Skip to content
- Liability for failing AI: the revised PLDHow the revised Product Liability Directive (PLD) and the AI Act fundamentally change the burden of proof for damage caused by failing AI systems.
- Agentic AI: The Shift to Autonomous AI Systems | LLMNet NewsDiscover what Agentic AI entails, why autonomous AI agents are emerging now, and what the main opportunities and risks are for modern organizations.
- The AI Act in practice: what companies must do in 2026Practical guide to the European AI Act in 2026. Read which obligations, risk classifications, and compliance steps now apply to organizations.
- AI Act timeline: what took effect in 2026The AI Act: what actually came into force in 2026 and what was postponed by the Digital Omnibus? A factual review, verified on 2026-11-02.
- AI assistants in the browser: the impact on browsingHow are AI assistants in browsers and extensions changing our browsing behavior? Read all about the impact on search traffic, publishers, and the associated…
- AI at Dutch municipalities: from pilot to practiceHow Dutch municipalities are deploying AI and language models for service delivery, including GDPR requirements, the algorithm register, and the EU AI Act.
- AI and Copyright: Frameworks, Opportunities, and Pitfalls inWhat is the status of copyright and AI in the EU? Everything about the text and data mining exception, opt-outs, rights to output, and contractual agreements.
- The Impact of AI on the Creative SectorHow image, music, and video generation are transforming the creative sector. An in-depth analysis of the opportunities, the copyright debate, and new…
- AI and Cybersecurity: The Digital Arms Race of the 21stHow attackers and defenders deploy artificial intelligence. An in-depth look at AI-driven phishing, deepfakes, detection systems, and the impact on the…
- AI and energy: why computing power and power consumptionThe impact of AI on our energy consumption is growing rapidly. Discover why AI models require so much power, the latest efficiency trends, and the future of…
- AI in Libraries and Archives: ConsiderationsAn in-depth analysis of deploying artificial intelligence in public libraries and archives, opportunities, and strategic decisions.
- Digital sovereignty in Europe and dependence on AI - llmnet.nlAn in-depth analysis of digital sovereignty and AI dependence in Europe. Explore the AI stack, legal risks, and strategic choices.
- AI and work: which tasks are changing the fastest?Discover which tasks are changing and automating the fastest due to artificial intelligence, which remain human, and what this means for the labor market.
- AI in the Dutch financial sector: opportunities and requirementsHow financial institutions are deploying AI within existing frameworks of model governance, explainability, outsourcing risk, and supervision.
- AI Language Models in Dutch HealthcareA comprehensive and factual analysis of the deployment of AI language models in Dutch healthcare, focusing on reporting, MDR, AI regulation, and privacy.
- AI in education: how schools and universities are using itHow primary education, vocational education (mbo), universities of applied sciences (hbo), and universities are using AI in 2026: from adaptive learning and…
- AI in education: opportunities and concerns — LLMNet NewsDiscover the impact of artificial intelligence in education. A balanced look at personalized learning, academic integrity, and the redefined role of the…
- AI in industry and manufacturing companies: digitalizationAnalysis of AI applications in Dutch manufacturing industry: from predictive maintenance to vision systems, OT integration, and the EU AI Act.
- AI in the Dutch government - state of playHow is the Dutch government using AI as a user and buyer? Read all about transparency, the algorithm register, procurement and societal safeguards.
- AI in the Dutch judiciary: use and limitsDiscover how AI is used in the Dutch judiciary, what legal boundaries apply, and what the trade-offs are for judges and hearings.
- AI standards and ISO norms: what they do and do not doA factual analysis of AI standards and norms. Discover what process standards regulate, how harmonized standards work and what certification means.
- AI Supervision in the NetherlandsDiscover how AI supervision in the Netherlands is organized under the European AI Act. What role do the AP, RDI, and ACM play? Read the full analysis.
- AI for SME services: the practice of small firmsDiscover how small offices and SME service providers use AI for case processing, client contact, and quotes without getting lost in high costs.
- AI legislation outside the EU: an overviewSee how the US, the UK, China and other countries regulate AI outside the EU. Discover the differences in approach and what this means for organizations.
- Anthropic on agent memory: what 'dreaming' addsAnthropic describes agent memory in four layers plus an offline 'dreaming' step. With memory, Opus 5.5 rose from 0.655 to 0.838 on CL-Bench.
