Michelangelo AI is now open source, with the material citing Uber-scale usage of 400+ active ML projects, 20K+ model training jobs per month, 5K+ models in production, and 45M+ real-time predictions per second at peak. It offers feature management, model training, evaluation, deployment, and monitoring, with feature management, evaluation, and monitoring marked coming soon. A Python API example shows distributed XGBoost training with Ray and an end-to-end training workflow.
The AI Security Wiki is a reference for securing enterprise AI adoption, covering governance, risk management, secure engineering, monitoring, vendor evaluation and assurance, with pages reusable as templates that specify the audit evidence each activity should produce. It positions itself as the implementation layer—control objectives, evidence, audit procedures and test cases—that works alongside frameworks such as FINOS AIGF, ISO/IEC 42001, NIST CSF, PCI DSS and CERT-In rather than replacing them. The repository includes a 682-case test library, regional compliance hubs for the GCC, EU and other jurisdictions, framework crosswalks, role-based playbooks and downloadable templates.
Modal argues that fixed-resource reservations fit poorly with bursty Sandbox workloads, since a typical Sandbox spends more than half its life using under 6% of its peak CPU, so users over-reserve and pay for peak constantly. Modal's Sandboxes now burst by default from a base request of 0.125 cores and 128 MiB, with active-resource pricing billed each second for what is used, and the company cites 58% savings versus a fixed-size reservation on an illustrative 20-minute coding agent task. It also describes an OOM SLO of below 0.0001% per minute for bursting Sandboxes under 8 GiB and tools like exit snapshots and the termination reasons SDK to reduce eviction impact.
This GitHub Changelog entry is titled "0.171.0-alpha.1" and lists only the site's navigation and product menus, including Platform, Copilot, Actions, Codespaces, Secure Code, and Pricing. The supplied material contains no details about the release itself.
Claude Code v2.1.323 adds an experimental opt-in Code Mode in which Claude writes JavaScript to call its tools, plus plugin engine version ranges in plugin.json and TypeScript hook type checking in claude plugin validate. The release also adds .docx support to Artifact tool asset uploads and fixes a wide range of issues across permissions, sandboxing, sessions, MCP, OpenTelemetry and Remote Control.
On a single NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x GB300 NVL72, reaching 2.52M at peak throughput and 1,810 tokens/s for a single user. The gains come from kernels tuned for Rubin's split memory domains and FP4 tensor cores, specialized low-latency decode kernels, and a rack-wide KV cache and Engram tables shared over NVLink. The post details four optimization groups and reports a 96.5–97% prefix hit rate at high concurrency with workload-aware routing.
Why it matters: It offers a detailed, measurement-backed look at how Rubin-specific memory and MMA features translate into serving throughput and latency for a frontier open model.
This post rebuts four common misconceptions about LTAP: that it is CDC replication, that a single copy forces slow lookups or scans, that open table formats mean eventual consistency, and that Postgres logical replication suffices. It argues Lakebase keeps one logical table with the lake as its durable tier and caches in front of it, so no layer adds a second catalog or source of truth. Each myth is traced to the same misunderstanding of treating every materialization of the data as a second database.
A Show HN post describes Bugsy, an AI automation system that began as a Sentry AI triager and grew into a system powering much of the day-to-day operations of a ~20-person company. It is used for reporting and fixing bugs, code review, metrics, and answering questions, and is available from Slack, the company forum, and channels.
The FAA has expanded its $875 million SMART AI system, which advises air traffic controllers, from Washington, DC airspace to all 22 centers nationwide after it helped reroute 180 flights and saved an average of 10 minutes and nearly 80 gallons of fuel per flight since September 21. The cloud-based tool combines weather, flight path, traffic flow and controller staffing data to recommend schedules, routes and arrival order, though an FAA team in DC validates its recommendations before they reach controllers. Airline executives support the rollout, but the union representing air traffic controllers says it was not involved in designing or testing the system.
