Mattt introduces Swift packages for speech and typed decisions
AnySpeechRecognizer swaps engines; AnyDecisionModel returns typed answers with probabilities.
Mattt AnySpeechRecognizer and AnyDecisionModel announcementRead brief
TuliAI / News Lab
AI news. Useful context.
Independent, AI-assisted briefs on models, agents and compute. Primary sources, practical context and clearly labelled analysis.
Google brings file editing directly into Gemini chats
Google announces in-chat creation, preview and direct editing.
Google Gemini in-chat file editing announcementRead brief
Liquid AI releases LFM3 as a specialized multimodal family
Liquid AI presents six variants built on a shared backbone.
Liquid AI LFM3 team announcementRead brief
Google previews Connected Apps distribution for Search and Gemini
Developers use cloud MCP or Android AppFunctions to expose app capabilities.
Google Connected Apps Platform preview announcementRead brief
Microsoft announces more languages for Azure LLM Speech
Microsoft lists 25 additional languages, including 11 Indic, for 09-2026.
Microsoft Foundry LLM Speech 09-2026 announcementRead brief
OpenAI releases GPT-6.1 Terra for Work, Codex and API
OpenAI lists Work and Codex access for Plus, Pro, Enterprise and Edu.
OpenAI GPT-6.1 Terra announcementRead brief
Hugging Face recaps shared caching and Jobs improvements
The October 26 post covers features shipped over recent weeks.
Hugging Face CLI and hub-library roundupRead brief
Ollama adds local Phonon speech with different model licences
Ollama says version 0.40.11+ runs Gradium’s speech models on Apple Silicon Macs.
Ollama Phonon speech-model announcementRead brief
OpenAI details containment changes for internal frontier training
OpenAI describes cached web resources and layered isolation for internal research workloads.
OpenAI internal frontier-training containment reportRead brief
Kimi rolls out K3.1 Pro in its hosted Code service
October 23 notes put Pro on Plus/Moderato and higher plans.
Kimi Code K3.1 Pro release notesRead brief
Microsoft adds interactive chart specifications to Flint
Its Flint-interactive language maps semantic events to updates across charts, tables and widgets.
Microsoft Research Flint-interactive language announcementRead brief
Anthropic turns selected fuzz tests into machine-checked proofs
With Theorem, Anthropic reports Claude-generated proofs about compiled RISC-V test entry points.
Anthropic formal software verification researchRead brief
OpenAI makes its hosted Decisions API generally available
GPT-6 Luna handles bounded questions from text, images or native JSON.
OpenAI Decisions API general availability announcementRead brief
IBM combines existing Granite forecasters in an ensemble recipe
The four-model recipe produces medians and quantiles without retraining or fitted ensemble weights.
IBM Research Granite ensemble forecasting recipeRead brief
OpenAI describes coordinated PHP vulnerability disclosure
Calif used GPT-5.6 Sol; Wiz also reported independent discoveries, according to OpenAI.
OpenAI coordinated PHP vulnerability disclosureRead brief
Ai2 tests demographic decision stability with FairStress
FairStress pairs demographic variants with matched persuasive prefixes to test whether decisions change.
Ai2 FairStress research announcementRead brief
Hugging Face outlines native TorchNeuron support for Trainium
Its October 19 post announces a public beta planned from October 20.
Hugging Face TorchNeuron integration announcementRead brief
Google explains MCP Apps deployment in Gemini Enterprise
Its October 19 guide connects sandboxed iframe interfaces to developer-hosted Cloud Run servers.
Google Cloud MCP Apps implementation guideRead brief
Anthropic outlines conditional equity for its Public Benefit Fund
Anthropic announces a commitment of 5% of its shares, conditional on stockholder approval.
Anthropic Public Benefit Fund announcementRead brief
DeepSeek updates Flash behind the existing API alias
October 17’s update reaches web, app and API as DeepSeek-V4.1-Flash-1017.
DeepSeek official API change logRead brief
BeforeShare case study: inspect a file before sharing it
TuliAI builds and maintains BeforeShare. This studio case study explains its inspect-first workflow, published synthetic tests and practical limits. It is not an independent product review.
