Search Practical AI

Search everything this publication has kept. The same list is available as JSON.

• 17 Sep 2026Ollama 0.30 adds GGUF support through llama.cpp alongside its MLX engineOllama 0.30 is out, with the company saying it improves performance and adds GGUF model compatibility through llama.cpp. The evidence is a short annou
• 17 Sep 2026Ollama updates MLX engine for Apple Silicon with faster, leaner local models claimedOllama has updated its MLX engine and says it now delivers its highest performance on Apple Silicon yet, with higher-quality responses, faster output
• 17 Sep 2026Ollama 0.31 speeds up Gemma 4 on Apple Silicon with MLX multi-token predictionOllama says Gemma 4 runs up to 90% faster on Apple Silicon in version 0.31 using MLX-backed multi-token prediction, with the figure scoped to coding a
• 17 Sep 2026Mistral's Robostral Navigate claims single-camera robot navigation at 76.6% on R2R-CEMistral has introduced Robostral Navigate, an 8B model that moves a robot from RGB images and a plain-language instruction using one ordinary camera a
• 17 Sep 2026Telecom fraud benchmark swaps easy negatives for near-domain calls, and classifier scores fall to 0.65-0.68 Macro-F1TeleAntiFraud 2.0 is an audio-based telecom fraud detection benchmark built from a Mixed-Tree Anti-Fraud Generation Pipeline and evaluated under a mon
• 17 Sep 2026A governed pipeline cuts guideline-to-artifact work from days to hours — and lowers hallucinationA paper describes SAGE, a multi-stage LLM pipeline that converts enterprise guideline documents into structured work artifacts. The evidence is a pape
• 17 Sep 2026NeMo Data Designer builds multimodal synthetic datasets from declarative column specs and a preview loopAn arXiv abstract describes NeMo Data Designer, an open-source framework for generating multimodal synthetic data by declaring each dataset column rat
• 17 Sep 2026Masking improved compositional generalization in preregistered test of sixty four-cell systemsA preregistered study with sixty four-cell systems sharing a frozen language-model backbone found that restricting what a module can read improved acc
• 17 Sep 2026GraphEcho tests whether graph agents count repeated paths as new evidenceA new arXiv benchmark holds evidence content fixed while varying how many paths an agent walks and where that evidence originated, and reports that re
• 17 Sep 2026Captions outperform direct video QA on long egocentric video, CapMem benchmark findsA new human-annotated benchmark, CapMem, tests whether text captions can serve as reusable episodic memory for wearable assistants. The reported gains
• 17 Sep 2026GVD links document versions and resolves rule conflicts locally, without an LLMA new paper describes GVD, a pipeline that treats versioning, duplicate detection and contradiction detection as one cross-document problem rather tha
• 17 Sep 2026EvolveTrade rewrites a trading agent's system prompt from its own decision traces, leaving the model fixedA new arXiv preprint describes EvolveTrade, a framework that revises a tool-using trading agent's system prompt after each update interval using
• 17 Sep 2026Mistral Studio adds versioned, traceable system of record for prompts and skillsMistral's Studio now offers a system of record for AI prompts and skills, versioned, owned, and traceable, according to a short announcement that
• 17 Sep 2026Inspect Evals v0.14.4 fixes scorer grade-binding and setup behavior in four evaluation suitesA point release of the UK AI Safety Institute's inspect_evals package corrects scoring and setup behavior in four evaluations; the release notes
• 17 Sep 2026METR proposes an 'expenditure horizon' to measure an AI agent's optimization abilityMETR proposes a measure of an AI agent's optimization ability, illustrated with the NanoGPT speedrun. The source is a short method note rather th
• 17 Sep 2026Mistral's Shieldstral puts moderation policy in the prompt, not the model weightsMistral has introduced Shieldstral, a 3B open-weights multimodal safety classifier released under Apache 2.0 that takes plain-language policies at inf
• 17 Sep 2026Meta's first open model from Superintelligence Labs lands on Ollama: Muse Glimmer, a 30B multimodal build for local coding agentsMuse Glimmer, described as the first open model from Meta Superintelligence Labs, is now available through Ollama under Apache 2.0. The evidence is a