Natera's lab engineering team built SWARM (Systematic Workflow Analytics and Resolutions Manager) on Amazon Bedrock AgentCore to capture and deploy institutional knowledge for lab instrument support, using two runtimes (swarm-agent and swarm-hive) that share a versioned library of 33 skills. The system reduced average first response time from 17.4 hours to roughly 8 minutes (about 130x faster) and delivered an automated first response on 96.2 percent of tickets, measured over a 90-day production evaluation period. The post details the architecture, including a custom document pipeline with Amazon Titan Text Embeddings and Amazon Bedrock Knowledge Bases, keyless OIDC-based deployment, Jira integration via EventBridge and Amazon SQS, and a continuous evaluation loop.
Amazon announced that SGLang-Neuron, released by RadixArk with Annapurna Labs, is publicly available through an AWS Deep Learning Container, letting users serve models and generate RL rollouts on Trainium via the same SGLang interface. The out-of-tree plugin runs on native PyTorch, is selected with SGLANG_PLATFORM=neuron, and supports PD disaggregation with KV-cache transfer via NIXL over EFA. The post includes a walkthrough on a Trainium3 EC2 instance using the pytorch-inference-sglang-neuronx container to serve openai/gpt-oss-120b.
Codex 0.168.0 adds TUI model provider browsing and default selection for new chats, an opt-in side conversation pane (features.tui_workspace) with resizable and zoomable panes, audio attachments with rejection recovery, and opt-in animated GIF image generation. It also fixes account-scoped local history, keeps threads on their saved model provider when defaults change, and improves TUI rendering of interrupted replies and reasoning summaries.
vLLM announces day-0 support for Thinking Machines' Inkling 2 Small, a multimodal text-and-vision MoE model with 594B total and 17B active parameters, up to 1M tokens of context, and a native DSpark speculator. The post details the architecture (48 layers, 384 experts with top-8 routing, block attention residuals, gated attention, a 23-layer ViT image encoder, and mixed NVFP4/MXFP8 expert precision) and vLLM's optimizations, including kernel fusion with PDL, fused MoE activation quantization, sequence parallelism for prefill, and a Gluon relative attention kernel, reaching up to 630 TPS/user at TP4 on GB200. It also covers LoRA and RL (via SkyRL) support, a serve command, agentic-workload benchmarks, and accuracy checks on NIAH up to 1M tokens, MMMU-Pro and HLE.
Why it matters: Useful for engineers evaluating or deploying a large multimodal MoE model, as it pairs concrete architecture and quantization details with vLLM-specific kernels, benchmark numbers and a ready serve command.
Nebius outlined $40 billion in planned 2027 spending on its Q3 earnings call, funded largely through customer prepayments and contract-backed loans. Its Microsoft contract, worth up to $19.4 billion, is already slipping, with only just over half of the New Jersey capacity planned for delivery by year-end and the rest in Q2 2027, after operator DataOne wound down temporary generation faster than expected. Nebius cut its 2026 revenue forecast by $200 million to $3 billion and shares fell 17%.
Databricks announced that Genie Migrate is now generally available, aimed at helping companies migrate to the Databricks Lakehouse as they race to build AI agents and applications. The supplied material does not include further details on the tool's capabilities or migration process.
CopilotKit released OpenBot 1.0 on November 9th, an MIT-licensed, self-hostable agent workspace where teams bring their own models, agents and infrastructure, and every Bot action is checked against a CEL policy through a single gateway before it runs. Each Bot gets its own container with a Chromium browser, logins and files, and any AG-UI agent plugs in, with support for OpenAI, Claude, Gemini, Grok and open models such as Kimi, DeepSeek, Qwen, Nemotron and Mistral. Personal use is free and runs locally, while the Business tier costs up to $29 per active user per month and adds SSO/SCIM, on-behalf-of tokens, spend limits, SIEM export and multi-tenancy.
The post explains how four request-level Amazon Bedrock settings—prompt cache duration, service tier, reasoning effort, and maxTokens—can change token processing and billing rates without altering the model or prompt. It reports tests showing that a 1-hour cache pays off from the second request per hour when gaps exceed 5 minutes, that Flex on Nova 2 Lite processed the same tokens at half the Standard rate in us-west-2, that lowering effort from high to low cut median output tokens by 169% on an agent test, and that a 155-token maxTokens fence avoided 1,893 output tokens in one test. It recommends confirming model support, changing one setting at a time, and verifying effects through response metadata, CloudWatch metrics, invocation logs, and billing data.
