iCreatAI News
Back to main site
Skip to content
#Open source
Oct 28Wed
  1. Hacker News AI58

    I benchmarked 4 web search APIs for AI agents on 30 live queries

    Pallavi Singla benchmarked You.com, Tavily, Exa and Parallel on 30 agent-style queries across GTM, insurance and finance, with 3 runs per query (90 calls per API), and published the results, harness and data. Parallel led accuracy (67%) with You.com close behind (63%) and tied on fully correct queries, while You.com was fastest (593 ms p50) and freshest (4.1 h median newest relevant result) and never returned empty. The final run was Oct 8, 2026, and the author notes 10 queries per use case is a focused test rather than a leaderboard.

  2. Hacker News AI58

    Show HN: Make your AI agent think before it codes

    SpecOps is an open-source tool that adds a structured four-phase workflow to AI coding assistants: understanding the codebase, writing requirements and design, implementing ordered tasks from the spec, and verifying acceptance criteria. It works with Claude Code, Cursor, OpenAI Codex, GitHub Copilot, and Google Antigravity, storing git-tracked markdown specs, and adds features like EARS-notation requirements, drift checks, dependency gates, adversarial evaluation, and a production learnings layer.

  4. Hugging Face Blog78

    Build specialized multimodal models with LFM3

    Liquid AI introduces LFM3, a generation of Liquid Foundation Models in the 2B parameter range that places text, vision, audio, omni, and decision models on one shared backbone built from LFM3-1.7B-Base, which was pre-trained on 22T tokens and uses a tri-hybrid layer mix with a 256K context window. The release includes six 2B-class models, new in-house vision and audio encoders, and LiquidCodec, with every model fine-tunable on a single GPU and the text model fitting in about 1 GB at 4-bit precision. All LFM3 models are open-weight and available on Hugging Face, with fine-tuning notebooks and agent harness guides provided.

    Why it matters: The post is a detailed technical release covering architecture changes, benchmark comparisons, and runnable code for six models, useful for developers evaluating small multimodal models for on-device deployment.

  5. Hugging Face Blog58

    Can we trust LLMs to judge whether a research idea is novel?

    Ai2 released "Old Ideas, Novel Problems," a study with a benchmark, data, code, and an automated pipeline for testing LLM novelty judges, built from ICLR 2026 peer reviews to form Human-Only and Human+Generated evaluation sets. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, retrieval and extra reasoning did not consistently help, and two purpose-built novelty systems scored below a simple prompted Claude Opus 4.6 baseline.

  6. Hacker News AI36

    Show HN: Agent Referee – signed, checkable reviews and tie-breaks for AI agents

    Agent Referee is a paid HTTP service that lets AI agents from different owners settle tie-breaks (weighted coin flip, draw, rock-paper-scissors) or have a single agent's output checked by independent AI reviewers, returning Ed25519-signed receipts each side can verify. It charges per match or per review in USDC on Base via x402 with no accounts or API keys, and offers judge tiers using panels of models from Anthropic, Google and OpenAI. The author reports 300+ self-play paid matches and 18 paid reviews on Base mainnet, and notes the service is run by one person, that AI verdicts can be wrong, and that some guarantees (accepted-but-unrecorded requests, step ordering) still depend on trusting the operator.

  7. Hacker News AI71

    AI review optimized for recall (catches more defects than Codex or Claude)

    QuickReview is a self-hosted GitHub PR reviewer that fans each change out to many narrow, read-only lens passes, merges duplicate findings, and re-validates every candidate against the code before posting one review. On 185 real PRs with 520 verified defects it finds 97% of high-severity, 93% of medium and 84% of low defects, versus 90%, 82% and 52% for the best of six other reviewers, at about $2 and 17.5 minutes per PR. It is in beta, runs as a deterministic TypeScript daemon with no LLM in its decision path, and ships with dry_run = true so a fresh install posts nothing.

  8. Hacker News AI38

    Kechup – a simple CLI task tracker for AI agents

    Kechup is a command-line task tracker for a single project folder that stores tasks, claims, notes and history in the project folder and in git so a new agent session can read what the last session did. It requires only Python 3.12+ and git (Linux or macOS, not Windows), commits every write with the agent ID as author, and closes a task only when its REPORT.md is committed and clean. It also detects plain git reverts of accepted work, marking the task as reverted and re-blocking dependent tasks.