• 17 Sep 2026Anthropic signs deal to lease site of its first Australian datacentre in western QueenslandAnthropic, the company behind the Claude large language model, has done a deal to lease the site of its first Australian datacentre, in western Queens
• 17 Sep 2026OpenAI opens a misalignment incident log, with six training-time cases disclosedOpenAI has announced a framework for tracking and publishing misalignment cases found during model development, and released six incident reports cove
• 17 Sep 2026OpenAI sketches an advertising push around Sponsored Agents and HubSpot, Shopify integrationsOpenAI has published a page describing new AI-powered advertising experiences, including something it calls Sponsored Agents, plus tools for marketers
• 17 Sep 2026LlamaIndex v0.14.14 adds a token-budget callback and LangChain 1.x supportThe v0.14.14 release notes for LlamaIndex list llama-index-core at 0.14.14 with a new TokenBudgetHandler callback filed under cost governance, plus La
• 17 Sep 2026LlamaIndex 0.14.15 adds multimodal core support and AgentMesh trust layerThe release, dated 2026-02-18, brings multimodal types and prompt templates to llama-index-core, and introduces llama-index-agent-agentmesh 0.1.0 as a
• 17 Sep 2026LlamaIndex 0.14.16 adds built-in rate limiters for LLM and embedding callsThe 2026-03-10 release notes for llama-index-core 0.14.16 list two new rate-limiting features alongside smaller fixes. The notes describe the features
• 17 Sep 2026n8n Gateway credits let Cloud users call AI models without a provider account or API keyStarting with n8n v2.36, select n8n Cloud plans can use supported AI models and tool services without creating a provider account or generating an API
• 17 Sep 2026LlamaIndex 0.14.24: parser, ingestion and memory fixes land in coreLlamaIndex published release notes for v0.14.24 dated 2026-08-19, led by fixes to core splitting, ingestion and tool handling plus a handful of small
• 17 Sep 2026Mistral's Agentic Search moves retrieval from one-shot chunks to a multi-step loopMistral introduced Agentic Search on 20 August 2026, a retrieval layer it says lets AI systems navigate, read and verify information inside complex do
• 17 Sep 2026Anthropic merges Claude chat and Cowork, and adds Docs and Slides for document outputAnthropic has integrated Claude's chat with its working environment, Claude Cowork, and added document and slide creation, according to a report
• 16 Sep 2026Sentence Transformers v6.0 adds training for multi-vector, ColBERT-style retrieval modelsThe Python library for embedding and reranker models now has a fourth model type, MultiVectorEncoder, with a full training path for finetuning or buil
• 16 Sep 2026LangChain puts Managed Deep Agents and its LLM Gateway into public betaLangChain's August 2026 newsletter lists Managed Deep Agents and LLM Gateway as reaching public beta, alongside Deep Agents v0.7, Tuned Evaluator
• 16 Sep 2026Cloudflare gives bot operators a dashboard home with submission tracking, editing and a behavior modelCloudflare's BotBase for Operators now lets bot and agent operators manage directory submissions from a dashboard. The source gives no pricing, r
• 16 Sep 2026A design-system CLI for coding agents: cheaper tokens, 8% faster tasks, wider tool reachAtlassian built a command-line tool that feeds AI coding agents design-system context from the terminal, and says it cut token costs, sped up tasks by
• 16 Sep 2026Hugging Face ships a WebGPU kernel loader and 207 versioned kernels for browser inferenceHugging Face has released @huggingface/kernels, a small JavaScript library that loads and runs optimised WebGPU kernels from the Hub, alongside an ini
• 16 Sep 2026BenchMIRT splits a single benchmark score into the capabilities actually driving itAI2 has published BenchMIRT, a method for auditing LLM benchmarks prompt by prompt rather than at the aggregate score. The evidence is a lab blog post
• 16 Sep 2026A coding model taught to paint: open RL recipe puts the reward in a hand-rated reference poolA Hugging Face blog post reproduces, with every artifact published, a viral demonstration in which a language model paints watercolours by writing Jav