This AWS post describes how to get per-identity Amazon Bedrock cost visibility and enforceable spend limits using only AWS-native features: IAM principal and session tags for attribution, Cost Explorer or CUR 2.0 exports (with the caller identity option) for reporting, and AWS Budgets with budget actions that attach a Deny IAM policy when a threshold is crossed. It notes that budget data refreshes up to three times a day, cost allocation tags are not retroactive, and a Deny action on a shared role blocks everyone using it.
Eclipse Aerospace worked with Databricks to replace a multi-stage AWS Lambda and Avro pipeline with an Arrow-native PySpark custom data source that wraps its existing Rust flight-data decoder, reading proprietary telemetry files directly in Spark on serverless compute. The approach, enabled by the Python Data Source API available from Databricks Runtime 15.4 LTS, cut latency, removed duplicate Avro storage and a Lambda tier, and brought decoding into the Spark UI and job metrics. Eclipse reports the connector can process records within a minute of upload and commonly handle multi-gigabyte batches in parallel in under two minutes.
Constellation, the largest US nuclear fleet operator, is building a governed Enterprise Data Platform on Databricks to give data consistent shared meaning across domains, choosing Databricks partly because it interoperates with existing architecture rather than requiring a rip-and-replace. The platform rests on conformed semantic definitions, a semantic knowledge layer with selective Data Vault modeling, and Unity Catalog for lineage, fine-grained access and ownership of meaning, with early use cases including a nuclear engineering assistant built with Databricks Genie, outage planning, and sales intelligence.
Headwig introduces "Wisdom" groups that bundle an app's services so they can be built and run together, such as a frontend and backend as two services in one Wisdom. Users create a Wisdom via New Wisdom, then add services by clicking the three dots beside the Wisdom and choosing Add Service. The post also points readers to Discord for questions.
This post describes a pattern for a self-service LLM gateway that runs LiteLLM on AWS Fargate to expose Amazon Bedrock models through an OpenAI-compatible API, using Bedrock Application Inference Profiles for per-model quota isolation and scoped IAM access. It covers configuration of developer tools, LiteLLM model mapping with per-token costs for budget enforcement, CDK constructs for inference profiles, Langfuse observability, CloudWatch TPM alarms, and lessons learned from scaling to several hundred developers.
This AWS post argues that model cost should be measured per completed task rather than per token, citing four hidden cost drivers—tokenizer efficiency, output verbosity, agentic loop efficiency, and reasoning overhead—and an illustrative example where a model with double the list price costs 65% less per task. It describes levers for reducing cost per task (prompt caching, intelligent prompt routing, output constraints, batch inference) and Bedrock features such as CountTokens, the usage block, and invocation logging for measuring token and turn counts. For ROI, it recommends documenting and measuring a human baseline before launch, instrumenting workflows as experiments with run records, joining cost to outcomes via Athena, and using a five-dimension scorecard with green/yellow/red thresholds.
OwnOwl is a time-limited test build that runs a large language model entirely on an iPhone, with no internet used for chatting and conversations kept on the device. It requires an iPhone with 6 GB of RAM or more and iOS 17 or newer, plus about 6 GB (GigaChat), 9 GB (Ling), or 14 GB (Gemma) of free storage for the chosen model, and is installed via AltStore or Sideloadly using your own Apple ID. Version 0.2 adds faster first replies, a Continue button for longer answers, memory of user-confirmed facts, and stability fixes; the app and its inference engine are proprietary and the build expires.
The author explains founding the Inference Cooperative, a member-governed AI inference co-op registered at inference.coop and fiscally sponsored by the nonprofit Metagov, after exploring inference as a point of intervention for democratizing AI. Built with open-source tools and help from an agent called InferenceBot, it grew past 50 members in about a month and adopted a charter letting members set binding policies and remove the general manager. The co-op draws on US nonprofit law, Swiss and EU providers for green energy and privacy, and Chinese open-weights models, while the author acknowledges the models remain stained by stolen data and dirty energy.