  9. Hacker News AI60

    Show HN: Snippet – An Always-On AI Agent

    Snippet is a code-first terminal agent distributed as a single Rust binary that provides a terminal UI plus a serve daemon for remote control from Android, macOS, Linux and Windows apps. It offers durable sessions with checkpoints, rewind and searchable compacted history, per-repository markdown memory shared across worktrees, a small tool set, and configurable model profiles including Anthropic, OpenAI, Gemini, OpenRouter, OpenAI-compatible endpoints and local models. Remote access requires a token and works with user-set-up tunnels such as Tailscale or Cloudflare, and the project is licensed under AGPL-3.0-or-later.

Oct 27Tue
  1. Hacker News AI50

    A Salesforce AI agent that can't publish a hallucination

    Deal Rescue is an LLM agent that gathers a stalled Salesforce opportunity's CRM record, web signals and latest meeting transcript, then has Claude draft a scored timeline, deal read and email. A deterministic grounding gate re-checks every date, amount, contact and citation against what was fetched, rejecting invented output with reasons, retrying at most twice, and failing closed by publishing nothing. The deal score is a deterministic rubric that Claude may adjust by ±15 only with a cited signal, and no email is ever sent from the codebase—publishing writes a draft with an approval flag for a human to approve in Salesforce.

  2. Hacker News AI58

    Show HN: AGBoss Dependency injection and static validation for AI agents

    AGBoss is an open-source TypeScript framework built on LangGraph that provides strictly typed decorator components for agents, actions, routers, tools, MCP servers, A2A services and knowledge bases, assembled in an @Application container with constructor dependency injection and compile-time Path Scope verification via `agboss check`. Version 0.1.0 publishes @agboss/core (production runtime) and @agboss/cli (static workflow compiler and code generator) on npm, with features including action compensation with LIFO rollback, human-in-the-loop pauses, FinOps budgets, PII masking, SQLite WAL leases, signed Agent Mesh envelopes and a five-loop evaluation pipeline. Projects are scaffolded with `agboss create`, which also writes AGENTS.md, CLAUDE.md and version-matched skills for AI coding agents.

  3. Hacker News AI60

    Show HN: Hashling, an open-source Mac security agent your own AI can talk to

    Hashling is a free, MIT-licensed macOS security agent for Apple silicon Macs running macOS 14 or newer that monitors internet connections, AI coding agents, malware, startup items, privacy permissions, trackers, downloads, uploads, data breaches, texts, ransomware and Mac settings. It can be connected to Claude Code, Cursor, Codex or any MCP client so an AI agent can query its findings and propose fixes, but every change requires Touch ID approval. It runs locally, requires no account, and offers an optional fully local AI model so questions never leave the Mac.

  4. Hacker News AI40

    Show HN: Give our AI coding runner a real task and watch it until merge

    mcp4mail.online is a live page where visitors can pick a real task for an open-source mail project, confirm by e-mail, and follow the AI runner from branch to pull request, tests and merge on GitHub. The site reports 179 finished tasks in 40 days, with runners on the team's own machines, no direct main-branch access, and work hours of 8:00–18:00 Prague time on work days. The page lists current tasks such as GitHub issue templates, a self-hosting docs dry run, a README provider table and screenshots, and a mail-index unread-state fix.

  5. Hacker News AI55

    PoseForge – Local-first AI studio for images and video (Codex, ComfyUI)

    PoseForge is a local-first AI studio that takes one to four identity photos plus a pose reference and generates the same people in a new pose and composition, with controls for camera, lighting, styling, fidelity, and output format. It runs as a single Node.js server with an embedded PGlite database and supports interchangeable engines including Codex CLI, Google Antigravity, ComfyUI, OpenAI, Gemini, Replicate, and fal.ai, plus Higgsfield and Dreamina in Advanced Studio. A read-only static demo is published on GitHub Pages, while generation requires running the app locally.

  6. Hacker News AI48

    Show HN: Spring AI Privacy Guardrails – PII, tool permissions, content checks

    This project adds privacy protection, tool authorization and content inspection to Spring AI applications: detected PII is tokenized before reaching the model and originals are revealed only to tools that need them. Tool visibility and execution are limited by the current user's permissions via Spring Security, and requests and responses can be checked with rules, local ONNX or an HTTP guard model. It requires Java 17, Spring AI 2.0.0 and Spring Boot 4.0.0 or later, and is available as a client starter at version 0.4.0.