• 16 Sep 2026inspect_evals v0.14.3: SWE-bench runs that miss a marker now error instead of scoring 0.0The inspect_evals v0.14.3 release reworks how SWE-bench Verified results are graded and how failed runs are classified, plus smaller changes to three
• 16 Sep 2026funes gives coding agents a memory layer you host yourselfA post on Hugging Face's site describes funes, a single-binary memory layer that indexes the coding-agent sessions already on your machine and ca
• 16 Sep 2026NeoMME ships a 260M and 800M multimodal encoder, with a retriever the team says cuts late-interaction index storage 255×H Company has released NeoMME, a family of 260M and 800M multilingual multimodal encoders trained from scratch without a separate vision tower or a ca
• 16 Sep 2026GPT-6 Astra claims state-of-the-art computer use and coding, with token prices up 2.5xOpenAI's GPT-6 Astra is described as a new frontier model class with new state-of-the-art computer use and coding, priced 2.5x higher per token b
• 16 Sep 2026SAGE scores literary quality in six layers, and finds the interpretive gap is roughly double the emotional oneA paper describes SAGE, a six-layer framework that splits rule-based checks of observable text properties from LLM-judged interpretive qualities. The
• 16 Sep 2026A Benchmark That Scores Markdown Format Failures Separately From ContentLatentMD is a new benchmark and CLI for diagnosing CommonMark fence-boundary failures in LLM-generated Markdown, and its authors report that 38.0% of
• 16 Sep 2026SQLMorph mutates text-to-SQL queries to expose join and wording failures, and scores partial answersA new arXiv paper describes SQLMorph, a framework that generates harder text-to-SQL evaluation sets by mutating queries and adds execution-level preci
• 16 Sep 2026Mistral describes a 40,000-line Fortran 77 to C++ migration built around a numerical parity harnessMistral has published an account of helping a European energy operator migrate 40,000 lines of Fortran 77 to C++, using AI agents for documentation an
• 16 Sep 2026Cloudera and Mistral partnership targets on-prem and air-gapped model deploymentMistral and Cloudera say Mistral models will run inside Cloudera's hybrid data platform, including in fully air-gapped environments. The announce
• 16 Sep 2026Atlassian says it is bringing governed agent loops to its AI-native SDLCAtlassian has announced governed agent loops for what it calls the AI-Native SDLC, supported by a single statistic from its own 2026 study. The announ
• 16 Sep 2026AWS pairs agent quality scoring with autonomous infrastructure investigation for production multi-agent systemsAn AWS blog post describes a dual-layer monitoring approach for production agents — Amazon Bedrock AgentCore Evaluations for continuous quality scorin
• 16 Sep 2026Perplexity runs GPT-6 Astra across systems work and checks in less oftenOpenAI's account says Perplexity uses GPT-6 Astra to write communications, change software and monitor production systems, and checks in much les
• 16 Sep 2026Fyxer's AI assistant leans on fine-tuning, memory and user feedback to sound like its ownerA vendor account describes Fyxer, an AI executive assistant that organises inboxes and drafts email in each user's voice. The description covers
• 16 Sep 2026AWS post outlines unattended replenishment loop with Databricks Genie and Amazon QuickAn AWS Machine Learning blog post describes a closed detect-decide-act loop that reconciles demand surges against live supplier availability and place
• 16 Sep 2026Amazon Bedrock AgentCore adds a consent portal for end-user OAuth with GitHub and Slack targetsAmazon Bedrock AgentCore Identity now offers a Consent portal, a managed web experience and a session binding endpoint for AgentCore Gateway. The acco
• 16 Sep 2026Abnormal AI runs Bedrock AgentCore Code Interpreter as an agent scratch pad for email threat detectionAbnormal AI has described deploying Amazon Bedrock AgentCore Code Interpreter as an ephemeral compute scratch pad for the agents behind its real-time
• 16 Sep 2026LangChain guide collects production agent lessons from Schneider Electric, Vodafone and monday.comLangChain has published a guide on scaling agents in Europe and the Middle East, drawn from how Schneider Electric, Vodafone and monday.com approach p