BeforeShare: published cleanup studyRead brief
Google prototypes lab agents with a live digital twin
VERID links laboratory state tracking, simulated dry runs and command checks.
Google Research wet-lab agent prototypesRead brief
Anthropic tests quantum circuit designs in an evaluator loop
Claude proposed circuits for eight computational primitives, with programmatic correctness checks and resource scoring.
Anthropic quantum circuit research reportRead brief
Reflection releases Beam weights for self-hosted agents
The October 16 release follows the earlier model announcement; weights use Apache 2.0.
Reflection Beam availability announcementRead brief
Mistral introduces CPU-capable Timestral forecasting models
Mistral’s October 15 announcement describes four sizes for probabilistic forecasts of related time series.
Mistral Timestral announcementRead brief
NVIDIA and Siemens assess chip-design agents through final files
An October 14 case study uses NeMo Gym to compare generated Liberty files with references.
NVIDIA Technical Blog Siemens NeMo Gym case studyRead brief
Google opens Deep Review for technical document critique
Google’s October 14 announcement extends its document-review system to public use.
Google Research Deep Review announcementRead brief
NVIDIA opens the full Run:ai orchestration platform
Its October 13 release covers GPU-workload management, with independently usable scheduling, sharing, quota and workload-description components.
NVIDIA Technical Blog release announcementRead brief
Google expands Gemini 4 Argon access in its app
Google AI Pro and Ultra subscribers can use Argon and give Gemini tasks.
Google Gemini app availability announcementRead brief
Meta introduces Muse Wonder with separate access routes
Meta attributes persistent context and sustained agent work to Active Memory and Active Focus.
Meta AI Muse Wonder announcementRead brief
Liquid AI releases LFM2.5 for local tool-using agents
LFM2.5-24B-A2B and its Base version are available as open weights.
Liquid AI LFM2.5 release announcement on Hugging FaceRead brief
Microsoft explains StudentSim’s learner-specific tutoring simulations
StudentSim adapts pooled learner records to individuals for AI-tutor research.
Microsoft Research StudentSim explanationRead brief
LightOn delegates repository search to ColGrep-Agent
Its local, read-only search agent returns file paths, line ranges and code.
LightOn ColGrep-Agent announcement on Hugging FaceRead brief
Mistral previews managed hosting for Workflows workers
Its October 9 release notes describe deployments from GitHub repositories.
Mistral October 9 Managed Deployments release notesRead brief
Cloudflare’s Clef-omni adds audio and video decision inputs
Cloudflare’s October 9 announcement extends its open-weight decision models beyond text and images.
Cloudflare release announcementRead brief
Anthropic reports unintended actions in testing and internal use
Anthropic describes unintended form submissions, gated-data access and tool workarounds.
Anthropic alignment research reportRead brief
Ai2 explains GPU scheduling through time budgets
Its October report describes a rollout begun in July, using GPU-time budgets and hierarchical fair share.
Ai2 infrastructure report on Hugging FaceRead brief
LightOnOCR-3 adds page locations to document extraction
LightOn released variants labelled 0.8B, 1B and 4B under Apache 2.0.
LightOn release announcement on Hugging FaceRead brief
Anthropic opens applications for OSS Scanner
Eligible open-source maintainers can apply for free periodic vulnerability scans.
Anthropic Frontier Red TeamRead brief
OpenAI reports disrupting influence operations using false identities
OpenAI says it banned accounts linked to Russia- and Iran-origin influence operations.
OpenAI investigationRead brief
Microsoft makes Execution Containers generally available
Microsoft’s containment layer limits agent access through externally enforced policies.
Microsoft Windows Developer BlogRead brief
Liquid AI releases open-weight d1 decision models
The models return structured decisions instead of generating prose.
Liquid AI release announcement on Hugging FaceRead brief
LASER looks for more useful safety-evaluation examples
OpenAI describes a method for selecting examples to test model safety.
OpenAI AlignmentRead brief
GPT-6 brings interactive responses to ChatGPT
OpenAI introduces answers that can include charts, forms and interactive components.
OpenAIRead brief
Haiku 5.5 targets focused, high-volume AI work
Anthropic’s new small model is aimed at repetitive tasks and scoped agent work.
AnthropicRead brief