Beam opens as a 501B agentic MoE
Reflection is putting real weight behind the open-agent race: Beam is a 501B-parameter model with 23B active parameters, trained end-to-end from scratch, with weights due this month.
Technical AI signals · Daily at 8 PM ET
Monday, October 5, 2026
Today was an operator-heavy mix of Reflection’s open-weight Beam release, agent memory and retrieval upgrades, and new verifier-centric papers for long-horizon task reliability.
Reflection is putting real weight behind the open-agent race: Beam is a 501B-parameter model with 23B active parameters, trained end-to-end from scratch, with weights due this month.
Microsoft’s CorpusMap turns repeated entities into linked pages, lifting answer quality by 6.4 to 11.7 points while trimming input tokens by 34% to 57%—exactly the kind of retrieval win that changes agent cost and latency.
Google’s VeriHarness pushes the current control pattern further by giving the verifier a workspace, evidence collection, and repeated sampling so it can check long-horizon agent outputs instead of scoring them in one pass.
Cognition is productizing persistent agent state: Devin now builds a memory graph across sessions and runs overnight cleanup plus latent-info discovery to keep that memory usable.
GitHub’s ReviewBench looks more useful than toy review sets, using representative PRs, multi-source ground truth, calibrated evaluation, and metrics aligned with how code review systems are actually deployed.
Interfaze-1-lite returns confidence scores and bounding boxes on every run.
Agr comes in 31B and 360M versions, scores 58.15 on Decision Index 0.2.1, and returns typed values with probabilities instead of text.
SimularAI says Sai reaches 79.38% partial success rate and costs $14.34 per task versus Claude Computer Use with Opus 5 at $24.11.
The post describes a spatial verification pipeline using TEE attestation on IMU streams and zero-knowledge anti-replay proofs.
Hooki watches buys, sells, volume, holders, and market cap, then rewrites the token hook using fees that fund its wallet.
OpenAI says the post explains when watermarking applies, how detection works, and why access starts with researchers.
The authors pretrained multiple 30B LLMs from scratch and found filtering subversion-relevant information sharply reduced that knowledge while retaining general ML knowledge.
The paper measures self-recognition and self-report, then shows interventions on either can modulate emergent misalignment.
Forethought says current AI systems are hard to benchmark for persuasion and defines persuasion as directed social influence and related abilities.
Together Link puts GLM 5.3 and Kimi K3 into the coding agent you already use and says model spend drops by over 50%.
Simon Willison reran the experiment on a DGX Spark using Qwen3.8-27B-Q4_K_M.gguf after a GPT-4o version had been shared on Bluesky.
The post discusses a proposal that AI R&D can only be assisted by AIs trained at least 9 months earlier.
The post argues that prosaic alignment work can speed capabilities advances and may reduce the chance of a warning shot that would slow RSI.
Chris Paxton discusses humanoid robots that can dynamically interact with the world and with people.
The post asks, "Who chooses what AI gets to do?"
Google Research examines privacy and security problems in agentic systems from a contextual angle.
Hugging Face describes an agent that reported completion while the database showed otherwise.
Commenters note that Dust is more computationally expensive than backpropagation but more easily parallelizable.
Commenters say the agents found two room-temperature magnetic semiconductor candidates, and one commenter compares the result with the LK-99 debacle.
Commenters focus on whether search APIs allow storing and resyndicating results, and one commenter says Gemini Flash Lite 2.5 offers 1,000 free Google searches per day.
Commenters say the update is not the newest status and point to Google’s future QPR1 or QPR2 changes.
Commenters say the rewrite happened in about three weeks and note mold had previously been written in C.
Commenters note the game logic is about 5,900 lines of SQL and praise the query planner being used as a state machine.
Commenters say swap can trigger page-fault problems and argue that latency-sensitive systems should disable swap.
Commenters describe ExplainDB as a didactic Python database with a hands-on DBMS course in Jupyter and note it is still a work in progress.
Commenters say the tool helps surface local news and that the team first tried to balance quality with the cost of searching.
The top comments ask how cgroups v2 and newer seccomp features would change the 2016 implementation.
Commenters argue about macro-based type-safe containers in C versus using C++ instead.
The paper asks whether a general-purpose VLM can operate a robot by reasoning directly from observations, issuing actions, and continuing across steps.
Tencent Hunyuan’s RSR uses Qwen-3.8-27B, planner-generated runbooks, critic screening, and fresh sandboxes to reconstruct successful trajectories.