• 16 Sep 2026Claude Fable 5.1 returns to Copilot after upstream provider outageDegraded availability for Claude Fable 5.1 across some Copilot products and integrated development environments was resolved on Sep 15, with the statu
• 16 Sep 2026Copilot budget increase requests are now generally availableGitHub has made Copilot budget increase requests generally available, so members who exhaust their Copilot AI credits are no longer simply blocked. Th
• 16 Sep 2026An agent that passed 77.4% of runs passed all five on only 53% of tasks. IBM's new diagnostic targets that gap.IBM Research published an article on Hugging Face introducing "consistency guidelines" in ALTK-Evolve and a diagnostic called the Consistenc
• 16 Sep 2026llama.cpp b10996 forces a newline at the end of the reasoning budget for qwen3-coder in chatThe llama.cpp release b10996 carries one behavioural change: chat now forces a \n at the end of the reasoning budget for qwen3-coder. The release note
• 16 Sep 2026llama.cpp b10997 extends MoE tile heuristic to RDNA3.5 on ROCmRelease b10997 widens the HIP MoE ncols_opt tile condition so that RDNA3.5, not just RDNA3.0, is included, with reported matrix-multiply gains measure
• 16 Sep 2026Mozilla's Firefox Smart Window beta runs on Mistral models, starting in France and North AmericaMistral and Mozilla have announced a partnership that puts Mistral models inside Firefox Smart Window, described as Mozilla's AI browsing assista
• 16 Sep 2026GitHub AI Scan for pull requests no longer requires CodeQL default setupCode scanning's AI Scan can now be used on pull requests to find security vulnerabilities even when CodeQL default setup is not enabled on a repo
• 16 Sep 2026llama.cpp b11009 fixes tensor-parallel split states for fused QKV in gemma4 and qwen35Build b11009 of llama.cpp carries a tensor-parallel fix for fused QKV in gemma4 and qwen35. The release notes give one short changelog line plus a mod
• 16 Sep 2026TypeSafe AI releases Jev, a model that returns classifications instead of textTypeSafe AI, founded by former OpenAI researcher Diogo Almeida, is releasing a model called Jev that returns pure classifications for software rather
• 16 Sep 2026AgentCore's system prompt optimizer proposes changes from production traces, then validates them before promotionAWS published a technical companion to its AgentCore launch post explaining how the system prompt optimizer's reflector engine works and sharing
• 16 Sep 2026BearQ brings an assignable testing agent into Jira workflowsAtlassian's Inside Atlassian blog introduces BearQ for Jira, describing an assignable agent placed inside the Jira workflow for autonomous testin
• 16 Sep 2026smartARM's bionic arm selects grips automatically using Meta's open-source DINOv2 vision modelA Toronto startup is prototyping a bionic arm that recognizes objects from a few reference photos and picks a grip on its own, instead of asking the w
• 16 Sep 2026Spain's AEPD receives first report of an AI-agent data breachThe Spanish Data Protection Agency was notified of an attack allegedly carried out with an AI agent powered by a known large language model — the agen
• 16 Sep 2026Benchmark Radar indexes 1,283 benchmark records and ships a CLI for prior-art searchesA paper describes a "living database and search engine" for AI benchmarks, combining daily discovery with a searchable catalog, score histor
• 16 Sep 2026Attackers are routing malware through shared AI artifacts, conversations and search adsThreat actors are treating trusted AI platforms as an attack surface — hosting malicious content, poisoning search results and steering users into mal
• 16 Sep 2026vLLM v0.29.0 makes Model Runner V2 the default and adds per-request speculative decoding metricsvLLM's v0.29.0 release completes the Model Runner V2 rollout that began with pooling models, and adds a flag that reports speculative decoding ac
• 16 Sep 2026AWS publishes 38 open-source agent skills aimed at HCLS reasoning errorsAn AWS machine learning blog post introduces 38 open-source agent skills across 11 healthcare and life sciences domains, aimed at agents that cite the
• 16 Sep 2026Security guide flags exposed API keys as a common way vibe-coded apps failA guide to vibe-coding security opens by naming leaked API keys as a frequent failure of newly shipped apps, with a $4,000 bill as its worked example.
