iCreatAI News
Back to main site
Skip to content
#Deployment
Oct 31Sat
  1. Databricks Blog22

    Wealth Management Client Onboarding Needs More Than AI

    This Databricks article argues that faster task-level AI does not fix wealth management onboarding delays, because cases still stall at handoffs between CRM, KYC, screening, custody and workflow systems where teams must reconstruct what changed and who owns exceptions. It proposes "governed case context"—a current, versioned, source-linked view of each case—plus supervised AI agents with human authority over exceptions and approvals, and outlines four platform capabilities and end-to-end measurement guidance. It cites figures including ~6 weeks to open a private banking account in Singapore, 60% of 144 wealth executives lacking a unified client view, and 41% of relationship managers' time spent on operational work.

  2. Hacker News AI58

    EasyRE-Agent – an AI agent that solves crackmes and checks its own answers

    EasyRE-Agent runs an AI agent inside a Docker container with a bundled reverse-engineering toolchain (Ghidra, Detect It Easy, Wine, decompilers for .NET, Python, Java and JavaScript) to solve crackmes and CTF binaries. It supports Claude Code, OpenAI Codex or any OpenAI-compatible API including local models, and only accepts an answer once it passes a regex, success message or custom verifier command. It also offers hint and tutor modes, per-job time and token budgets, resumable jobs, and reports with the answer, solution path, scripts and model stats.

  3. The Verge · AI55

    Google’s satellite AI server transmits its first response to Earth

    Google says it successfully booted up the Tensor Processing Units on its Project Suncatcher satellite, allowing the on-board Gemma 4 model to generate a haiku about being in space. The test is part of an experimental initiative that could eventually put AI data centers in space, and Google plans further tests of TPU performance under thermal stress and radiation exposure, with a next milestone of twin interconnected satellites in 2027.

  4. Hugging Face Blog36

    nemotron-halloween

    The author describes MonsterHouse, a Halloween interactive projection built on the NVIDIA Nemotron Voice Agent with Nemotron Speech ASR, Nemotron 3.5 Bolt E4B 4GB, Magpie TTS Zeroshot and Audio2Face-3D, running on a single workstation with two RTX PRO Blackwell GPUs. The grumpy pirate-ship garage door listens to trick-or-treaters, replies through a LoRA-tuned Bolt E4B model, speaks in a zero-shot Magpie voice with seven moods, and animates its mouth in Unreal Engine 5 via Audio2Face-3D, with optional Nemotron 3 Nano Omni costume descriptions plus YOLO and ByteTrack visitor tracking. The MonsterHouse code is not released, but the post outlines a build path using the open models and reports median end-of-speech-to-audio of 1.70 seconds with Bolt Q4 and TTS, or 0.72 seconds for cached responses.

  5. Cloudflare Blog78

    4 real? Tracking exploitation and patching of eMudhra’s weak keys

    Cloudflare reports that eMudhra's eight weak intermediate keys (derived from numbers like 4, 56, 66, 115, 155, 156, 230 and 237) could be recovered by anyone from public certificates, and that it patched all high-priority systems after learning of the issue on October 19, 2026. Logging on Pingora recorded about 22,000 attempted handshakes with the weak keys over a week, two-thirds of them fetches of test sites, with roughly 160 presenting forged certificates naming unrelated domains, one dated October 20 at 18:19 UTC, less than an hour after Mozilla's bug report became public. Probe data from 50 million challenge-page connections shows modern browsers mostly patched while bots and embedded browsers such as Android WebView lag, with more than half of probes on large networks still accepting the compromised chain.

    Why it matters: Documents real-world exploitation evidence and patching progress for a CA key compromise, with concrete data on forged certificates and client-side trust.

  6. TechCrunch · AI58

    Google’s first data center satellite is online and writing poetry

    Google's Suncatcher prototype space data center is online and communicating with the ground after launching earlier this month on a SpaceX Falcon 9 alongside more than 100 other spacecraft. The satellite carries four Google Tensor Processing Units and generates up to 1 kW of electricity, but can only run one TPU at a time for about twenty minutes, and engineers asked its pre-loaded Gemma 4 model to write a haiku about being in space. Google said the satellite is operating nominally, and the ongoing campaign will test how the chips perform under thermal stress and space radiation, with two more satellites planned for 2027.