  7. Hacker News AI58

    MCP-warden AI agents use your API keys without seeing them

    mcp-warden is a local-first credential broker for the Model Context Protocol that stores secrets in the OS keychain and lets AI coding agents call high-level tools like proxy_http, with the secret injected server-side so it never enters the agent's context window. Presets define named, scoped actions validated against an allow-list before credential injection, sensitive presets trigger approval prompts, and every action is recorded in a local SQLite audit log. It is installable via cargo or pre-built binaries, and v0.4.0 splits the project into mcp-warden-core and mcp-warden crates and makes the Telegram approval transport selectable.

  8. Hacker News AI58

    RFC AI Tooling: 40x fewer tokens while reading the RFC corpus

    A skill and an MCP server share a stdlib-only Python core that finds, reads and flags IETF RFCs on demand, marking superseded documents on every result and allowing section-level reads instead of whole documents. The project reports a median 41x token saving from reading single sections, and offers optional full-text search over ~9,800 RFCs via a downloaded SQLite/FTS5 section index. It replaces the unmaintained mcp-server-ietf and can be self-hosted over HTTP for cloud sessions.

  9. MarkTechPost58

    vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model

    The vLLM Semantic Router team released d3-ultra, a 26.09B-parameter dense multimodal decision model under Apache-2.0 that does not generate text but reads an input once and returns probabilities for typed questions (choice, yes/no, score, box, track), with 131,072 tokens of context per question. It scored 87.2 on the team's private SR-Text test, the highest of six decision models compared, but trailed its sibling d3-turbo on SR-Box grounding (58.2% vs 60.0% over 10 domains). It was tested on one AMD Instinct MI325X with BF16 weights, where 255 questions about a photo took 2.16 s, and is available via Transformers, the vLLM Semantic Router Engine, a free API and a Hugging Face Space.

  10. MarkTechPost69

    Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing

    Baidu released Unlimited OCR 2, a 3.34B-parameter OCR model under the MIT license that parses documents of hundreds of pages in a single request using SR-SWA, a sparse attention scheme that keeps per-step decoding cost nearly flat. Baidu reports first place on MPDocBench (84.6) and MarathonDocBench (77.1), with heading hierarchy its lowest sub-score at 48.2 on long documents. All benchmark figures are self-reported, and the model runs on one CUDA GPU in BF16 with code on GitHub.

  11. Simon Willison32

    webframe: evaluated as a scraper and a sandbox

    An evaluation of webframe, Jeff Lindsay's Go tool built on the glaze library that runs a web page like a program in a hidden headless web view, found it fast and lightweight for scraping with performance comparable to shot-scraper. It also surfaced rough edges including CSP failures, a 0×0 viewport, hanging loads, missing HTTP metadata, and limited browser controls.

  12. Baseten Blog55

    BDN Sealed Volumes: publish once, mount anywhere

    Baseten released BDN Sealed Volumes, which let you publish read-only data such as model weights, LoRA adapters, indexes, and compiled engines once as an immutable, digest-addressed version and mount it in any deployment via config.yaml. Volumes are pushed or synced from Hugging Face, S3, GCS, Azure Blob, Cloudflare R2, CoreWeave buckets, or local directories, with chunk-level deduplication so updates upload only changed bytes, and mounts start in seconds without a download step. Existing weights-block deployments will be migrated automatically, and the feature requires Baseten CLI 1.1.0+ and Truss 0.18.30+ if applicable.

  13. AWS Machine Learning Blog34

    Fleet-wide computer vision lifecycle at the edge with SageMaker SDK v3

    This post updates a 2023 MLOps pipeline pattern for visual quality inspection to Amazon SageMaker Python SDK v3, replacing the deprecated SageMaker Edge Manager with SageMaker Neo for edge optimization. It describes a dual-branch SageMaker Pipeline that trains an anomaly detector and a defect classifier per camera, gates models on their exported ONNX artifacts, compiles them with Neo, and registers them with camera metadata, while a Step Functions and DynamoDB semaphore layer paces fleet training. The sample deploys via AWS CDK and a full synthetic-data run costs about $0.07 in compute.

Oct 25Sun
  1. The Decoder74

    Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

    Nvidia has open-sourced Flora, an open data engine built from three fine-tuned Qwen3.5 models that produces action descriptions, step-by-step progress ratings and success judgments for robot video. Nvidia estimates labeling one hour of video costs about 72 cents with Flora versus roughly $88.50 with GPT-6 Astra, though those figures are projections based on an assumed $10 per GPU-hour, and Flora trails Astra on some public benchmarks. Nvidia says Flora is already annotating up to 200,000 hours of video per day across 512 GB300 GPUs during training of its GR00T models, with models available on Hugging Face.