• 16 Sep 2026New reporting framework separates implementation from evaluation in AI-assisted systematic reviewsA new framework called PRISMA-LLM has been proposed to improve reporting for AI-assisted systematic reviews. It separates implementation disclosure fr
• 16 Sep 2026Zapier highlights AI model flexibility to avoid lock-inZapier published a blog post on AI model flexibility, but the available excerpt is brief. The post argues that different AI models suit different task
• 16 Sep 2026AI agents add decision-making to business automationA Zapier blog post describes AI agents for business automation, but does not announce a specific new tool or price. Its useful detail is that agents c
• 16 Sep 2026Open-source harness measures cost per correct answer for OpenAI models on Amazon BedrockAn AWS blog post describes an open-source benchmarking harness for OpenAI models on Amazon Bedrock. The post claims the harness measures cost per corr
• 16 Sep 2026Shopify CLI 4.8 adds terminal commands for dev store managementShopify CLI 4.8 introduces shopify store commands to create, list, and delete development stores from your terminal or coding agent. The update also s
• 16 Sep 2026llama.cpp fixes Metal NaN for activations beyond f16 rangeA Metal bug in llama.cpp turned entire vocabularies into NaN for models whose activations exceed f16 range. The fix targets the mul_mm_id path, and a
• 16 Sep 2026Small correction module fixes some language model errors without degrading base skillsA new preprint describes a lightweight module that corrects errors in a frozen language model's outputs while preserving its base capabilities. I
• 15 Sep 2026Claude Code adds forking of remote-control sessions as background jobsClaude Code v2.1.273 adds the ability to fork a session started with claude --remote-control or /remote-control from the Claude app, running the fork
• 15 Sep 2026AI Trained on Railroad Game Improved at Financial Research, But Only With the Right Training DesignGood Start Labs trained an AI on a railroad game, and one version improved at financial research. The source provides few details, but the key differe
• 15 Sep 2026China's Spy Chief Flags US AI Models for Faster Cyber Vulnerability DiscoveryChina's spy chief identified Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber as signs of a 'disruptive upgrade' in cyber
• 15 Sep 2026Mistral 7B: a 7B model that outperforms Llama 2 13B, released under Apache 2.0Mistral AI has released Mistral 7B, a 7.3B parameter language model under the Apache 2.0 license. The company says it outperforms Llama 2 13B on all b
• 15 Sep 2026Llama 3.2 comes to Ollama with small local text models and vision models on the wayMeta's Llama 3.2 is now available to run through Ollama, adding 1B and 3B text-only models built for local mobile or edge use, with 11B and 90B v
• 15 Sep 2026OFASys Enables Multitask Learning with One Line of CodeA new system called OFASys aims to simplify multimodal multitask learning by reducing setup to a single line of code, addressing difficulties the team
• 15 Sep 2026Bespoke-Minicheck, a new factuality checking model, arrives in OllamaBespoke Labs has released Bespoke-Minicheck, a grounded factuality checking model now available in Ollama. It can fact-check responses from other mode
• 15 Sep 2026Guide Introduces 'Critique Shadowing' for LLM EvaluationThe guide 'Using LLM-as-a-Judge For Evaluation: A Complete Guide' describes a technique called Critique Shadowing, which aims to help teams
• 15 Sep 2026Llama 3.2 Vision Runs Locally in Ollama, Adding Image Analysis to Your MachineOllama now offers Llama 3.2 Vision in 11B and 90B sizes, letting you run image recognition tasks like handwriting OCR and chart reading on your own ha
• 15 Sep 2026Minions: a method for local and cloud LLM collaborationA Stanford Hazy Research lab team and collaborators have developed a way to shift a substantial portion of LLM workloads to consumer devices, though t
• 15 Sep 2026AI Evals FAQ Draws a Line Between Model Benchmarks and Product EvalsA curated FAQ on AI evals collects common questions from teaching 700+ engineers and PMs, and distinguishes model benchmarks from product evals.