The final installment of Wiz's series argues that AppSec's bottleneck is remediation, not detection, citing Verizon DBIR data that median patch time has risen to 47 days from 32. It describes how Wiz Code maps cloud findings to code owners via VCS connections and correlated data, and how the Wiz Green Agent generates tailored remediation plans that can open pull requests or hand off to coding agents like Claude Code. Wiz Workflows then routes tickets and alerts to Jira, ServiceNow, Slack, or Teams with full scanner and runtime context.
Databricks introduces Lakehouse//RT, a real-time analytics engine built to serve high-concurrency, low-latency workloads directly on governed lakehouse data without costs scaling proportionally with query volume. It claims sub-100 millisecond latency at 12,000 queries per second, node-level autoscaling, and elimination of separate serving layers like Druid or Pinot, with cited customers reporting 6x to 10x lower infrastructure costs. Lakehouse//RT is available in beta, with a 30% discount through January 2027.
The paper proposes AIR-LLM, an edge LLM inference architecture in which a central radio such as a 5G base station broadcasts LLM weights over the air and edge users perform the general matrix-vector multiplication of inference directly in the RF domain using RF mixers, avoiding weight storage and loading. It uses MIMO spatial multiplexing with an energy-efficient precoder-postcoder pair to shorten airtime, and because the central radio stays user-unaware, one broadcast can serve unlimited users in coverage. Evaluated on NVIDIA Sionna ray-traced channels of two urban scenes with a real RF mixer profile, AIR-LLM shows 4.0% WikiText-2 perplexity degradation on LLaMA-3.1-8B while saving 157.7x/40.4x energy versus FP16 and weight-only quantization baselines, and with 20 users its airtime is 104.1x/26.0x shorter.
AWS announced updates to the Amazon Bedrock AgentCore runtime that keep each session in its own hardware-isolated Firecracker MicroVM while starting sessions up to 15x faster, scaling to 4x more compute (up to 8 vCPUs and 32 GB, with container images up to 10 GB), and running up to 68% more efficiently. The runtime now reclaims memory within minutes of release and scales idle agents to zero, and agents can dynamically mount and save work to a filesystem such as an Amazon S3 bucket with FUSE support so data stays inside the trust boundary. The next generation runtime is available today in AWS Regions across North America, South America, Europe, and Asia Pacific.
Why it matters: Announces concrete performance, scaling, billing, and data-isolation changes to a managed agent runtime, with customer metrics and availability details useful to teams planning production agent deployments.
Mastra now lets agents route requests to self-hosted open-weight models through a custom model gateway, configured with a custom ID and referenced as a model prefix such as private/ollama/gpt-oss:20b. The post walks through creating a gateway for a local Ollama server, registering it on the Mastra instance, and setting AUTO_BLOCK_EXTERNAL_PROVIDERS=true to hide external providers in Studio's model picker. It requires @mastra/[email protected] or later and cites privacy/compliance and fixed infrastructure costs as motivations.
Amp is retiring its Dial and its four modes (low, medium, high, ultra) in favor of a model selector that lets users pick the model and its effort per thread, including mid-thread, and shows which subscription or API key will be billed or whether Amp credits are used. The change follows the addition of bring-your-own subscriptions and keys, which made the selected model determine billing as well as token production. New threads start on Staff Pick (Claude Opus 5.5 today), and existing threads keep running on their prior model.
H6.ai automates accounts payable on top of customers' existing ERPs by combining AI-assisted invoice understanding with deterministic, auditable policy execution. To make regression testing reliable despite constantly changing financial data, it gives each customer a dedicated Lakebase Postgres database and uses copy-on-write branches to freeze snapshots, run each test on its own disposable branch, and judge results with deterministic assertions plus a panel of AI judges. The company says branching provides isolation, frozen state, cheap disposable copies and schema-safe replays, and it plans to use Lakehouse//RT queries on Lakebase tables for read-only analytics without copying data.
