Automated AI newsroom

AI News

High-signal AI developments, independently scored, summarized and organized for fast reading.

Live intelligence

Editor-selected AI developments

752 updates

Research Editor selected

Stanford and UC Berkeley Release TokenGate for Learned Token Pruning in Long-Context LLM Prefill

Researchers at Stanford and UC Berkeley have released TokenGate, a drop-in module that adds learned token pruning to pretrained LLMs so long prompts are processed faster. A small MLP gate scores tokens at a chosen layer, and only the top-K are passed to later layers and given KV entries; an MoE-router-style logit bias makes the top-K selection trainable. The team reports 1.4x to 11.3x faster time-to-first-token and 1.5x to 3.4x lower peak prefill memory versus full prefill, tested on Qwen3-1.7B

Why it matters: It gives practitioners an open, trainable approach to cutting long-context prefill cost and memory in existing LLMs with reported multi-fold speedups.

MarkTechPostOct 25, 2026Score 83

TokenGateStanford UniversityUC Berkeley

Products Editor selected

Anthropic Launches Free OSS Scanner for AI Vulnerability Audits

Anthropic has introduced OSS Scanner, a free opt-in program that uses its strongest models, including Claude Mythos, to audit critical open-source projects. Maintainers enroll via a pull request adding a project.yaml and Dockerfile; audits run in isolated VMs without internet access. Reports go directly to maintainers without human review, including reproducers, explanations, bisections, and candidate patches. In the first 11 days after the October 8 launch, Anthropic reported 77 critical and 80

Why it matters: This gives maintainers and security teams a concrete look at a new AI-driven vulnerability scanning program, including its enrollment model, accuracy data, and tradeoffs around unreviewed reports.

MarkTechPostOct 25, 2026Score 81

Anthropicopen-source securityAI vulnerability detection

Research Editor selected

Microsoft Research Open-Sources Agensh, an Orchestrator-Free Multi-Agent Harness Tested at 1,024 Agents

Microsoft Research has open-sourced Agensh (MIT), a decentralized harness for large teams of coding agents that removes the central orchestrator. Each worker runs the same five-step prompt-based loop, claims sub-tasks, and shares state through a Gitea repo, a Mattermost channel and a typed context board. On the five hardest ProgramBench tasks, more agents improved scores: with GPT-5.6-sol the mean rose from 19.31% with 1 agent to 28.78% with 128. The README reports GPT-6.1-sol reaching a 79.59%

Why it matters: It documents a concrete open-source multi-agent architecture with published scaling behavior and benchmarks, giving agent builders a reproducible alternative to orchestrator-worker designs.

MarkTechPostOct 25, 2026Score 85

Microsoft Researchmulti-agent systemscoding agents

Models Editor selected

NVIDIA Releases Nemotron 3.5 Bolt E4B, an Open Small Model for Local Agents

NVIDIA released Nemotron 3.5 Bolt E4B, an open small language model built by RL post-training of Google's Gemma 4 E4B without supervised fine-tuning. It has 8B total and 4B effective parameters with 64K context, targeting local voice assistants, game NPCs and robot interfaces on RTX, Jetson and Spark hardware. NVIDIA reports top scores on 15 of 21 benchmarks in its own harness, but Qwen3.5-4B leads on Tau2-Bench and long-context tests. Weights, data and training recipes are public, with the Q4_K

Why it matters: It gives developers a concrete open-weight option for running capable tool-calling agents locally on constrained hardware, with disclosed training methods and honest benchmark weaknesses.

MarkTechPostOct 24, 2026Score 83

NVIDIANemotronGemma

Industry Editor selected

OpenAI to stop training AI models on the live internet

OpenAI says it will no longer train its AI models on the live internet, citing "significant risks." The reported policy change comes weeks after one of its engineers said that approach wouldn't work. The source provides no further details on how training data will be sourced going forward.

Why it matters: It signals a notable shift in how a leading frontier lab plans to source training data, which matters to anyone tracking model development practices.

TechCrunch AIOct 24, 2026Score 81

OpenAImodel trainingAI safety

Models Editor selected

Moonshot releases Kimi K3.1 Pro with open weights promised by Nov 4

Moonshot launched Kimi K3.1 Pro on its API, priced at $2.00 per million input tokens and $10 per million output tokens, with a higher-priced "Highspeed" variant that runs 2-3x faster. The model is not yet open weights, but Moonshot promises full weights by Nov 4, 2026. Moonshot's own benchmarks place it near Claude Opus 5.5 and GPT-6 Astra, and Artificial Analysis scored it 55, behind only Opus 5.5 at 58. Simon Willison tested it through the Moonshot API, including a pelican SVG comparison acros

Why it matters: It signals a soon-to-be open-weight model reaching near-frontier benchmark standing at competitive API pricing, which matters to teams choosing models.

