, now runs Hercules, a seven-person company he says is doing tens of millions in revenue by letting non-technical people build production-grade business software such as ERPs and CRMs. He describes Hercules's bottoms-up, milestone-based approach, its use of low-CAC Google Ads in emerging markets for product feedback, and his view that custom software is a revenue-maximizing strategy rather than a cost-saving one.
Databricks announced that Lakebase Postgres is generally available on Google Cloud, offering fully-managed, serverless Postgres on the Databricks Data + AI Platform with compute separated from storage and a shared storage layer with analytical data. Features include serverless autoscaling and scale to zero, instant branching, point-in-time recovery, standard Postgres with pgvector-compatible search, and unified governance via Unity Catalog. GA regions are asia-southeast1, europe-west3, us-central1, us-east4, and us-west1.
The article argues that AI agents, being powered by probabilistic LLMs, cannot reliably recognize risky instructions, so behavioral safeguards like system-prompt rules are insufficient and architectural safeguards are needed. It describes emerging infrastructure—sandboxes, credential brokers that attach secrets on egress via an HTTPS_PROXY, and gateways for granular permissions—that deterministically prevent harmful actions. The piece is published by Infisical, which promotes its Agent Vault product at the end.
The Biological Computing Co. (TBC) launched TBC mischief 1, an image-to-video model built on Wan 2.2 that generates video 5x faster than its base model and is available on fal and OpenRouter at half a cent per second. TBC says the speed and cost gains come from efficiencies it discovered by measuring living neurons in the lab and applying them as proprietary software that adds less than 0.1% to the base model's parameter count. The company says it is already applying the process to newer open-weight video models and developing applications in language models and physical AI.
A 30-day analysis of a headless Shopify fashion storefront found that ChatGPT and Claude fetched pages 1,264 times for users while GA4 recorded only 122 visits from AI answers, with ChatGPT alone reading pages about nine times per visit. The logs also showed assistants repeatedly requesting discontinued products that returned soft 404s, Claude probing standard Shopify paths the storefront doesn't serve, and roughly 10,600 requests from fake AI crawlers that failed IP verification and mostly probed for secrets. The authors note their own dashboard initially overstated ChatGPT arrivals by more than tenfold due to utm tag handling in referer headers, and recommend checking 404 status codes, adding redirects for standard Shopify paths, and verifying bots by IP.
devTopBars is a paid macOS menu-bar app built on the free, open-source cpuTopBars, adding bars for Claude Code's 5-hour and weekly usage limits and for each live session, plus LAN/WAN/VPN address rows and a keep-awake option. It requires macOS 27 or later and a free open-source usage-feed plugin installed once in Claude Code, and it is currently in open beta via TestFlight ahead of an App Store release. The app sends no session or user data anywhere, making only an ipify WAN lookup on click and a per-launch GitHub check for plugin updates.
HiveOS is presented on Hacker News as an AI chief of staff featuring sub-1ms PII masking and offline swarms. The submission links to gethiveos.com and has drawn 1 point and 1 comment so far.
Anthropic announced that message threads is in public beta on the Claude Platform, letting agents send only new messages plus the previous response ID instead of resending full conversation history each turn. With message threads on by default for most Claude Code users, requests of 500k to 1M tokens saw median time to first token drop by about 35%, though token usage stays the same since Claude still reads the full conversation each turn. Threads are held encrypted in memory, expire 5 minutes after the last request, and are available on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed).
Why it matters: It gives developers concrete latency figures and setup details for a new API feature aimed at long-running agent sessions.
Startups including Ornn, Silicon Data, Architect and Liquid Compute are working to create standardized GPU pricing and commodities-style futures and brokerage services, with Ornn reportedly in a new funding round at a valuation of around $1 billion. The sector faces hurdles: none of the companies has obtained financial regulatory approval, CME's planned listing tied to Silicon Data's indexes has been pushed back from its October 5 target, and the CFTC has requested public comment on whether compute is standardized and liquid enough for futures. Rapid GPU depreciation and uncertainty over whether the compute shortage is permanent have led some investors to pass on deals in the space.
Deep Principle has released Hydrogen-1.0, an MIT-licensed 753B-parameter open-weight model built by post-training Z.ai's GLM-5.3 on chemistry, materials science and life sciences data, laboratory records and agent discovery trajectories, with a 1M-token context and 128K max output. Deep Principle reports it scores 38.3% on FrontierScience-Research versus 33.3% for GPT-6-Astra and 93.0% on GPQA-Diamond, while trailing GPT-6-Astra by 3.33 points on NMR structure elucidation. Weights are on Hugging Face and ModelScope, with BF16 serving tested via SGLang on 8 GPUs, NVFP4 builds for NVIDIA B200/B300 and MXFP4 for AMD MI350X/MI355X, though full datasets and recipes are not released.