• 15 Sep 2026Qwen3-Coder: New Agentic Code Model with 1M Token ContextQwen has announced Qwen3-Coder, a new agentic code model. Its most powerful variant, Qwen3-Coder-480B-A35B-Instruct, is a 480B-parameter Mixture-of-Ex
• 15 Sep 2026Qwen-Image-Edit adds precise text editing to image editingQwen-Image-Edit is an image editing version of Qwen-Image, built on the 20B Qwen-Image model. It extends text rendering capabilities to editing and co
• 15 Sep 2026Ollama's new scheduler targets out-of-memory crashes and GPU utilizationOllama has rolled out a new model scheduling system that it says reduces out-of-memory crashes and improves GPU performance, especially on multi-GPU m
• 15 Sep 2026MiniMax M2 arrives on Ollama’s cloud for coding and agentic workflowsMiniMax M2 is now available on Ollama’s cloud, positioned by the source as a model built for coding and agentic workflows. The source does not detail
• 15 Sep 2026Gemini 3 Flash promises frontier performance at lower cost and global accessGoogle's DeepMind blog says Gemini 3 Flash offers frontier intelligence built for speed at a fraction of the cost, and the announcement headline
• 15 Sep 2026Three Years from GPT-3 to Gemini 3: The Move from Chatbots to AgentsA new piece by Ethan Mollick traces three years from GPT-3 to Gemini 3, framing the change as a shift from chatbots to agents. The evidence available
• 15 Sep 2026Gemini app adds Lyria 3 music generation for 30-second tracksThe Gemini app now features Lyria 3, a music generation model that the source says can make 30-second tracks from text or images. The evidence does no
• 15 Sep 2026Gemini 3.1 Flash TTS Adds Granular Audio Tags for Precise Speech ControlGoogle DeepMind has released Gemini 3.1 Flash TTS, a new text-to-speech model that introduces granular audio tags for precise control over AI speech.
• 15 Sep 2026Claude Code v2.1.267 Adds Effort Cap and Fresh System Prompt OptionThe release introduces a maxEffortLevel setting that caps effort on every provider, and a flag to render the system prompt fresh on each request.
• 15 Sep 2026Claude Code v2.1.268 adds gateway pricing setting for consistent cost trackingClaude Code v2.1.268 was released with a new gateway pricing setting. The setting lets signed-in clients receive the same rates through managed settin
• 15 Sep 2026Claude Code v2.1.271 adds per-command domain allowlists for sandboxed commandsAnthropic released Claude Code v2.1.271, which adds per-command allowed_domains to Bash, PowerShell and Monitor in auto mode with sandboxing, among ot
• 15 Sep 2026OdoBot cuts web-agent token use by 44-80% on a 45-task testA new web-agent architecture called OdoBot completes tasks with fewer tokens than two state-of-the-art competitors, according to a paper on arXiv. It
• 15 Sep 2026Judge feedback narrows the gap between cheap and costly patent-drafting modelsA new study tests an LLM judge for evaluating and revising AI-generated patent drafts. It finds that judge-guided revision improves quality and can he
• 14 Sep 2026GitHub Copilot auto model selection adds efficiency, balance, and intelligence tiersGitHub Copilot auto model selection now offers three tiers: efficiency, balance, and intelligence. The tier you choose reflects how you want auto to w
• 14 Sep 2026Competence gate improves hybrid forecasting by weighting language models per domainA new arXiv preprint describes a method that estimates when a language model adds value beyond an existing forecast. The evidence is a single preprint
• 12 Sep 2026Autonomous AI Agents: What They Can Do and How to Manage the RisksAutonomous AI agents are systems that pursue goals with little to no human intervention, planning and acting across tools. They differ from regular au
• 12 Sep 2026OpenRouter's GPT Sol Latest Alias Always Points to the Newest GPT Sol ModelOpenRouter has listed a model entry called OpenAI: GPT Sol Latest, which the listing says always redirects to the latest model in the OpenAI GPT Sol f
• 12 Sep 2026Agent can study an unfamiliar environment before test time, paper findsA new arXiv paper formalizes task-agnostic environment preprocessing: an agent explores an unfamiliar environment under a budget, before test time, an
• 12 Sep 2026Nemotron specialists reach IMO gold threshold with open natural-language pipelineTwo Nemotron 3 Ultra specialist checkpoints, trained with supervised fine-tuning and reinforcement learning, scored 30 out of 42 points at IMO 2026, t
• 12 Sep 2026Program-solve interface improves clinical math for 32B model, not 7BA preprint tests whether a language model can write case-specific Python for a restricted executor instead of doing clinical arithmetic itself. The ha