Laya
An Apache-2.0 System One model for classification, routing, scoring and guardrails.
- By
- ConvAI Innovations
- Access
- Not recorded
- License
- Not recorded
- Deployment
- Not recorded
103 entities · 33 projects · 23 tools · 9 models · 38 news
An Apache-2.0 System One model for classification, routing, scoring and guardrails.
Deterministic verification for AI output: it proposes, deterministic tools decide, every claim checked against ground truth.
An MIT-licensed desktop document-to-Markdown workbench with a local REST API, multiple extraction engines and original-page comparison.
A November 2026 practitioner debate about unintended edits, test integrity and the human review cost of inexpensive coding agents.
An Apache-2.0 framework for coding agents to build custom BI apps on reviewed models, metrics and a warehouse-backed server.
A public test-input export for multimodal agents making 3D mechanical assembly decisions, GUI actions and bounded minimal-rework choices.
A community training-free weight edit of Prism ML’s Bonsai 2 9B, distributed as GGUF variants with author-reported reductions in repeated reasoning.
Kling Team’s November 3 technical report explains how multimodal understanding connects training captions, prompt enhancement, text encoders, reward models and video evaluation.
A web recommendation index whose creator uses local Gemma inference to extract Reddit and Hacker News opinions, with source-linked excerpts and attributed ranking methods.
An alpha Python pipeline that compiles organizational policy into reviewed rules, reproducible scenarios and decision-model training data.
Hugging Face's proposed testing format combining small frozen model fixtures, executable Markdown usage paths and human-reviewed output baselines.
Apache-2.0 conformance fixtures and a verifier harness for checking records of AI-agent actions, approvals, completion and supply-chain evidence.
Business Insider reports that internal Claude consumption valued at customer prices exceeded monthly employee compensation excluding stock, a comparison distinct from actual cash spending.
Firmware simulation tools and a GitHub reviewer that let engineers and AI agents run embedded code against virtual microcontrollers.
Partnership on AI researchers argue that reliable agent oversight needs persistent identities, human-intervention records and traces of permission and memory changes.
A research evaluation package pairing multi-turn embodied-understanding questions with source images and stored rollout videos across ten task subsets.
A research framework that alternates active interaction with offline consolidation and synthetic rehearsal to retain knowledge in model parameters.
An open research project and command-line scanner measuring disclosed AI attribution in public Git history, with aggregates from a 96-repository study.
Surge AI's benchmark of 50 professional tax assignments, testing whether agents investigate records and deliver complete work under detailed rubrics.
Google DeepMind's frontier model announced October 28, with phased access for selected cyber defenders before broader product availability.
A robotics research framework that learns future physical features during training and removes the auxiliary video model at execution.
A research benchmark testing whether complete coding-agent systems can deliver rendered videos from natural-language briefs within one hour.
A Python security decision engine combining source-aware rules, SQL validation, PII checks and optional transformer classifiers.
A security scanner combining cross-file static analysis, a Z3-based reasoning engine and an optional language-model layer for detection, triage and fixes.
A standalone local desktop assistant that searches an offline emergency reference library using models run through Ollama.
Parents allege GPT-4o encouraged their daughter to stop epilepsy medication; TechCrunch reports a new October 27 lawsuit.
An announced application-security agent that combines deterministic scans, AI-assisted fixes and human approval, with a Claude Code plug-in for Coverity users.
A host-agnostic CLI that lets coding agents author and execute JSON mobile UI scenarios through mobile-mcp.
OpenAI's October 26 research announcement combines three scientific claims with public proof artifacts and different forms of external review.
Ai2 and University of Pennsylvania researchers describe a workflow that helps authors critique and revise ideas against selected papers.
Published traces describe a blocked network action; intent, prevalence and the meaning of continuation tests remain disputed.
Hosted human-approval service announced by iProov, combining action-bound credentials and genuine-human-presence checks with agentgateway and Cerbos integrations.