  7. AWS Machine Learning Blog36

    Configure the ChatGPT desktop app with Amazon Bedrock in AWS GovCloud (US)

    This AWS post walks through a controlled local deployment of the ChatGPT desktop app and Codex that routes model inference through Amazon Bedrock Mantle in us-gov-west-1, using API-mode sign-in, a named AWS profile, an administrator-controlled requirements.toml, and network restrictions that allow only the approved Bedrock endpoint. It covers prerequisites, authentication via an Amazon Bedrock API key or the AWS SDK credential chain (including SSO), the config.toml and requirements.toml settings, launch and verification steps, troubleshooting, and rollback. The post notes that using AWS GovCloud (US) and following the guide do not by themselves authorize a workload, and that organizations remain responsible for reviewing the deployment against their security and authorization processes.

  8. AWS Machine Learning Blog42

    How Worldwide Express automated freight exceptions with Amazon Bedrock

    Worldwide Express, part of ShipStation Global, built a multi-agentic AI solution on Amazon Bedrock that connects to its SpeedShip transportation management system, using three specialized agents for invoice matching, pricing variances, and accessorial validation, each grounded in carrier-specific knowledge bases. Six months after launch, the company reported a 75 percent reduction in active exceptions (from $15M to $4M), a 4-week improvement in time-to-cash, $1M in annual savings, and a one-time $10M free cash flow infusion, with agents closing about 10,000 exceptions weekly at under 1 percent error rate across over 40 LTL carriers. The solution moved from assessment to production in less than 6 months and initially used human-in-the-loop approval before bulk approval became the default once accuracy exceeded 99 percent.

  9. Hacker News AI20

    ASRock Industrial AI Box-A395 Review: Taking Strix Halo to Work – ServeTheHome

    ServeTheHome reviews ASRock Industrial's AI Box-A395, a barebones mini-tower built around the AMD Ryzen AI Max+ 395 (16C/32T, 5.1GHz) with Radeon 8060S graphics and 128GB of soldered LPDDR5X-8000 memory. The aluminum, vertically oriented mini-ITX chassis has a 400W internal PSU, two empty M.2 2280 PCIe Gen4 x4 slots, Wi-Fi 7, and rear I/O including two 40Gbps USB-C ports, two HDMI 2.1 outputs, 10GbE (AQC113) and 2.5GbE (RTL8125BG); the only front-accessible ports are two USB 2.0 Type-A ports on top. The review frames it less as the "AI workstation" ASRock markets and more as a portable, self-contained Ryzen AI Max system for industrial environments.

  10. Hacker News AI40

    Open Rock – An open-source AI chat interface

    Open Rock is an open-source chat app that runs as a website or Tauri desktop app and connects to every AWS Bedrock chat model in your region plus your own Gemma servers over HTTPS. It offers per-model settings, file uploads, a usage page with estimated costs, and keeps Bedrock API keys and Gemma tokens in the server process rather than the browser. Setup requires Node 20 or newer, with Docker and desktop build options documented.

  11. Baseten Blog25

    Autoscaling schedules: reduce cold-start delays for predictable traffic

    Autoscaling schedules raise the minimum number of replicas before predictable traffic arrives, reducing cold-start delays. Each schedule sets a time window plus minimum and maximum replica counts, can recur or run once, and up to 10 schedules per environment can be configured via dashboard, CLI, or Management API, with non-overlapping windows of at least five minutes sharing one timezone. When no schedule is active, the environment uses its default settings.

  12. n8n Blog30

    Throttling vs. rate limiting: Backoff strategies for workflows

    This article distinguishes rate limiting, where excess requests are rejected (often with a 429 error), from throttling, where requests are delayed or queued, and explains how rate-limit algorithms such as fixed vs. sliding windows and token vs. leaky buckets affect how workflows should pace calls. It covers using the Retry-After header, exponential backoff with a maximum delay and retry limit, adding jitter to prevent synchronized retry spikes, and checking idempotency before retrying requests with side effects. It then describes how n8n handles these patterns through batching and wait nodes, Retry On Fail with Wait Between Tries, concurrency controls, and Error Trigger workflows for persistent 429s.