WEFT couples agentic system construction, execution-driven self-evolution, and stable post-training instead of scaling environments in isolation.
HyperBrowseComp contains 423 manually authored and human-validated questions across 13 languages and requires evidence from videos, scanned documents, images, or maps.
SimuVerity evaluates 101 text-to-executable Simulink tasks with a hierarchical scorer that checks delivery, executability, and engineering qualification.
The benchmark includes 8,000 controlled simulation cases from 80 families and tests text-video retrieval, physical-property regression, and video-description classification.
The authors build FoldingCorpus and Fold2Reason, then post-train with discrete structural answers and continuous 3D geometry.
Meta’s NAVA-WAM pretrains an action policy directly from observation-only videos instead of using latent-action transfer or separate robot trajectories.
Latent-MOPD transfers both teacher predictions and hidden states using late-layer targets, a shared projection, and domain-grouped updates.
Pivot-SD trains only the high-impact commitments selected during denoising instead of supervising the final text or whole denoising steps.
ProWAM predicts actions plus an ordered sequence of sparse visual sub-goals instead of generating dense video rollouts.
Triadic linear attention writes a triadic outer product of a key, a second key, and a value into a three-dimensional recurrent state.
Jess Wang reports that agentic search and vector search reached the same accuracy on TypeScript Go fix-PR tasks, while vector search cost about four times as much.
Adam Schiff argues that voluntary AI rules mean little while discussing AI oversight, tech self-dealing, and a White House meeting with AI CEOs.
Harrison Chase outlines three pillars for AI products: an open model-neutral harness, a learning loop from user behavior, and governance for internal agents.
Paige Bailey demos a browser-run model, then introduces Gemma 4 as an open model family ranging from 2 billion to 31 billion parameters.
Laurie Voss demos an improvement loop that adds tracing to the Wonder Toys app, diagnoses empty results and narrow filters, and adds price filtering to a toy search tool.
The episode shows a 96-line Python agent loop that ran four trips in 13 seconds and solved five of five tasks.
Qwen3.8-27B ships weights and config in Hugging Face Transformers format and is planned for a hosted Qwen Cloud version with 1M context by default.
LTX-2.5 is tagged for image-to-video, text-to-video, video-to-video, and audio-to-video generation.
Clef is a 27B multimodal decision model that reads text, JSON, images, or video and returns probabilities for every allowed option in one forward pass.
Clef-Flash is a 9B multimodal decision model that returns probabilities for every allowed option without free-form text generation or output parsing.
Kolibri is a 78B mixture-of-experts reasoning model with explicit reasoning mode, tool calling, and a 1,048,576-token context length.
ZDTaichu5.0-9B combines a Qwen3.5-9B backbone with a C-RADIOv4-H vision encoder for text, image, and video input.
Nemotron 3 Diarization handles up to eight speakers and supports both streaming and offline inference.
The dataset is tagged for agents, llm-agents, computer-use, and ai-safety.
The dataset contains 10,000 multi-turn coding and reasoning traces and about 500 million tokens.
The space packages a research article template with chapter MDX files, embedded HTML/D3 visualizations, and hosted data assets.
The dataset packages RL training environments for LLM agents, including executable tests, cyber vulnerability reproduction, rubric-based judging, and Docker images.
The framework lets an agent drive web or mobile apps from natural-language goals, then replays recorded agent steps with no model calls until the app changes.
Claude-Mem captures tool usage, generates semantic summaries, and preserves context across sessions for Claude Code.
The plugin generates STEP, GLB, STL, or 3MF files and adds DFM checks, engineering drawings, and fabrication-service connections.
Agent Reach adds browser-based internet access for AI agents and says BrowserAct can extract data from sites like Amazon, LinkedIn, X, and Google Maps.
OpenMontage is an open-source agentic video production system, and Bloome lets multiple AI agents collaborate in one conversation.
Cloudflare OS is an internal AI productivity environment that also acts as an operating system for AI workloads.
The DNS ad-blocker stores 140,000+ domains as 40-bit hashes in flash, fits in about 0.7 MB, and uses about 50 KB of RAM.
AnyPS5 converts executables to native Linux or Windows formats without emulation or a separate runtime process.
The project bundles specialized AI agents and now has a native app that installs them into Claude Code, Cursor, Codex, Gemini, and more.
T3 Code controls agents on your machine through iOS, Android, web, and Electron apps, and it works with Claude Code, Codex, Cursor, Grok Build, OpenCode, and Google Antigravity.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.