English synthetic dataset of matched tool-response triplets for studying indirect prompt injection and fictional recovery-token disclosure.
Future Chunk Prediction research with released Wan-based video generation checkpoints, teacher-forcing baselines, training code and demo data.
Oliver Grabner's Pokemon Showdown environment for studying private communication and coordination between two LLM teammates.
October 21 PHP security releases fix two critical engine vulnerabilities whose practical exploitation was established through AI-assisted research and coordinated disclosure.
A game matching-decompilation tracker whose maintainer disabled its public index and new project submissions while preserving existing project tracking and historical pages.
Dharma-AI's previewed method for adapting small-model vocabularies around inference cost, combining domain subwords and repeated multiword spans.
Agent plugins for reproducing bugs, proving scoped patches and checking regressions with independently replayed evidence.
A community GGUF conversion of Google's Gemma 4 12B QAT model, rebuilt as MXFP4 after its author identified an earlier mislabeled hybrid.
A joint protein-sequence and all-atom structure generation model from the University of Washington's Institute for Protein Design and NVIDIA GenAIR.
An MIT Claude Code boilerplate combining container execution, deterministic workflow generation, Bash command rules and test gates under human supervision.
A vendor-run survey benchmark comparing five bot detectors against scripted browsers, a stealth browser and three AI browsing agents.
Anthropic's October 20 study of Claude-assisted configuration and code reachability analysis across 25 open-source projects.
A Cosmos3-based research system that jointly forecasts multi-view videos, persistent 3D point trajectories and optionally robot action chunks.
Redact and WIRED report that WhatsApp help pages exposed support-chatbot instructions, including directions against recommending regulators; Meta says a bug was fixed.
A confidential Anthropic proposal reported by Politico would set catastrophic-risk thresholds for model suspension, require outside evaluations and preempt some state AI safety rules.
A 140M typed-decision model and calibration library that return probabilities with an automate or escalate flag.
A community 321M CPU-oriented classifier preview for prompt injection, tool-result inspection and output checks in English, Arabic and Persian.
A local browser console for wicked-crew that puts coding-agent jobs, approval questions, memory and evidence into one workspace.
DefuzeX's open-source harness for running agents in isolated cases, capturing execution evidence and evaluating behavioral defects.
TokenSave's published two-run comparison of Claude Code and Codex startup overhead, MCP tools and a browser-game coding task.
A Git-hook comprehension checkpoint that asks developers to predict the behavior of agent-assisted code before a protected push continues.
Infatoshi's community measurement bundle for a locally served, attention-edited DeepSeek V4.1 Flash 1017 checkpoint on four Blackwell workstation GPUs.
A community bundle of Prism ML's PQ2_0 Bonsai 2 model and a separate inference-time refusal-steering control vector.
Meta revised disputed commercial-terms language following concerns that it could authorize AI training on publishers' websites.
An October 16 internal study describes changing junior hiring shares, agent management and meeting time at Anthropic.
A community benchmark dataset comparing total agent-session prompt tokens and answer correctness with raw, compressed and truncated MCP tool outputs.
Mercor's provisional map of how 29 selected AI benchmarks cover tasks across the US economy.
Dualverse AI's public research community where independently operated agents discuss problems and submit papers for autonomous review.
A research method that trains agents across context compaction and subagent sessions using session-aware reinforcement learning.
A benchmark and practice platform for agents delivering working systems under ambiguous customer requirements.
A research method that explores temporary attention routes during reinforcement learning of image and video generators, then restores dense attention at inference.
NVIDIA and AI4Bharat's announced synthetic Indic pretraining corpus, paired with a public NeMo Data Designer translation and evaluation plugin.
A video-generation research method that accepts converged token drafts early and spends further denoising computation on the remaining tokens.
PostHog's public view of model and harness usage, tool calls and errors on its own MCP server, with explicit identification and sampling limitations.
A standalone orchestration harness that coordinates coding workers, independent acceptance and code reviewers, bounded repairs and optional Git delivery.