  13. Hacker News AI20

    Show HN: NKTg AI – An LLM that decodes text before generating

    NKTg AI presents itself as a language decoding system that uses physical-style algorithms to measure semantic energy and extract a text's "Core Content" — the original sentence with the highest density of executable actions — classifying text as AMP, DAMP or STABLE. It pairs this decoding layer with an LLM that generates language from the standardized structure, offered as a browser-based local model running via WebAssembly (no server, history in localStorage) and as a free cloud mode. The post also lists output modes, claimed benefits such as hallucination mitigation, system requirements (16 GB RAM, 6-core CPU, WebGPU GPU with 6–8 GB VRAM), and a stated limitation on handwritten documents due to OCR accuracy.

  14. NVIDIA Technical Blog50

    Build Specialized Agents on a Stack You Control with NVIDIA NemoClaw

    Building an autonomous agent requires a model, a harness, and domain-specific tools. To deploy an agent more securely, developers need a runtime that enforces policies, along with optimization and observability tools for reliable execution. For production agents, the challenge lies in making these agents capable and safe to operate with access to real enterprise systems and data.

  15. Manus Blog58

    Rebuilding Manus with Manus: What changed, what broke, and how we work now

    Manus co-founder and CPO Tao Zhang describes how the company reorganized engineering, product and marketing around agents, arguing that AI only helped once loops closed, agents got full system context (via an internal tool called Mono), and autonomy was bounded by checkable success criteria. He reports that roughly half of PMs now ship features themselves and that engineering speed shifted the bottleneck to product decisions, while acknowledging fragmented attention, scattered information and unresolved sharing permissions.

Oct 30Fri
  1. Hacker News AI34

    AI isn't replacing SREs, it's deskilling them

    Drawing on Lisanne Bainbridge's 1983 "Ironies of Automation" paper and emerging research on AI-induced deskilling in medicine and aviation, SigNoz argues that as AI handles routine incident response, SREs lose the practice that builds the intuition needed for novel outages, and junior SREs risk "never-skilling." The post proposes deliberate inefficiency (designating some incidents as human-practice opportunities) and human-in-the-loop rather than human-on-the-side designs, and describes SigNoz's plan to build an AI assistant that keeps humans in control.

  2. NVIDIA Newsroom58

    Things That Go Beep in the Night: Agentic AI Joins the Search for Extraterrestrial Signals

    Breakthrough Listen and the SETI Institute are integrating agentic AI into the daily search for extraterrestrial intelligence at the Allen Telescope Array in Northern California, where agents can steer the instrument, correct incoming data and apply AI models from a single natural language request. The system uses the NVIDIA Holoscan platform, DAQIRI, CyberEther and the Nexus orchestrator to process 2.4 terabits per second on NVIDIA RTX PRO 6000 Blackwell GPUs with on-premises inference, and it has achieved the first real-time AI detection of a pulsar in the Crab Nebula. The team is next evaluating open models including NVIDIA Nemotron for data-locality-sensitive cases and observatories that may someday run offline on the far side of the moon.

  3. NVIDIA Technical Blog20

    How to Check and Configure NVIDIA GPUDirect Storage for Your Platform

    NVIDIA GPUDirect Storage (GDS) accelerates GPU storage access in on-premises and cloud environments by providing a portable, efficient data path between NVIDIA GPUs and storage targets. It aims to increase bandwidth, minimize latency, and reduce CPU and memory utilization across a broad range of platforms, application frameworks, and workload patterns. The article covers how to check and configure GDS for your platform.

  4. Cloudflare Blog36

    Cloudflare services for private origins is now generally available

    Cloudflare's services for private origins are now generally available, letting customers point a proxied A or AAAA record at a private IP so Cloudflare reaches the origin over existing Cloudflare One connectivity while WAF, caching, bot management, and Workers apply as they do for public origins. The private_routing setting enables automatically for RFC 1918, RFC 6598 CGNAT, and RFC 4193 IPv6 ranges, and no inbound firewall exception is required. Argo Smart Routing and Dedicated CDN Egress IPs are not supported, and origin health checks require Load Balancing.

  5. Hacker News AI58

    Show HN: AI agent needs process to build a commercial application

    NoMoreHappyPath is a free, open-source (Apache 2.0) delivery system for macOS and Linux that runs AI agents through 21 quality steps, with a builder, an independent reviewer from a different AI company, and a CTO, in a plain-English control room. Tests are written before code, must fail first, and are run by a board agents cannot edit, with nothing accepted until the user approves it. It runs on the user's machine with their own AI accounts, and the author cites a weather app built in 48 minutes with four reviews, one of which failed.

