iCreatAI News
Back to main site
Skip to content
Only Tutorials
Oct 30Fri
  1. Meta Engineering58

    Catching Bugs Before They Land: JiTTesting’s Human-on-the-Loop Approach at Full Scale

    Meta reports that its Catching JiTTests system, which uses LLMs to generate executable tests that fail on risky code diffs before they land, is now part of its production infrastructure rather than a research prototype. The system combines intent-aware and "dodgy diff" test-generation workflows with an LLM-as-Judge ensemble and a rule-based assessor, and uses a human-on-the-loop model where engineers review outcomes at an aggregate level. Meta says 589 engineers (64% of outreach) engaged with the system, which showed a 60% true-positive rate and helped catch 74 serious privacy bugs at the diff stage in the first half of 2026, lowering privacy SEVs by 20%.

  2. Hacker News AI58

    Make your AI skills more predictable with scripts

    This newsletter explains that AI skills are folders containing a SKILL.md file with a name, description and instructions, and can also bundle scripts and other resources that agents run when needed. It walks through a lead-generation skill that starts as a single prompt and is then refactored with three scripts (web_search.js, qualify_lead.js, store_lead.js) plus updated instructions, moving repeatable steps out of the agent's judgement and into code. The author argues this yields more predictable outcomes and is easier to change, since each part of the process lives in one place.

  3. AWS Machine Learning Blog41

    Accelerate agentic AI with NVIDIA Nemotron models on Amazon EKS and SageMaker HyperPod EKS

    NVIDIA Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts model activating 12B parameters per token that handles text, image, document, and video input, joining other Nemotron models already on AWS. The post describes its hybrid Mamba-2, MoE, and attention architecture with a RADIO vision encoder and up to 1M-token context, then shows deployment on Amazon EKS, EKS Auto Mode, and SageMaker HyperPod EKS using NVIDIA Dynamo and vLLM, with validation and performance results on B200 and H200 instances.

  4. NVIDIA Blog40

    Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents

    NVIDIA describes a series of developer projects in which frontier AI models such as GPT-6 Astra are directed through natural-language instructions to build simulations using Omniverse libraries for GPU-accelerated physics, rendering and sensor simulation. Examples include recreating an OP-1 synthesizer in Houdini, testing robotic disassembly in Isaac Sim, reconstructing buildings from site scans, and a gas-turbine what-if application. Other projects cover SimReady asset scene assembly, warehouse performance scaling, cable routing, CAD-to-SimReady conversion, humanoid simulation, autonomous-driving testing, sensor-based digital twins, Robo Olympics, an ISS browser model and room reconstruction.

  5. Hacker News AI34

    Running Self-Hosted AI Agents on Oxide

    Oxide explains that AI agents should run in virtual machines rather than containers for stronger isolation, with an orchestrator VM provisioning ephemeral VMs for each agent session and cleaning them up afterward. The post walks through Claude Code self-hosted environments, where a spawn-runner hook creates ephemeral VMs and a separate reaper deletes stopped ones, and Amp self-hosted orbs, where a dispatch script handles spawn, pause, and delete operations. It also outlines future work on security, multi-rack scale, cold-start latency, and a possible microVM primitive.

  6. Hacker News AI22

    Serving Prefill and Decode for Agentic AI (On Tenstorrent Blackhole)

    Tenstorrent describes how its Galaxy systems, built on Blackhole chips, serve both prefill and decode phases on the same hardware so operators can shift capacity between them as agentic workloads change. The post explains how agent loops produce growing, largely cacheable context, models how faster decode only partly shortens dependent-call latency, and contrasts prefill's compute-bound long prompts with decode's memory-bound small batches. It cites a Wormhole FlashAttention kernel comparison (9 to 44 times faster than a basic implementation) as a measured example of data-placement control.

Oct 29Thu
  2. Hacker News AI58

    You can just ask AI to do things

    The author argues that a key skill for working with AI is agency—being willing to try ambitious things—and lists everyday examples from his work: having agents start his dev server, run curl requests against it, poll metrics and logs during production changes, get second opinions from other models, do a 6am pass over unread messages and logs, improve their own setup, and manage their own sessions. He notes he still writes his own messages and reports, makes his own design decisions, and reviews code himself, and is often ruthless with agent suggestions.

  3. Hacker News AI36

    Slop-Free AI Programming

    Karl Zylinski describes how Claude's verbose comments overwhelmed his review process and led him to accept code he didn't fully understand in his Odin library Karl2D, so he adopted strict CLAUDE.md rules: no comments in generated code, imperative procedural style, and no commits or pushes without explicit instruction. He argues AI has encyclopedic knowledge but no taste, recommends hand-writing architecture and reviewing changes as separate commits, and estimates roughly 5x speedup for new code with these rules versus 20x for unchecked vibecoding.