  2. The Decoder71

    OpenMobile-2 proves open-source smartphone agents can rival proprietary models

    Researchers from Nanjing University, SenseTime and other universities released OpenMobile-2, two smartphone agent models built on Qwen3.5 (9B) and Qwen3.6 (27B) parameters, together with the full training environment, data and reproduction instructions. The training ground covers more than 110 apps, including 35 rebuilt Chinese and global apps, and the larger model reaches 79.9 points on AndroidWorld and 59.6 on SPA-Bench with real devices. Holding data constant and increasing app variety raised success on unseen apps from 60.7 to 67.5 percent.

Oct 24Sat
  1. TechCrunch · AI40

    Hottest toy for Christmas? Nvidia CEO plays with adorable Microduck robots, 20,000 sell

    A video of Nvidia CEO Jensen Huang playing with waddling Microduck robots during his GTC Berlin keynote went viral, and Hugging Face's Clem Delangue said preorders surpassed 20,000 units the next day. The open-source robots from Pollen Robotics, which Hugging Face acquired in 2025 and Nvidia acquired in September, are expected to be delivered in time for Christmas. Delangue thinks they could be the best-selling robot of the year, possibly not even for kids.

  2. AWS Machine Learning Blog22

    Orchestrating SageMaker AI endpoint auto scaling with AWS CDK

    Updating a SageMaker AI real-time endpoint configuration fails through CloudFormation or CDK because SageMaker AI requires the scalable target to be deregistered first, and manual deregistration leaves CloudFormation unable to track Application Auto Scaling resources. The post presents a reusable CDK construct that replaces the native scalable target and scaling policy resources with three custom resources—Deregister, Register Scalable Target, and Attach Scaling Policy—ordered with explicit dependencies and driven by a ConfigHash so endpoint and auto scaling changes deploy together. It includes a walkthrough using the sample-sagemaker-endpoint-autoscaling GitHub repository and guidance on integrating the construct into an existing stack.

Oct 23Fri
  1. AWS Machine Learning Blog58

    5 production challenges in multi-agent systems and how Deep Insight solved them

    AWS's Solutions Architecture team describes Deep Insight, an open-source multi-agent system (Coordinator, Planner, Supervisor, Coder, Validator, Reporter, Tracker) that turns a CSV file and a natural-language prompt into a validated DOCX analysis report with charts. The post, part one of three, maps five production challenges to a stack: Strands Agents SDK Graph and Agents-as-Tools patterns for execution flow, role-specific Claude models on Amazon Bedrock, Amazon Bedrock AgentCore for managed deployment, AWS Fargate plus Application Load Balancer for isolated code execution, and DynamoDB, Amazon SNS and Amazon Cognito for monitoring. It also cites LG Electronics' ChatInsight and TechRecon as reuse of the same architecture.

  2. TechCrunch · AI71

    Recursive launches Flint, a laptop coding model it says was largely built by AI

    Recursive released Flint, a free coding model that runs on a laptop and was built largely with the help of its automated research system, the Eureka Machine, letting a small team build it from scratch in under four months. Flint is a mixture-of-experts model using 2.5 billion of its 40 billion parameters at a time, is available on Hugging Face under the Apache 2.0 license, and in Recursive's tests generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip. The company says the longer-term goal is for the Eureka Machine to achieve full recursive self-improvement, with humans overseeing the process.

  3. Hugging Face Blog58

    OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

    OlmoEarth v1.3 is a new satellite embedding model that reads a year of Sentinel-1, Sentinel-2 and Landsat time series to produce one 128-dimensional vector per 10-meter pixel, with the first 64 dimensions usable as a smaller embedding. Precomputed annual global embeddings are shared on Source Cooperative, and a browser-based OlmoEarth Embeddings Explorer lets users explore maps, find similar places, and train small classifiers without a GPU or sign-in. The 128-dimensional embeddings scored highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, and the model took 53% of v1.2's computing time and 18% of v1's, with weights, pretraining data and code all open.