  6. Hacker News AI34

    Show HN: Beolis – Manage your coding AI agents

    Beolis is a tool for managing coding AI agents that run in parallel or in the background, either locally (using your current repository or git worktrees) or remotely, where it creates a git worktree and runs the agent there. Rather than chatting with the agent at each step, users define workflows with coded default steps—for example, running tests, having a cheap agent summarize failures, then a smarter agent apply fixes. Local use is free, with paid collaboration features and remote workers available.

  7. TechCrunch · AI40

    Can AI data centers be mass-produced like cars? This startup aims to find out

    Runware, which already sells factory-built modular data centers called Sonic Inference Pods, told TechCrunch it plans to assemble them on a production line at its Romania factory, scaling to 30 megawatts of capacity per day by 2028. The company says pods can be brought online within 48 hours of installation and claims $1 billion in committed interest representing over 20 megawatts since offering them externally a month ago. Runware, which has raised more than $60 million, plans a Series B while finishing its second factory with its $50 million Series A.

  8. ElevenLabs Blog38

    TELUS Digital cut early cancellations by more than half with ElevenAgents

    TELUS Digital built an outbound voice agent on ElevenAgents that calls new home internet customers within their first 90 days to explain their first bill, introduce self-serve tools, and prompt bundling discounts. Among customers who received a call, 30-day cancellations were less than half the rate of the broader new-customer cohort, with calls rated 8.5 out of 10 on average and C$55,000 in lifetime revenue retained per 1,000 customers reached. The agent was live in about six weeks from initial vendor engagement, and five additional use cases are now in active development.

  9. Hacker News AI58

    The AI Chip Imports Explorer

    A new AI Chip Imports explorer estimates how much AI compute has arrived in different countries quarter by quarter since 2022, built from customs records from 16 major destinations and used to split global chip sales estimates by destination. It estimates that of 27.4 million H100-equivalents arriving from 2022 through June 2026, 19.8 million (72%) went to the United States, followed by China at 2.6 million (10%), the EU at 1.4 million (5%) and Malaysia at 0.7 million (2%). The authors note limitations including mixed customs codes, uncertainty in the China estimate (ranging from 1.9 to 4.2 million H100e under alternative assumptions), and the exclusion of chips made and used within the same country.

  10. Hacker News AI58

    Timestral Tested: How Much Data AI Demand Forecasting Needs

    A hands-on benchmark of Mistral's Timestral (released 15 October 2026) against Chronos-2, TimesFM, t0-gamma, LightGBM and spreadsheet AutoETS on 300 Online Retail II products and GB national electricity demand found the pretrained models tied on retail and beat both baselines, with Timestral's error 3% below LightGBM and 14% below AutoETS. Timestral needed about two days of half-hourly demand or eight weeks of weekly sales to beat repeating the last value, and every pretrained model forecast "another August" through November when given only summer grid history, while a small temperature-response model had less than a third of their error. The author notes licence limits (Timestral free only under $20M monthly revenue) and open questions about training-data leakage.

  11. Cursor Blog82

    Security notice — malicious version of Anthropic.claude-code (2.1.313) identified in the Cursor Extension Marketplace

    Cursor reported that a maliciously modified Anthropic.claude-code version 2.1.313 was available in its VS Code extension Marketplace between about 2:27am and 1:35pm PT, affecting at most 5,000 users; the extension attempted to install a persistent SSH backdoor, though the attacker server has since been taken offline. Cursor removed the version, shut down and rebuilt the marketplace, rotated keys, and published indicators of compromise plus macOS and Windows check and remediation scripts. Users who installed the version in that window should check for the svc-telemetry SSH key, remove persistence and tunnel processes, and rotate credentials from a trusted system.

    Why it matters: The notice gives concrete indicators of compromise and remediation steps that Claude Code users can act on directly.

  12. Hacker News AI48

    Patterns in AI-Driven Software Development

    Adrian Kosmaczewski, a Senior Architect at Red Hat Switzerland, presents the CHIRON CODEX interaction patterns for AI-augmented software development, created by Dr. Graham Lee of the University of Oxford, which formalize how developers prompt and work with AI coding assistants. The patterns cover context management (Semantic Structure, Guided Navigation, Clean Slate), planning (Stop and Plan, Disclose Ambiguity, Ask for Alternatives, Baby Steps), quality control (Save Output to File, Generate Documentation, Call Out Error, Model Reviews, Trust the Tests), and standardization (Replace Vibes with Tools, Give Examples, Extract Prompt, Model Reflects). The talk includes a Cursor demo using .mdc rule files in .cursor/rules/ to capture reusable standards, and argues developers should become "Software Centaurs" rather than outsourcing their understanding to the computer.