  4. Hacker News AI22

    How to Make an Outstanding Tiny Desk AI Video with Your Own Photos

    This guide explains the Outstanding Tiny Desk AI trend, inspired by Charlie Wilson's NPR Tiny Desk performance of "Outstanding," in which users upload 1–5 photos of people or pets to be placed into a 15-second vertical 9:16 concert video. It walks through the steps at outstanding-tinydeskai.com: assigning photos to the lead vocals, keys, backup left, backup right, and sax positions, then choosing 480p (50 credits) or 720p (100 credits) and generating the video, which includes the original audio and no watermark. It also offers tips such as using clear, front-facing photos without sunglasses or heavy filters.

  6. ElevenLabs Blog25

    How we use ElevenAgents to save 1,200 hours a month in daily operations

    ElevenLabs says its internal agents save teams over 1,200 hours a month, answering more than 5,000 questions monthly and taking actions such as building customer demos and opening over 150 tickets in the last month. The agents span sales (GTM and FDE bots), support and engineering triage, and product data analysis, with lessons including schema curation over prompt engineering, injecting Salesforce picklist values into tool type definitions, pruning overlapping procedures, and verifying tickets after creation.

  7. Hacker News AI35

    Open Source or Custom in Enterprise AI Systems?

    This article argues that as AI makes code generation cheaper, engineering judgment becomes more valuable, and that velocity metrics like PR volume or lines changed are weak signals of progress. It explains where quality breaks in AI-assisted systems—at integration boundaries, retries, observability, and governance—and recommends building on proven open-source components, applying Twelve-Factor and annotation discipline, then adding AI-assisted implementation with strict review, testing, and monitoring. It cites an advisory project for a U.S. industrial manufacturer where an internally built agentic system was rebuilt on the OpenAI Agents SDK for greater robustness and maintainability.

  8. Hacker News AI47

    Control Planes of Agentic AI Infrastructure for Reliable Software Delivery

    The article proposes five control planes—Context, Containment, Validation, Rollout, and Accountability—for safely scaling agentic software engineering, and argues that AI-generated code changes should be classified by production risk, with high-risk areas such as authentication, billing, permissions, and data migrations requiring more context, guardrails, and human judgment. It describes practices including architecture-focused context documents and specialist agents, feature flags and scoped cluster roles for containment, behavioral and feature-flag testing, rollback plans designed before deployment, and retained human ownership of high-risk decisions. A running example applies the workflow to modifying authentication systems across multiple deployment types.

  9. Hacker News AI22

    Building a Banana Sorter AI Machine

    The author tests four vision models (Qwen2.5-VL 3B, Gemma 3 4B, Liquid d1-2B, and gpt-6.1-sol) on five banana images to decide BASKET or TRASH over 100 runs each. The smaller offline models trashed any banana with a black dot, while the frontier model sometimes trashed the heavily spotted banana (16% at low reasoning, 38% at high) and Liquid d1 gave probabilistic answers. The author concludes he would not trust a high-end model with a critical decision and that designing a real banana sorter is more nuanced.

  10. Hacker News AI12

    Ask HN: Techniques for productive AI use in the workspace

    A software engineer asks Hacker News how others use AI productively at work, describing their own approach of using Opencode as a pair programmer to work out a solution's general structure before doing a roughly 35% AI and 65% personal pass of writing the code. They say they enjoy writing well-designed code and are weighing whether their style of development has a sustainable place in the field.

  11. Hacker News AI50

    Build a better AI product with explicit classification

    Dhruv Dang's newsletter explains AI classification, contrasting general-purpose models like GPT and Claude with purpose-built decision models such as Jev and Laya, and argues that explicit classification—making classification a separate step—gives teams granular control and flexibility over agentic workflows. A detailed example refactors an AI blogging agent so an intent-agent classifies each prompt into draft, save, publish, or unsupported requests, letting application code enforce rules like saving before publishing and route work to a sandboxed drafting-agent.

  12. Hacker News AI46

    Tools, 17 with write access: How AI platforms bound agent authority

    Antalpha counted its own MCP server's 120 tools and found 17 can write, 16 of them in Trading, and proposes a five-grade scale from architectural to no guardrail for judging agent authority. It compares Binance AI Pro's isolated-account boundary, OKX's post-hoc staked evaluator network and Minara's procedural confirmations, noting each answers a different question. The tool inventory is verifiable via the public tools/list endpoint at mcp-skills.ai.antalpha.com/mcp.

