FDE-Gym: The Last Mile of AI Delivery
FDE-Gym contains 95 end-to-end delivery tasks drawn from 20 customer scenarios across 18 sectors and uses scripted stakeholders in simulated customer environments.
Technical AI signals ยท Daily at 8 PM ET
FDE-Gym contains 95 end-to-end delivery tasks drawn from 20 customer scenarios across 18 sectors and uses scripted stakeholders in simulated customer environments.
The post describes agent-assisted CuTe DSL changes that speed up Kimi K3 decode on NVIDIA GB300 with correctness tests and reproducible benchmarks.
DeepSeek says V4.1-Flash-1017 improves benchmark performance, speeds up task completion, and produces shorter outputs than the previous version.
The recipe keeps SFT unchanged and adds a midtraining stage on 500B tokens at 256K context, lifting SWE-bench Multilingual from 55.8 to 62.8 on Nemotron 3.5 Lightning.
The harness routes routine choices to a distilled 4B decision model only above a fitted confidence threshold, cutting cost by 14.2% on 500 tasks without changing outcomes.
Tencent says Hy-Memory-30B-A3B ranks first on four evaluated harnesses on fixed 20% benchmark subsamples, with a 55.49% macro average versus 54.62% for Opus 5.
The post points to an open-source humanoid robot project on GitHub that builders can study and modify.
It reports Jev matching Claude Sonnet 5 accuracy on 500 SST-2 examples while cutting median latency from 2,076 ms to 376 ms and boosting throughput.
The setup proxies all agent traffic through a transparent HTTP proxy so CLI, local MCP, and LLM provider traffic can run without holding API keys.
It describes an open-world agent harness built from eight components, including memory, tool guidance, subagents, skills, scripts, and plugins.
AMD is acquiring World Labs for $8.2 billion in stock, and Fei-Fei Li will become AMD's chief scientist.
Unity adds real stock identifiers and 1,040 features, and Numerai plans to merge Numerai and Numerai Signals into one tournament.
Dirac Labs is adding diamond-based sensors and an AI model that turns magnetic-field readings into a position fix without satellites.
The model runs on-device in about 3 GB of memory and read a humidifier panel's LCD and switch from one photo.
The thesis highlights that refusal-direction ablation can raise introspection scores, but the result comes with significant model degradation.
The larger experiment covers 10 models and more than 5,000 games with a four-parameter fit for deception, divisiveness, accuracy, and teammate assistance.
The project finds that Qwen 3.8 27B can use reward laundering to improve chain-of-thought controllability and generalize to unseen controllability tasks.
The study says Qwen3.6-35B-A3B and Qwen3-8B can compose fabricated propositions when the bridge entity is real, with success depending on how well known the bridge is.
Muse Spark Open adds 816B MoE support, 1M context, multimodal inputs, and initial TPU support on vLLM.
vLLM says Beam-1.0 has 501B total parameters, 23B active parameters, and DFlash2 speculative decoding with adaptive verification.
Modal says VM Sandbox filesystems got faster by speaking in blocks instead of files.
The paper proposes FBS Debate, which gives worst-case correctness guarantees instead of average-case guarantees for stable decompositions.
Modal added Muse Spark Open to Auto Endpoints for coding-optimized multimodal inference.
Magnetar supports standard programming languages, faster verification, and quantum-resistant security for zero-knowledge proof computing.
Hugging Face is introducing Muse Spark Open, an open-weight multimodal mixture-of-experts model from Meta.
NVIDIA says RTX Spark lets developers code, test, and deploy on a single machine.
Google Research outlines building blocks for autonomous wet labs grounded in biological reality.
The LessWrong post says replay outperforms projection sampling on the capability-alignment tradeoff, while projection recovers much of the lost capability and also reverses much of the reductio effect.
The vLLM blog says the Kimi K3 optimization work found the bottlenecks that made inference faster.
The top comment links a short summary of the transformer-memory article.
Commenters discuss Claude's claim that NIST P-192, P-224, P-256, P-384, and P-521 seeds came from the C library's lrand48 PRNG seeded with a Unix timestamp.
Commenters say the image model seems weak at generating people and text details, according to the discussion.
Commenters note that the benchmark graph compares Ultramarine only against models it beats on every metric.
Commenters say the tool reports around 46 tok/s and note that benchmark numbers can differ from real usage.
Commenters say the project runs Qwen 3.8 27B at 2-bit quantization on a jailbroken PlayStation 5.
Commenters describe it as a browser-based ESP32 simulator that can run real ESP32 C code compiled to WebAssembly.
Commenters focus on the garage-built setup and a 3D-printed platform for holding the chip.
Commenters say agents help fix bugs by reproducing, debugging, and patching reports, while others note Greg KH pushes back on low-quality LLM-generated submissions.
Commenters say Estonia has used e-voting since 2001 and widely since the first parliamentary e-voting in 2007.
Commenters say the code borrows flow-based programming ports and discuss avoiding concurrency bookkeeping and traps.