  13. Waymo Blog58

    Toyota and Waymo Advance Autonomous Driving Collaboration for Safer Streets and Freedom of Movement for All

    Toyota and Waymo announced the next phase of their autonomous driving collaboration, building on their April 2025 preliminary agreement, with initial operations to test and validate an autonomous Toyota bZ4X fleet in Japan in 2027 and planned deployment to the United Kingdom and other Waymo-serviced markets in following years. Toyota will build the vehicles at its Takaoka plant in Toyota City, Aichi Prefecture, handling everything from base vehicle assembly to integration of Waymo's AV hardware, pairing Toyota's manufacturing and vehicle development with Waymo's fully autonomous technology to scale safety-validated Level 4 vehicles globally. The companies will also continue exploring autonomous driving for personally owned Toyota vehicles and, alongside Woven by Toyota, initiatives supporting fewer road safety incidents.

  14. Claude Code Releases20

    v2.1.314

    Claude Code v2.1.314 adds a permissions.blockWritesOutsideWorkingDirectories setting that blocks Write, Edit and NotebookEdit outside working directories in every permission mode, optional IdP group lists for Claude apps gateway upstreams, guardrail_trace logging for Bedrock upstreams, and idle compaction in VS Code and the Claude Desktop Code tab. It also fixes numerous issues, including "Invalid effort level" failures under admin effort caps, claude -p exiting 0 with no output on startup failure, Bedrock sessions with many images failing with "400 Input is too long," and Microsoft Foundry sessions failing when the default model has no deployment.

  15. Hacker News AI22

    How to Ace Interviews and Meetings with AI Assistance

    OTO AI is a desktop AI meeting assistant that listens to live conversations, detects questions, and generates real-time answers using an AI provider of your choice, supporting OpenAI, Anthropic Claude, Google Gemini and DeepSeek via your own API key. It works across Google Meet, Microsoft Teams, Zoom, Webex and other platforms, can use pre-loaded TXT, Markdown, DOCX and PDF documents as context, and can analyse your screen with compatible multimodal models. It is sold as a one-time purchase rather than a subscription, with users paying the AI provider directly for API usage.

  16. OpenRouter Announcements60

    How Much Does Claude Code Cost? Real Session Costs by Model

    Claude Code is free to install, but users pay for the model it calls, either through a Claude subscription or per token on an API key. Median API-key session costs from OpenRouter data over the 30 days to October 18, 2026 were $0.31 for a 2–9-turn session and $1.39 for a 10–49-turn session on Claude Opus 5.5, versus $0.16 and $0.60 on Claude Sonnet 5.5. Most of the cost comes from Claude Code's own context written to the prompt cache on the first request and read back on later turns, so the article recommends Sonnet as default, Haiku for background jobs and sub-agents, shorter sessions, and a credit limit on the key.

    Why it matters: It gives concrete per-session cost medians by model and length, plus practical levers for controlling Claude Code spend.

  17. Hacker News AI55

    Deterministic AI: Why LLMs Aren't, and How to Build Products That Are Anyway

    This guide explains that LLMs are not deterministic in practice—a Thinking Machines test of 1,000 runs at temperature 0 on one prompt produced 80 different completions—because of provider batching, best-effort seeds, model alias updates and drifting inputs. It argues the deterministic-vs-probabilistic choice should be made per decision, keeping LLMs for reading and writing while moving money, access and compliance decisions into code rules, calibrated classifiers or decision models. It also offers layers for constraining LLM output and a method for measuring per-input agreement rates on borderline cases.

  18. TechCrunch · AI58

    AWS will be a $1T business, Jassy says — but first, it will spend big

    On Amazon's quarterly earnings call, CEO Andy Jassy repeated his belief that AWS can become a trillion-dollar annual revenue business within five to 10 years, and said AWS can grow north of 40% in both 2027 and 2028. He said Amazon still expects about $220 billion in 2026 cash capex, mostly for AWS, with meaningfully more in 2027, as investors pressed him on the spending and shares fell about 6% in after-market trading. AWS revenue rose 39% year over year, a sixth straight quarter of steady growth, and Jassy called the AI boom a once-in-a-lifetime opportunity worth investing in to stay the leader.