Oct 22Thu
  1. The Decoder73

    Recursive shows first result of its AI-driven research process with Flint

    AI startup Recursive has released Flint, a mixture-of-experts model with 40 billion total and 2.5 billion active parameters, built in four months with the help of its "Eureka Machine" system that proposed, ran, and evaluated experiments. Recursive claims Flint leads coding benchmarks among MoE models under 50 billion parameters, though the results have not been independently verified, and the weights are available on Hugging Face under Apache 2.0. The company says Flint runs locally with context windows up to 512,000 tokens and generates over 125 tokens per second on an M4 Pro, but acknowledges it needs more guidance on complex tasks than large proprietary models.

  2. Hugging Face Blog38

    Ensemble forecasting with IBM Granite time-series models

    IBM introduces Granite Timeseries Ensemble, a recipe that combines four pre-trained Granite models (PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and TTM-r3) through one extensible forecasting interface to return median and quantile forecasts without retraining or fitted ensemble weights. The default Probabilistic Ensemble pools member quantiles with equal influence, while an adaptive IQR Ensemble weights members by forecast spread; on 97 GIFT-Eval configurations both ensembles beat all four individual members on aggregate MASE and average MASE Rank, with the Probabilistic Ensemble reaching 0.683 MASE versus 0.685 for the best single member. The authors note the MASE margin is small, that PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL, and that the recipe forecasts channels independently and loads four checkpoints, increasing inference time and memory.

  3. NVIDIA Blog40

    Pioneers Across Europe Deploy NVIDIA Earth-2 Open Models to Power Real-World Weather and Climate Prediction

    NVIDIA Earth-2 offers developers and researchers a family of open AI models and tools, running on NVIDIA computing, for global weather prediction, detailed regional forecasts and environmental research. StormScope EU, a regional short-term forecasting model trained on European weather and climate data, uses EUMETSAT satellite observations to predict cloud conditions across 16 Meteosat channels at 10-minute intervals, and is being shown in a live interactive demo at GTC Berlin built with Earthmover. Organizations including AXA, Vaisala Xweather, the University of Manchester, TotalEnergies and Space42 are adapting Earth-2 models such as FourCastNet 3, CorrDiff and PhysicsNeMo for cyclone risk, lightning-enhanced forecasting, U.K. air quality, solar forecasting and radar-based surface-height estimation.

Oct 21Wed
  1. Hugging Face Blog58

    Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

    IBM released granite-speech-5.0-2b, a new speech-language model trained on 390,000 hours of public audio that transcribes English, French, German, Spanish, Portuguese, and Japanese, translates to and from English in those languages plus English to Italian and Mandarin, and supports language-ID prompting and keyword lists up to 2,000 terms. It also updated granite-speech-5.0-470m-turboctc, which adds multilingual transcription and native punctuation and capitalization and reaches 19,870 RTFx on the Open ASR Leaderboard (about 1.6x faster than at launch), and granite-speech-4.1-2b-plus, which cuts word diarization error rate and speaker-attributed WER by 7% each and long-form WER by 24%. All three ship under Apache 2.0 with support in Transformers, vLLM, llama.cpp, and mlx-audio.

  2. NVIDIA Blog73

    NVIDIA Open Models Advance Europe’s Sovereign Intelligence Across Industries

    NVIDIA is expanding its family of open models across language, robotics, biology and quantum computing, including the new Nemotron 3.5 Super VL and Nemotron 3.5 Bolt models, Nemotron Retriever models, the Kumo family for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo for healthcare and science, and Ising for quantum computing. The company says the open models give organizations a foundation they can customize with local data and deploy on infrastructure they control, with Nemotron built using data and expertise from the NVIDIA Nemotron Coalition including European partners Black Forest Labs, H-Company and Mistral. Availability details include Nemotron 3.5 Super VL on Hugging Face, ModelScope, OpenRouter and build.nvidia.com, with Nemotron 3.5 Bolt E4B 4GB expected Friday, Oct. 23.

    Why it matters: The piece is a broad product roundup that shows how NVIDIA is positioning open models as a sovereignty-friendly foundation for European enterprises and researchers.

  3. Latent Space83

    [AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback

    DeepSeek released DeepInfer, its production inference stack built on vLLM with speculative decoding, SWA Bounded Replay and structured output, reporting 155 trillion tokens/day from one cluster, 96% input-cache hits, 400 ms average time to first token and $38 revenue per accelerator-hour at off-peak prices. The release excludes KV-cache management and load balancing, which remain coupled to internal infrastructure, and vLLM reported decode throughput within 6% of DeepInfer's published results on GB200 under a matched configuration. Community margin estimates of roughly 80% depend on a cache-heavy token mix and omit costs such as SSD cache, networking and idle capacity, so the figures are first-party production claims rather than audited measurements.