AWS has released AWS Emulator, a free, CC0-licensed local emulator of AWS APIs that runs on a laptop, dev server or in CI without an AWS account, built from 438 AWS API models covering 200+ services. Its code was written and reviewed entirely by autonomous agents, and release v1.0.322 passes 24,243 of 42,857 scored fidelity tests (56.6%), with KMS at 93.9% and CloudFormation at 15.8%. AWS lists limitations including no IAM evaluation, a single region and account, in-memory short-lived resources, and telemetry on by default.
Near AI Gateway is a macOS app that routes AI requests from Claude Code, Codex and OpenCode through NEAR AI private models once you sign in, connecting one tool at a time. It reads only folders you choose, redacts identifying information on your Mac, and lets you review every trace before deciding whether to share it, with accepted traces earning NEAR AI Cloud inference credits. It requires macOS 14 or later and is version 1.0.7 at 119 MB.
Anthropic announced a private beta of self-hosted Claude Managed Agents, which runs the full service—harness, sandboxes, credentials, and session history—in a customer's own cloud or on-premise, with model requests routed to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or a customer gateway. The design separates the harness from the sandbox, keeps credentials out of the sandbox via vaults or a customer secrets manager, allows per-environment host allowlists, and exports session records over OpenTelemetry. Access is available by request through an Anthropic account team.
Factory introduced the Cyber Defense Network, an ecosystem of cybersecurity companies whose findings connect into Security Factory, with founding partners including Snyk, Palo Alto Networks, Semgrep, Cogent, Casco, DepthFirst, and Detail. Partner findings flow into Security Factory, where Factory Droids investigate vulnerabilities, implement code changes, and validate fixes through testing across signals, triage, and continuous remediation stages. Factory plans to expand the network to additional cybersecurity innovators and model providers.
Brian Woods argues that AI agents make producing a working first change cheap, but accepting subsequent changes requires preserving what must stay true—such as no duplicate payments or cross-tenant data leaks—through specification, validation, provenance and authorization. He describes a synthetic experiment against Uncle's verifier where a signed record of a build with failed tests and audit returned VERIFIED, showing that attestation validity and release readiness are separate questions. He proposes a design in which each accepted change is attested by a trusted workflow, linked to prior attestations in a DAG, and verified by walking the graph, while noting the approach cannot prove tests were sufficient or the software correct.
Block developed Buzz, an open source, model-agnostic collaboration space where employees and AI agents work together in group discussions, focused conversations, and one-on-one exchanges, with users choosing the model or coding harness behind each agent. Databricks Unity Gateway provides the model access layer, letting Buzz users select Databricks as their inference provider and choose from frontier models from OpenAI, Anthropic, and Google plus select open-weight models through a single integration. Block has also been experimenting with multi-agent workflows, such as a more capable model coordinating agents powered by smaller, more efficient models, and developed an internal benchmarking approach for these agent teams.
Helion now uses compile-time heuristics that analyze a kernel's IR and dataflow to classify it as matmul, reduction, or pointwise and emit block sizes, warps, and stages without autotuning. On H100, the computed configs reach 1.18x hand-written FLA Triton for linear attention, 1.27x vLLM's handwritten CUDA reduction kernels, and 1.09x torch.compile max-autotune on a general reduction suite. The heuristics are merged into Helion and enabled by default on H100 and B200 with the Triton backend.
AWS authors propose the Agent Development Lifecycle (Agent DLC), which ties six activities—Define, Build, Evaluate, Release, Observe, and Feed Back—to an explicit "gold standard" of criteria, thresholds, and approved examples, with an inner loop (score, inspect the path, change one thing, re-evaluate) and an outer loop that updates the standard from production feedback. A reference HR policy assistant shows chunk-level retrieval precision of 1.00 on all three test questions while fact-level precision was 0.59, 0.53, and 0.09, and a prompt-only change improved one question but worsened another, which the authors attribute to missing source coverage. The post also describes a benefits-enrollment answer that scored 0.90 groundedness yet made a false claim about the user's contact preference, and outlines a one-week AWS getting-started plan using Amazon Bedrock AgentCore Evaluations and Observability.