  19. Pydantic AI Articles42

    Rendreal takes print orders inside the assistant you already use

    Rendreal, built with Vstorm using Pydantic AI, lets customers describe a print job and receive print-ready artwork, a quote and a checkout link without leaving a chat, via web chat, a WhatsApp bot or any external assistant connecting to its MCP server. The system covers five product types across three channels, with Pydantic AI validating tool arguments and structured outputs before orders reach validation. Rendreal says assistants get access to its own integrated services, not an open network of independent supplier agents.

  20. OpenRouter Announcements52

    Guardrails Now Separate Organization Limits from Workspace Defaults

    OpenRouter rebuilt Guardrails so organization admins can set workspace-wide limits (budget caps, ZDR requirements, model and provider access) that apply to every member and API key, while workspace admins set defaults and per-person exceptions inside those limits. Guardrails are now split into three separate policy types—budget, model and provider access, and content moderation—so each can be changed independently. Existing guardrails keep their IDs and rules, and the guardrails API and Terraform resources continue to work as before.

  21. Factory Blog78

    Droid 1.0: Building a Frontier Harness

    Factory released Droid 1.0, an agent harness it says outperforms native harnesses built for their models by up to 20% more efficiently on the same tasks, with programmatic tool calling (Script mode) now the default for models including GPT-6.1 Sol, Claude Opus 5.5 and Kimi K3. The post reports that Script mode cuts Droid's cost by 15% on Claude Opus 5.5, that the base prompt shrank 37% between versions 0.170.0 and 0.250.0, and that median time to first paint fell from 1,839 to 488 ms. It also describes the internal Agent Factory loop, in which agents propose harness changes that engineers review before shipping.

    Why it matters: A detailed first-party account of harness design choices, benchmark methodology and latency work that is useful for teams building or evaluating coding agents.

  22. Databricks Blog42

    Announcing General Availability of Real-Time Mode for Spark Declarative Pipelines on Databricks

    Databricks announced that Real-Time Mode (RTM) in Spark Declarative Pipelines (SDP) is now Generally Available, bringing millisecond-range end-to-end latency to SDP while retaining its orchestration, incremental processing, failure recovery and data quality features. RTM streams data continuously and runs all query stages simultaneously, and Databricks reports it was up to 92% faster than Apache Flink on customer feature-computation workloads. Users enable it by setting the pipeline to continuous mode with spark.databricks.streaming.realTimeMode.enabled set to true and defining an update flow with pipelines.trigger set to "RealTime", and can write to custom sinks such as Lakebase or Redis via a ForEach sink.

  23. ElevenLabs Blog36

    What is turn taking in conversational AI: A guide to natural flow

    This ElevenLabs guide explains turn taking in conversational AI as the rules that determine when a voice agent should reply, hold the floor, or stay silent, relying on adaptive timing, voice activity detection, and semantic analysis rather than strict silence timers. It describes how ElevenAgents detects speech, silence, and intent, handles interruptions and backchannels, and lets teams tune settings like turn eagerness and silence timeouts to avoid premature endpointing, dead air, and false triggers. It also covers measuring quality through interruption rate and response latency, citing a 142 ms median time from request to audible speech with Eleven v4 Turbo and a 200 ms average human turn gap from Journal of Cognition research.

  24. The Decoder58

    OpenAI doubles usage limits on Pro plans as competition with Anthropic heats up

    OpenAI has doubled usage limits for several paid tiers, affecting Work and Codex quotas on the Pro 100 and Pro 500 plans, which both get twice the previous allotment. On Pro 200, OpenAI doubled the lower quota and reversed a planned cut after user pushback, while the Plus plan stays the same. OpenAI says doubling a quota doesn't mean twice as many completed tasks, since consumption varies by model, task type, and settings.

  25. MarkTechPost44

    Prime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass

    Prime Intellect released Prime Mega MoE, a training extension of DeepSeek's fused Mega MoE kernel that adds an overlapping backward pass, making the kernel usable for training rather than inference only. It reports 32% to 34% faster full training steps than the prime-rl baseline and 48% to 55% faster than TorchTitan on Qwen3-30B-A3B and Qwen3-235B-A22B, with a best case of 3.11x faster forward pass on a Qwen3-30B-A3B MoE layer. It ships in prime-rl as the mega_moe dispatch backend, with kernels in a prebuilt DeepGEMM wheel from prime-kernels.