  13. Hacker News AI36

    Agentic CI/CD with Azure Cosmos DB: Isolated Envs for a Swarm of AI Agents

    The post describes giving each AI coding task its own disposable Azure Cosmos DB emulator environment, seeded from versioned fixtures, so parallel agents can use identical document identities without colliding. It walks through a two-environment isolation proof, a GitHub Actions matrix workflow with per-task emulators, and a swarm check that runs three worktrees concurrently, merges branches into an integration worktree, and runs combined tests against a fresh emulator. A real Copilot CLI run passed all five jobs, with three tasks passing 18, 18, and 57 tests and the integrated revision passing 75 combined tests in 140.802 seconds, though the author notes this does not establish semantic conflict resolution, arbitrary scale, or production readiness.

  14. Databricks Blog28

    How Procore is powering construction analytics with OpenSharing

    Procore rebuilt its Procore Analytics product on a Databricks medallion architecture and OpenSharing, replacing a SQL Server backend that capped refreshes at one to three per day and relied on costly, brittle recursive API calls. The new design shares hundreds of curated, Unity Catalog-governed tables per customer at schema, view, table, and partition level, with Change Data Feed for incremental syncs and automated per-user, per-project entitlements. Procore says the catalog expanded to hundreds of tables in less than half the time the first generation took, and customers can now refresh on their own cadence.

  15. OpenRouter Announcements65

    How to Prevent Prompt Injection: 13 Attacks, Tested

    The article describes testing 13 self-written and 12 model-written prompt injection attacks against four current models in a document-summarizing assistant with an email tool, finding that text-only attacks rarely worked but that one model (GLM 5.3 Flash) sent 58 of 60 attack emails without defenses. It recommends labeling untrusted content as data (cutting attack emails from 60 to 13 of 720), using a guardrail with a model allowlist and budget limit, checking every tool call with Armor (which blocked all 73 attempted attack emails at about 0.3 seconds added latency), and keeping secrets out of the prompt. Armor blocked real work in one case: a user approving a document's request without naming the address.

    Why it matters: It offers a concrete, measured comparison of layered prompt-injection defenses, including where a tool-call guardrail blocks legitimate user requests.

  16. OpenRouter Announcements62

    How to Extract Tables from a PDF with an LLM

    The guide shows how to send a PDF to a file-reading model together with a strict JSON schema describing one row, so the model returns typed rows with cleanup such as parenthesized negatives and footnote markers handled by field descriptions. In tests over five PDFs (two scans, 644 cells), every model that read the PDF directly got all cells right, while failures came from the parser route and output token limits, including two silent failures. It recommends rejecting empty results and any response whose finish_reason is "length", and comparing routes on your own PDFs before choosing one.

    Why it matters: Offers a concrete, tested recipe with code for schema-driven PDF table extraction, plus specific guidance on silent failure modes worth knowing before building a pipeline.

  17. Databricks Blog28

    How Procore is powering construction analytics with OpenSharing

    Procore rebuilt its Procore Analytics product on a Databricks medallion architecture and OpenSharing, replacing a SQL Server backend that capped refreshes at one to three per day and relied on costly, brittle recursive API calls. The new design shares hundreds of curated, Unity Catalog-governed tables per customer at schema, view, table, and partition level, with Change Data Feed for incremental syncs and automated per-user, per-project entitlements. Procore says the catalog expanded to hundreds of tables in less than half the time the first generation took, and customers can now refresh on their own cadence.

  18. Hacker News AI58

    Our AI Agent Executed the Same Successful SQL Query 254 Times

    Funnel compared the open-weight model GLM-5.2 against Claude Opus 4.8 in a randomised A/B test on its internal data agent platform Reef, finding GLM averaged 7.8 seconds per model call versus 18.1 seconds for Opus but made 66% more tool calls (19 vs 11) and took 14% longer end-to-end. In two separate runs GLM executed the same successful SQL query 113 and 254 times, in one case repeatedly querying a tool that only exposed a 20-row preview. The author attributes much of the problem to the agent harness—nonexistent Athena tables in saved instructions, ambiguous tool contracts and missing safeguards against repetitive calls—and recommends explicit tool contracts, bounded agent behaviour and whole-system benchmarking.

  19. ElevenLabs Blog23

    Conversational AI in banking: Use cases, benefits, examples

    This guide explains how banks use conversational AI—NLP-based chat and voice agents that understand natural language, pull account context, and escalate to humans when needed—for 24/7 support, onboarding, disputes, and fraud alerts. It cites CFPB findings on early bank chatbots, McKinsey figures on reduced handling time and call volume, and ElevenLabs customer examples from Customers Bank and Klarna, while detailing security standards such as SOC 2 Type II, ISO 27001, PCI DSS, and Zero Retention Mode.