Anthropic's October 15 arrangement gives six METR employees internal access comparable to its risk teams and publication rights subject to defined redactions.
Community-maintained, MIT-licensed dataset and explorer for free AI API offers, with dated source links, constraints and documented provider data policies.
Geoffrey Huntley's example of a reusable Nix overlay that patches Git and supplies the same variant to a developer shell, container image and NixOS test.
Vendor-neutral OpenTelemetry references for coding agents, including dated records of the configuration schemas supported by specific SDK and agent releases.
A benchmark that asks models to build a coding-agent harness from documentation, then tests the frozen product with scripted model responses.
Beta services combining isolated coding-agent workspaces with continuous verification and signed test attestations.
Google DeepMind and Singapore GovTech developed an on-device model intended to warn users about manipulative scam conversations without sending message text to the cloud.
An open-source benchmark that measures the net syntax added by AI coding agents in published code patches.
A community-maintained Jev reproduction index corrected its same-skill scoring and added a public-versus-private comparison.
The Typed Decisions maintainers replaced a test-exposed model score with a clean re-score after a public challenge.
An independent Hugging Face model explorer plots download history and derivative families, with comparisons, rankings and author summaries.
Kore.ai's announced engine for evaluating, diagnosing and optimizing enterprise AI agents against business-defined goals.
A developer's English retrospective says autonomous coding agents repeatedly found ways around a byte-matching checker during a private game decompilation project.
A Python library for assessing and improving the fairness of machine-learning models; version 0.15.0 adds group-level plots and fixes weighted fairness calculations.
Berkeley highlights randomized trials linking brief AI assistance to reduced persistence and performance when assistance is withdrawn.
Anthropic opened a free opt-in service that periodically scans open-source projects for vulnerabilities and sends model-generated reports to maintainers.
Arena announced a $200 million Series B at a $3.1 billion valuation and previewed an index for observable AI-agent failures.
The Financial Times reported that OpenAI told investors its annualized revenue was approaching $50 billion, below a prior $70 billion media estimate.
OpenAI's mathematics repository records the withdrawal of three related manuscripts after a sign error invalidated a key argument and lists revisions to other papers.
Senator Maria Cantwell outlined six AI safety priorities that she says should guide future US legislation, including frontier-model testing and oversight.
Google opened a dedicated SynthID Detector site for wider watermark checks on supported images, video and audio, following an earlier limited test.
Meta Oversight Board members told NBC News that AI safety committees need real independence, internal access and authority to avoid becoming symbolic.
Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an unacceptable risk after testing parental alerts and crisis responses; OpenAI disputes the methodology.
GitHub released a public research preview of ReviewBench, an open benchmark and leaderboard for AI code review agents.
HackerRank made its Chakra AI interviewer generally available to customers after a beta, with a coding workspace and automated follow-up questions.
Google stopped accepting new product-vulnerability reports through its open-source reward program on October 1, citing a surge of mostly invalid automated submissions.
OpenAI plans to phase textGrain watermarks into eligible ChatGPT and Codex output in the EU, with optional API access worldwide and limited detector access.
Microsoft and Hugging Face published an October 3 analysis of ThinkingBox results on agent outcomes and consistency across repeated business workflows.
Ai2 released an open-weights 8B model for generating cited scientific reports and added it as Fast mode in Asta.
OpenAI says it disrupted a coordinated attempt to extract protected model reasoning and strengthened controls against adversarial distillation.
Google DeepMind introduced research methods for marking AI-generated protein sequences and predicted structures so their origin can be checked.
OpenAI formed an external group to review and communicate emerging mathematical results produced with AI.
TypeSafe AI's System One model for fast, typed, confidence-aware decisions.
Replaces LLM judges with typed Jev decisions for evaluation.
A short brief on a recurring agent architecture: one system plans, another acts, and deterministic checks verify the result.
An argument that prompts are not a stable object of study, at 114 points on Hacker News.