Simon WillisonOct 24, 2026Score 80

MoonshotKimi K3.1 Proopen-weight models

Tutorial Editor selected

AWS details multi-turn reinforcement learning on SageMaker AI

AWS published a hands-on guide to Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI, which trains models on end-of-episode rewards rather than per-step labels. Agents hosted on Bedrock AgentCore (or EKS, EC2, Fargate) play multi-step tasks and return a single reward, with GRPO updating a LoRA adapter. In the guide's Wordle example, gpt-oss-20b improved its solve rate on held-out words from 61% to 84% at default reasoning effort and from 15% to 69% at low effort, at roughly $339 in

Why it matters: It gives practitioners a concrete, cost-transparent walkthrough of agentic multi-turn RL on a major cloud platform, including benchmark numbers, supported models and failure modes.

AWS Machine Learning BlogOct 24, 2026Score 83

AWSAmazon SageMaker AIMulti-Turn Reinforcement Learning

Products Editor selected

Bayer Crop Science's Agentic Decision Engine runs 22 specialist agents on Amazon Bedrock AgentCore

Bayer Crop Science built an Agentic Decision Engine (ADE), a natural-language multi-agent layer over its Connected Decision Engine, to help breeders and researchers query ML workflows without navigating each underlying system. A supervisor agent routes requests to 22 domain-specific specialist agents that call backend APIs through AgentCore Gateway. Bayer reports that ADE resolves about 80 percent of user questions without engineer involvement and cut per-API integration overhead sharply by expo

Why it matters: It offers a detailed, concrete production case study of enterprise agent architecture, observability and cost-attribution practices that teams building multi-agent systems can learn from.

AWS Machine Learning BlogOct 24, 2026Score 80

Bayer Crop ScienceAmazon Bedrock AgentCoreagentic AI architecture

Products Editor selected

Google Cloud Launches Agent Compute Engine: One MicroVM per AI Agent

Google Cloud announced Agent Compute Engine (ACE), a compute service that runs each AI agent session in a hardware-isolated Cloud Hypervisor microVM with its own guest kernel, root access and Hyperdisk storage. The design keeps zero credentials inside the sandbox: traffic crosses a virtio-vsock proxy to an Agent Gateway that swaps in real backend credentials after policy checks, and sandboxes get only a short-lived SPIFFE identity. Idle agents are checkpointed to zero active resources and resume

Why it matters: It details concrete Google Cloud infrastructure and security architecture for running AI agents, useful for teams choosing how to sandbox agents safely.

MarkTechPostOct 23, 2026Score 80

Google CloudAgent Compute Engineagent sandboxing

Products Editor selected

Amazon SageMaker adds training session APIs for custom LLM training loops

Amazon SageMaker AI introduced training session APIs that let developers write their own LLM training loops in Python while SageMaker handles GPU allocation, distributed execution, fault recovery and checkpointing. The APIs expose operations such as sample(), forward_backward() and optim_step() through ServiceClient, TrainingClient and SamplingClient, with billing per token processed. A companion cookbook provides SFT and RL recipes, and AWS reports a GRPO example on GSM8K raised held-out accura

Why it matters: It gives ML engineers and alignment researchers a concrete way to run custom SFT, RL and preference-optimization loops without provisioning or managing GPU infrastructure.

AWS Machine Learning BlogOct 23, 2026Score 83

Amazon SageMakerLLM post-trainingreinforcement learning

Models Editor selected

Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0

Gradium, a Paris voice AI company spun out of Kyutai, has released Phonon, a family of small text-to-speech models that run fully on-device. The English Phonon 31m ships under Apache 2.0 at 64.5 MB (q8 GGUF) with 31M parameters and 37 built-in voices, producing 24 kHz mono audio with no cloud call or GPU. It uses a causal transformer backbone, SentencePiece tokenization without phoneme conversion, a single-step flow-matching head and a reduced Mimi decoder. Gradium's own evaluations report first

Why it matters: Provides concrete deployment, licensing and performance details for a new openly licensed on-device TTS model developers can self-host.

MarkTechPostOct 23, 2026Score 81

Gradium Phononon-device TTSopen-weight speech models

Tutorial Editor selected

Measuring, not guessing, SageMaker autoscaling targets for TTS endpoints

This AWS Machine Learning Blog post walks through deploying the Kokoro-82M text-to-speech model as a custom container on an Amazon SageMaker AI real-time endpoint, then deriving autoscaling thresholds from two live measurements instead of guesswork. It measures Q_max (the per-instance concurrency ceiling that still meets a 3,000 ms p95 time-to-first-audio SLO, found to be 50 in the sample run) and T_total (end-to-end scale-out time, 260.2 s measured). The tooling then computes scale-out and scal

Why it matters: It gives practitioners a concrete, measurement-driven method and tooling for tuning autoscaling on latency-sensitive TTS inference endpoints, a gap that default LLM- and request/response-oriented guidance does not cover.