- Anthropic S-1/A: $17B Q3 revenue, $895B in commitmentsAnthropic's amended S-1 shows $16.96B Q3 revenue (+1,127% YoY), $895.4B in commitments and an RSP warning. What builders should know.
- ChatGPT Images 2.6: GIFs, references, Flare and SunburstChatGPT Images 2.6 adds GIFs and reference images and is 50% faster. In the API you choose Flare (fast) or Sunburst (consistent across edits).
- ChatGPT Pro and Business: up to $500 a month in API creditsEligible ChatGPT Pro and Business subscribers can claim up to $500 a month in API credits, depending on Pro tier or Business seats. How it works.
- The race for AI chips and computing powerWhy are AI chips so scarce and expensive? Read all about the GPU race, geopolitical tensions, ASML, Big Tech alternatives, and the impact on the Netherlands.
- Claude Code 2.1.319: MCP tasks move to the backgroundClaude Code 2.1.319 moves MCP tasks to the background, can deny instead of waiting on a permission prompt and closes a sandbox escape. What changes.
- Secrets in Claude Code settings: how to check your repo'Don't ask again' in Claude Code stores the full command, password included. What Apiiro found on GitHub and how to check your own repository today.
- Claude Fable 5.2 and Mythos 5.2: what changesFable 5.2 and Mythos 5.2 are one model with different access. Price $8/$40, 1M context, 30-day retention, and Anthropic advises starting with Opus 5.5.
- Claude for Legal gets a gallery of ready-to-use artifactsClaude for Legal adds ready-to-use artifacts such as Cite check and Tabular review, plus 60+ legal connectors. A product expansion, not a new model.
- Claude Meetings: Claude joins Google Meet callsAnthropic launched Claude Meetings in beta: Claude joins Google Meet calls as a participant, takes notes and builds documents during the call.
- Claude Personal Finance: read-only bank access (US)In the US, Claude can read your bank accounts through Plaid: balances, spending and transactions, without moving money. What is shared and what to check.
- Claude auto-connects connectors with Enterprise-managed authTeam and Enterprise orgs can now auto-connect connectors to Claude. IdP-managed auth is GA; Claude-issued is in beta and checks membership only, not SSO/MFA.
- Workers AI KV cache leak: GLM-5.3 exposed customer dataCloudflare says Workers AI model GLM-5.3 returned 155 responses with other customers' data from Aug 27 to Sep 9 on prompts over 681,280 tokens.
- Codex 0.169: --trust for exec, side chats, custom providersCodex rust-v0.169.0 keeps project config and MCP servers off in untrusted folders until you pass --trust, and adds side-chat tabs and custom model providers.
- Consolidation among AI start-ups: acquisitions and closuresAn analysis of the consolidation phase among AI startups: acquisition structures, margin pressure per stack layer, contract risks, and spreading business risk.
- Context caching: smart storage halves token costsDiscover how context caching works with LLM APIs, why KV caches structurally lower input costs, and how latency and architecture improve as a result.
- Context Mode 2.2 keeps tool output out of the context windowThe Context Mode plugin keeps coding-agent tool output in a local SQLite index and returns only what the agent needs. Version 2.2.0 supports eleven agents.
- CrowdStrike: Fiddler, Dreadnode and an AI Kill SwitchCrowdStrike plans to acquire Fiddler and Dreadnode and adds an AI Kill Switch: gateway first, endpoint soon. What it means for teams building AI agents.
- Cursor incident: a leaked key led to a fake extensionCursor explains how a key exposed at Fireworks led to stolen agent traces and a fake Claude Code extension in its marketplace cache on October 29.
- Data poisoning in LLM fine-tuningData poisoning is a structural risk in fine-tuning LLMs. Read how manipulation of training data works, how it can be detected, and how to prevent it.
- Datacenters and AI in the NetherlandsAn in-depth analysis of the debate on AI datacenters in the Netherlands: from power consumption and grid congestion to the call for European digital autonomy.
- Data Diversification: Training on Non-English SourcesHow multilingual data diversification improves language models and its impact on tokenizer efficiency, reasoning, and cultural representation.
- NPU chips in the workplace: local AI computing powerNPU chips move AI tasks from the data center to laptops. Read about architectures, power consumption, memory bandwidth, and practical deployability.
- Decentralized Compute Networks for AI Inference ExplainedDecentralized compute networks distribute AI inference across dispersed machines instead of hyperscale data centers. Mechanics, rise, and limits.