  20. AWS Machine Learning Blog23

    Accelerating software delivery with agentic QA automation using Amazon Nova Act – Part 3

    This third post in the series describes two QA Studio capabilities for scaling agentic QA across an organization: applications, which group use cases and test suites under an owner and report health via pass rate, a flaky-adjusted 0–100 score, trend charts, and failure triage built on Amazon CloudWatch custom metrics; and team-based user management, which assigns five cumulative roles (Admin, Team Lead, Editor, Executor, Viewer) through Amazon Cognito groups, with cross-team sharing capped by the recipient's role. The backend enforces access through a Lambda authorizer resolving Cognito groups into scopes plus a resource check on application ownership or sharing, and the reference solution is deployed with the AWS CDK.

  21. AWS Machine Learning Blog36

    Distributed training with Ray Train DLCs and EFA on EKS

    This AWS post shows how to fine-tune Qwen2.5-1.5B with PyTorch FSDP across eight NVIDIA L4 GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS, using the published Ray Train DLC (public.ecr.aws/deep-learning-containers/ray:train-ml-cuda-v1.1) without building any image. KubeRay creates the Ray head and worker pods from one RayCluster manifest, while hardware, transport and per-job Python packages are configured from outside the image via resource requests, environment variables and Ray runtime_env. The sample repository includes idempotent deploy scripts that locate g6.12xlarge capacity, provision the cluster and node group, install the GPU and EFA device plugins and KubeRay, and submit the job from inside the head pod, with the script failing if NCCL did not use EFA.

  22. AWS Machine Learning Blog42

    Building an Agentic AI Platform on Amazon Bedrock AgentCore

    This AWS post describes an enterprise Agentic AI Platform blueprint on Amazon Bedrock AgentCore that separates a centrally governed Platform control plane from repeatable, team-owned Workstream cells, with separate inference and tool gateways and a federated operating model. It details multi-account topology, service control policies, a CI/CD pipeline with deployed-runtime evaluation gates and adversarial denial tests, and dependency-ordered retirement, using the eu-west-1 (Ireland) Region. The companion GitHub blueprint lets teams deploy a representative Workstream cell and validate its security and lifecycle contracts end to end.

  23. Google Developers Blog58

    Hill-Climbing Agent Quality with Agents CLI Evaluation

    A walkthrough of the Agents CLI eval loop using the LangChain template in Agents CLI v1.8.0, covering a baseline run (mean 4.0), a system-prompt regression that dropped the mean to 2.75, per-case diffs via eval compare, and a gemini-3.8-flash to gemini-2.5-flash swap whose mean drop traced to a single off-topic case. It also shows scheduling eval runs with cron or Cloud Scheduler, a script that exits non-zero when scores fall past a threshold, and eval optimize for automated prompt hill-climbing on ADK projects.

  24. Hacker News AI58

    C# and NATS.io for agentic AI orchestration

    This post argues that production agentic AI systems are distributed systems, and introduces the open source Synadia Agent Protocol for NATS, a contract (not a framework or runtime) that lets agents of any kind be discovered, prompted, and monitored on a single NATS fabric. The protocol defines three responsibilities—discovery via NATS micro service ping/info on the "agents" service, conversation via typed JSON chunks streamed to a reply inbox with a mandatory first ack and a zero-byte terminator, and liveness via heartbeats on agents.hb.{agent}.{owner}.{name}—along with a subject hierarchy, request envelope, and error reporting conventions. Reference SDKs exist in TypeScript and Python, with chunked attachments, JetStream at-least-once delivery, stronger identity, and audit traces deferred to future extensions.

  25. Pydantic AI Articles74

    Code mode: make your agents 2x faster and 2x cheaper on long tasks

    A benchmark of 5,276 graded runs across eight models on five small Pydantic AI apps found that code mode, where the model writes a program that calls tools inside a sandbox, cut cost by 58% and wall time by 51% on long tasks while raising success by 16 points. Gains grew with task length: at 200 items code mode was 47% to 75% cheaper and 42% to 67% faster, whereas at 5 items it cost 21% to 38% more on three of five tasks. The author attributes the win not to fewer round trips, since models already batch calls, but to less content per request, and runs the model's code in Monty, a Rust Python interpreter with no host access by default and 30-second, 256 MiB limits.

    Why it matters: Offers a detailed, methodology-heavy benchmark with per-model cost, latency and correctness numbers, plus an honest list of where code mode costs more.