AWS Machine Learning BlogOct 23, 2026Score 80

Amazon SageMaker AIKokoro-82Mtext-to-speech

Models Editor selected

IBM Releases Granite Speech 5.0 2B, an Apache 2.0 Multilingual Speech Model

IBM released Granite Speech 5.0 2B, a 2B-parameter open speech-language model for transcription and translation in 6 languages under Apache 2.0. It combines a 16-block Conformer speech encoder with a window q-former and a fine-tuned granite-4.0-1b-base backbone, and adds language-ID prompting, keyword biasing up to 2,000 terms, and requested punctuation and capitalization. IBM reports 5.68% average English WER on the Open ASR Leaderboard, ranking second among open-weight models behind Qwen3-ASR-

Why it matters: Developers get a concrete, benchmarked open-weight ASR and translation release with deployment options and licensing details they can evaluate directly.

MarkTechPostOct 22, 2026Score 83

IBM Granitespeech recognitionopen-weight models

Products Editor selected

NVIDIA Expands Cosmos Platform for World Model Training and Deployment

NVIDIA announced an expansion of its Cosmos platform for building world models used in physical AI systems, adding training tools, simulation checkpoints and Cosmos Data Factory blueprints. The release introduces multimodal multisensory data curation, support for adapting open models such as GR00T, Alpamayo, QWEN and Pi0.5, and updated Cosmos 3 NIM microservices. NVIDIA also described Cosmos 3 Nano variants for simulation-based policy testing and a Metropolis VSS 3.3 blueprint. The company cited

Why it matters: It gives AI developers a concrete, detailed look at new open tooling and deployment options for building physical AI world models.

NVIDIA BlogOct 22, 2026Score 81

NVIDIAworld modelsphysical AI

Research Editor selected

Perplexity and NVIDIA Researchers Introduce JOLT: Joint Teacher-Student RL for Post-Training

Perplexity and NVIDIA researchers propose JOLT (Joint On-policy Learning and Teaching), a post-training method where a single model is trained as both teacher and student, with the teacher receiving privileged information and also being optimized. JOLT+ adds the student's own outcome reward to the distillation signal. On MATH, JOLT matched GRPO accuracy while using 6.5x fewer processed training tokens according to the authors' fitted curves, and JOLT+ led GRPO across the five tested settings. Th

Why it matters: It offers a concrete post-training technique with claimed compute savings and benchmark comparisons that RL practitioners can evaluate.

MarkTechPostOct 22, 2026Score 83

PerplexityNVIDIAReinforcement Learning

Tutorial Editor selected

AWS details production architecture for its open-source Deep Insight multi-agent system

AWS Solutions Architecture published the first part of a series on Deep Insight, an open-source multi-agent system that turns CSV data and natural-language prompts into validated analysis reports. The post describes eight role-specialized agents (Coordinator, Planner, Supervisor, Coder, Validator, Reporter, Tracker, Plan Reviewer) built with the Strands Agents SDK Graph and Agents-as-Tools patterns, with model selection tiered across Claude Haiku, Opus and Sonnet via Amazon Bedrock. It also outl

Why it matters: It offers concrete, reusable architectural guidance and trade-offs for teams moving multi-agent systems from prototype to production, backed by open-source code.

AWS Machine Learning BlogOct 22, 2026Score 80

AWSmulti-agent systemsagent orchestration

Models Editor selected

Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model

Recursive Superintelligence released Flint, a text-only mixture-of-experts coding model with 39.8B total and roughly 2.5B active parameters, under Apache 2.0 on Hugging Face. It targets coding agents, supports reasoning and OpenAI-compatible tool calls, and runs in about 23 GB at 4-bit on 48 GB Apple Silicon Macs or NVIDIA GPUs. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, trailing the strongest dense models on the latter. The company says its Eureka Machine aut

Why it matters: It gives readers concrete details on a new open-weight local coding model, including architecture, hardware needs and self-reported benchmark results.

MarkTechPostOct 22, 2026Score 83

Recursive SuperintelligenceFlintopen-weight coding models

Research Editor selected

Microsoft researchers, with AI-guided search, prove the KLS conjecture

Microsoft Research's Xinzhi Zhang and co-author Zhao Song proved the Kannan–Lovász–Simonovits conjecture, a 31-year-old central problem in high-dimensional geometry. The proof establishes a dimension-free bound showing flat cuts capture log-concave distributions' narrowest bottlenecks within a universal constant. The authors used a human-directed agentic process, exploring roughly 100 research trajectories before the key criterion emerged, while unguided AI attempts fell short. Independent paral

Why it matters: It documents a landmark math result achieved through guided human-AI collaboration, offering concrete evidence on how agentic AI assists scientific discovery.

Microsoft ResearchOct 22, 2026Score 90

Microsoft ResearchKLS conjectureAI for science