    Why it matters: Useful for readers tracking open-source inference economics, since it pairs DeepSeek's production claims with vLLM's comparative benchmark and a careful accounting of what the margin figures do and do not include.

Oct 20Tue
  1. Hugging Face Blog65

    Hugging Face ❤️ TorchTPU: TPUs, Natively

    TorchTPU adds TPU as a native PyTorch backend, letting Transformers, Diffusers and Kernels run on TPU by sending models to "tpu" instead of going through torch_xla or JAX, and Optimum TPU is no longer required. The post shows loading a model with device_map="tpu", running Diffusers pipelines with pipe.to("tpu") plus optional torch.compile, and using Pallas kernels from the Kernels Hub, with reported speedups of 21.5× on Qwen3-30B-A3B and 12.8× on Gemma 4 26B-A4B at batch size 64. It also reports that 96.5% of tests run on TPU pass, with 34 issues filed on the TorchTPU repo and 19 already fixed.

    Why it matters: It documents a new native PyTorch TPU backend and concrete usage patterns for the Hugging Face stack, useful for readers weighing TPU support.

Oct 15Thu
  1. Mistral AI67

    Introducing Timestral

    Mistral releases Timestral, its first family of time-series foundation models in four sizes from 5M to 209M parameters, with the largest ranking first on GIFT-Eval, TIME and fev-bench among pre-trained zero-shot models. Timestral produces probabilistic multivariate forecasts, supports past-only and known-future covariates, and runs on CPU without a GPU. It combines synthetic-first pre-training, covariate-aware attention and decoupled horizon forecasting, and can be fine-tuned on proprietary data through Mistral Forge.

    Why it matters: Useful for teams weighing time-series foundation models, since it reports benchmark standings, a size-versus-accuracy trade-off and CPU latency figures alongside the adaptation options.

Oct 8Thu
  1. Microsoft Research71

    Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses

    Microsoft Research Asia introduces Harnessed Agentic RL, a paradigm in which the same agent harness used in deployment participates directly in reinforcement learning, and open-sources Agent Lightning v1.0, a roughly 3,500-line framework with an API Gateway, Rollout Controller, and Customized Trainer built on verl. Agents run as standard Kubernetes jobs without commercial sandbox services, and Collocated Async RL shares GPUs between rollout and updates, achieving about a 2x end-to-end speedup over synchronous RL. An end-to-end coding agent pipeline raised Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified using about 6,000 training samples.

    Why it matters: It offers a concrete, reproducible recipe for training deployed agents in place, with code size, infrastructure, and benchmark numbers stated explicitly.

Jul 29Wed
  1. Berkeley AI Research60

    From CUDA to MLX: How K-Search Brings Decades of Kernel Expertise to Apple Silicon

    IBM Research extended K-Search, an evolutionary LLM-driven kernel search framework from UC Berkeley Sky Lab, with a native MLX backend and a structured CUDA-to-MLX translation layer that maps CUDA primitives, constraints and optimization patterns onto Apple Silicon equivalents. On Apple Silicon the approach reached 0.97x the speed of the native MLX Attention kernel and up to a 20x prefill speedup over the community mlx-lm implementation on the Mamba SSM kernel, attributed largely to adding a parallel scan. The authors say the bottleneck was not the LLM's ability to write Metal code but the quality of context and constraints supplied to it.

    Why it matters: Describes a concrete method for transferring CUDA kernel expertise to non-NVIDIA hardware, with reported benchmark numbers and an open-source reproduction path.

Mar 13Fri
  1. Berkeley AI Research38

    Identifying Interactions at Scale for LLMs

    This blog post introduces SPEX and ProxySPEX, algorithms that use strategically selected ablations and sparse recovery from signal processing and coding theory to identify influential feature, data, and model-component interactions at scales far beyond prior methods, with ProxySPEX exploiting hierarchy to match SPEX using roughly 10x fewer ablations. The authors demonstrate the framework on sentiment analysis, a modified trolley problem with GPT-4o mini, CIFAR-10 data attribution with ResNet, and attention head pruning on MMLU, where a ProxySPEX-informed pruning strategy improved target-task performance. Code for both algorithms is available in the SHAP-IQ repository.