  26. Baseten Blog55

    The fastest, lowest-cost batch transcription with Qwen3-ASR

    Baseten optimized Qwen3-ASR 1.7B, an Apache 2.0 open-weight speech recognition model supporting over 30 languages and 22 Chinese dialects with a 5.4% average WER on the Open ASR Leaderboard, for batch transcription on a single RTX PRO 6000. The optimized stack transcribes 7x more audio per second than out-of-the-box vLLM at high concurrency, reaching 2,645x real-time factor on 5-minute clips and 3,128x on hour-long files, with median latency for 5-minute clips at 64 concurrent requests falling from 50 seconds to 7.3 seconds. Baseten reports a cost of about $0.0013 per audio hour, 8x lower than the same model on vanilla vLLM and roughly 277x less than popular closed-weight providers.

  27. Hacker News AI42

    Rent the model, own the harness: how to build a model-agnostic AI agent

    Werk describes building a model-agnostic agent harness for mobile teams, arguing the model understands and proposes while the harness decides what it reads, touches and records. It outlines four practices: calling models through an in-house interface, having agents request a level of work rather than a model, using a consistent input format, and placing stable content first for caching. It notes that agent SDKs handle the model-tool loop but not team-specific concerns like shared records, permissions and approvals tied to what a person saw.

  28. Hacker News AI55

    Progressive Determinism for AI Tasks

    The post describes a workflow in which an AI agent first explores a web application dynamically with Playwright to complete a scenario, then saves a deterministic Playwright test (e.g., tests/add-task.spec.ts) that is preferred on later runs. If the saved test fails after a UI change, the agent falls back to discovery mode, diagnoses the failure, and proposes a minimal repair that is kept only after the original assertion passes. The author notes that Playwright has built-in planner, generator and healer agents supporting this pattern, which they have not yet tried.

  29. Hacker News AI60

    Who Is the Agent Acting As? Why AI Agents Become Confused Deputies

    The article explains that AI agents given a single broad service account act with the platform's authority rather than the user's, so a prompt injection buried in a support ticket can produce a fully authorized write to another tenant's data. It recommends authorizing every tool call as the intersection of user, agent, and tenant permissions, decided in a gateway outside the model that fails closed and logs each decision.

  30. OpenRouter Announcements58

    Hit Your Claude Code Usage Limit? How to Keep Coding

    Claude plans have a five-hour limit and a weekly limit, and when you hit one you can keep coding by running Claude Code on an OpenRouter API key, which has no such windows and bills per token instead. The setup uses three environment variables (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, and an empty ANTHROPIC_API_KEY) placed in a shell function so plain `claude` still uses your plan, and you switch back by typing `claude` when the plan resets. The article reports median session costs of $0.305 on Opus 5.5, $0.155 on Sonnet 5.5, and $0.007 on Haiku 5.5, and warns that Claude Code's printed cost is unreliable because it prices OpenRouter models with Anthropic's table.

  31. OpenRouter Announcements38

    Claude 529 Overloaded: What It Means and How to Fix It

    A 529 overloaded_error means Anthropic's API is temporarily out of capacity for the requested model across all users, unlike a 429, which reflects your own usage limits. The article recommends retrying with exponential backoff and jitter, then routing the same model to another provider via OpenRouter, which serves Claude Sonnet, Opus and Haiku 5.5 through five providers each, and optionally listing a fallback model with `models`. For Claude Code, it suggests switching models with `/model` or running Claude Code on OpenRouter via Ori Harness or three environment variables.

  32. Allen Institute for AI Blog38

    How researchers are adapting olmOCR 2 for historical and archaeological archives

    Ai2 describes how researchers adapted its open olmOCR 2 model: historian Maurice Casey built ArtificeOCR to run olmOCR 2 locally and make his Tropy photo library of archival sources searchable, while TU Wien and Austrian Academy of Sciences researchers fine-tuned olmOCR 2 with LoRA on 57 annotated Carnuntum pottery records, lowering transcription error on 450 test records from 15.58% to 1.11% through their CENTURIA benchmark.

Oct 28Wed
  1. Hacker News AI28

    How to Build a Slow Product in the Fast AI Era

    The developer of Mole, a Mac and Windows cleanup tool, describes spending two months on unscalable work: installing and uninstalling 850 Mac apps one by one to build cleanup rules, testing more than 30 AI tools in real use, and redrawing dozens of icons over three rounds. Mole went from 1.13 to 1.18, added AI tool and System Data cleanup, shipped a Windows version on October 14, and passed 72.9K GitHub stars. The author argues that as AI makes code cheaper, the details learned only by doing the work yourself increasingly set products apart.