Waypoint-OPD uses an intermediate RL checkpoint to compare early and late policy shifts before distilling them to a student.
MRTPO assigns each reward's advantage only to the tokens its verifier checks instead of spreading it across the whole response.
AutoSFT defines a data action space for adjusting proportions, pruning, repairing, replacing, and synthesizing training data.
The paper introduces SAE Additivity Residuals to measure the behavioral variation left unexplained by a chosen interaction order.
VLA-ฮ supervises physical state transitions between current and future observations instead of mapping observations directly to actions.
EdgeWise keeps an SLM on device as the primary solver and sends only critical steps to a cloud LLM to reduce inference cost.
VHA uses Q Premix and Linear Postmix to improve KV cache size, training and prefill FLOPs, and model quality together.
FlashMask-3 extends column-wise sparse masks to Hopper GPUs and distributed context parallelism while keeping structured attention masks compact.
Athanor lets frontier agents post-train a small raw base model on real GPUs with full research autonomy.
Terminal-RSI alternates Recursive Self-Rewrite of verified trajectories with critic-inheriting Recursive Policy-Refine and introduces TB-GAE, FDR, and SPR for terminal agents.
The paper builds AgentDecisionBench to test decision models from isolated questions to control of a running agent and says harnesses can delegate many multiple-choice decisions to models like Jev.
The paper builds a slide-native dataset with over 13M continual-pretraining records, 1.9M supervised fine-tuning records, and VisSlideBench with 1,527 screened instances.
The episode features Stanford professors Clark Barrett and Mykel Kochenderfer discussing automated reasoning and aircraft collision avoidance as a safety benchmark.
The session applies short- and long-term memory plus skill distillation to an agent that flags suspicious bank transactions around the clock.
Harrison Chase says LangChain processes more than one billion events a day through LangSmith and serves 35% of the Fortune 500.
Six AI users compare OpenAI DevDay products with coding agents, personal assistants, decision models, subagents, and context management.
The workshop shows a Claude-powered financial analyst agent traced in Arize Phoenix, then gated in CI with npm test.
The episode covers autonomous password lockouts, hidden valence signals, and Lean proofs used to find Linux kernel bugs.
The episode says an agent made 200 calls in 90 minutes and got its user banned from Resy, then discusses OAuth scopes, budgets, and separating an agent's brain from its hands.
EmbeddingGemma 2 maps text, images, video, and audio into one 768-dimensional vector space with 740M total parameters.
The model report says DeepSeek-V4.1-Flash-1017 improves reasoning, coding, agentic tasks, and visual benchmarks while reducing task time and output length.
MiniCPM-V 4.7 is a 35B-A3B multimodal MoE model with lossless on-device inference using Lookahead Expert Scheduling.
Drama 3 is a 35B-parameter text-to-speech model with 3B active per token, built for multi-character scenes and long-form audio.
LFM2.5-24B-A2B is a 24B-parameter reasoning model with 2B active parameters that runs on-device and targets agentic tool use.
Unlimited OCR is positioned for one-shot long-horizon parsing and now supports vLLM inference.
The model card tags SAM3 for video feature extraction and mask generation.
Saluki packages Qwen3.8-27B into a 7.89 GB GGUF that keeps tool calling intact and adds an optional vision adapter.
The dataset is tagged for forward-deployed engineering benchmarks and agents.
EgoDemo is a video dataset for egocentric video classification.
The model weights are packaged for Hugging Face Transformers and compatible with vLLM, SGLang, and TokenSpeed, with a hosted Qwen Cloud version listed as coming soon.
Anthropic's directory includes internal and third-party Claude Code plugins and warns users to trust plugins before installing them.
CLI-Anything ships a CLI hub that installs community-built CLIs and shows agent demos for CAD, 3D scenes, diagrams, gameplay, and subtitles.
The repo now supports any agent using the open Agent Skills standard and adds K-Dense BYOK with 40+ models and 177 skills.
ai-memory lets coding agents carry long-term notes across more than twenty harnesses, including Claude Code, Codex, Cursor, Gemini CLI, and OpenCode.
Google's Android skills repo provides modular SKILL.md instructions for workflows where evaluations show LLMs underperform.
Orca runs Codex, Claude Code, OpenCode, or Pi in separate worktrees and includes a mobile companion for steering agents from a phone.
Claude Code is a terminal coding agent that can explain code, execute routine tasks, and handle git workflows through natural language commands.
VoiceStudio supports voice cloning, dubbing, dictation, transcription, and audiobook creation in 646 languages.
OpenCode is an open-source AI coding agent that also ships a desktop app for macOS, Windows, and Linux.
XERJ automatically indexes folders for local code search and says a controlled study cut output tokens 2.7x at the same 16/16 solve rate.
The repository is an AI-powered job application framework that runs locally, uses Claude Code to evaluate postings, tailor CVs, write cover letters, and prepare interviews.
The project warns to install ECC only from verified channels, including the GitHub repo, npm packages, GitHub App, plugin slug ecc@ecc, and project website ecc.tools.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.