- Deepfakes and disinformation: the downside of generative AIDiscover how deepfakes and generative AI fuel disinformation. Read about the technology, risks, detection, and legislation in this in-depth background…
- DeepSeek API adds search, files, link reading and speechSince November 13 DeepSeek offers its app functions in the API on deepseek-flash: search, files, URL reading, OCR and speech. Search costs $0.75 per 1,000.
- Direct Preference Optimization: beyond classical RLHFDirect Preference Optimization (DPO) replaces complex RLHF pipelines. Discover the mathematics, stability gains, and practical trade-offs in LLM training.
- Energy and AI: the state of Dutch data centersAn in-depth analysis of the impact of AI on Dutch data centers, grid congestion, energy consumption, and the national infrastructure in 2026.
- Energy use of AI data centers: reading the figuresA thorough analysis of energy figures for AI data centers. Read how PUE, GPU consumption, cooling and grid congestion are measured.
- The EU AI Act at a glance: what changes for creators andA clear explanation of the European AI Act. Learn about the four risk categories, transparency requirements, and the timeline for creators and users of AI.
- Frontier Model Forum rates jailbreaks on cyber riskThe Frontier Model Forum publishes a framework that scores cyber jailbreaks on four points, from Low to Critical. Comments are open until December 21.
- Gemini 4 Carbon: Google's frontier model for cyber, agentsGemini 4 Carbon tops the AA Intelligence Index (59) at $2/$10, but goes to cyber defenders via Fairwind first. What it means for builders.
- Gemini 4 Hydrogen: Google's fast model at a flat priceGoogle released Gemini 4 Hydrogen on November 12: 1 million tokens of context at 1.50 and 7.50 dollars per million tokens across the whole window.
- Gemini app: create and edit files right in the chatThe Gemini app now lets you create, preview and edit documents, PDFs, slides and sheets in the chat. For all users; on mobile, PDF only.
- Gemini Omni 1.2 Flash: 15-second video, 30% cheaperGoogle's video model Gemini Omni 1.2 Flash makes clips up to 15 seconds, edits conversationally and is 30% cheaper, Google says. Prices and what is known.
- GitHub builds Git storage for thousands of agents at onceIn internal tests GitHub's new Git storage sustained 5,000 agents and 7,200 pushes per second on one repo. What changes and when it rolls out.
- 15 free Google AI tools you can use right now | LLMNet NewsGoogle gives away a surprising amount of AI tooling for free: from NotebookLM and AI Studio to Jules and Flow.
- GPT-6.1 Terra: cheaper in the API, mixed in practiceGPT-6.1 Terra costs $1/$5 per million tokens in the API. What changes for builders, and why is the first hands-on test mixed?
- Grok Build: 94.7% of shared apps open without loginClaroc found 35,578 Grok Build apps that opened without signing in, some with internal documents and personal data. What was found and what you can do.
- The algorithm register examined: what public bodies report1,530 algorithms from 336 organizations, over 60 percent municipalities, 41 high-risk systems: a repeatable measurement of the register, checked on 2026-08-07.
- The algorithm register examined: what governments reportA thorough analysis of the Dutch Algorithm Register in 2026: what governments register, where the blind spots lie, and how oversight works.
- AI and language models in Dutch greenhouse horticultureHow computer vision, local language models, and autonomous agents are transforming Dutch greenhouse horticulture: from climate control to harvest robotics.
- How Mixture of Experts lowers LLM production costsHow mixture of experts (MoE) lowers the compute and inference costs of LLMs in production without giving up model capacity.
- Inkling 2 Small: open weights not released yetThinking Machines promised Apache 2.0 weights for 'next week' on November 5. On November 10 that sentence is still there. What the model is and what works.
- The rise of small language models (SLMs)Discover why small language models (SLMs) are on the rise. Read about the benefits in terms of efficiency, on-device processing, cost, and privacy.
- How OpenAI and Anthropic detect AI model misuseOpenAI monitors frontier models in layers; Anthropic has Claude assess conversations offline. What the systems found and what it means for privacy.
- Lara Mini: a multilingual open base model, non-commercialTranslated puts Lara Mini on Hugging Face: a 7.2B-parameter MoE base model with 0.6B active, trained in 45 languages. The weights are non-commercial only.
- Open-weights vs truly open sourceDiscover the crucial difference between open-weights and truly open source AI models. Learn about licensing, user freedoms, and why this debate matters.
- llama.cpp v1.0.0: stable C API, FP8 cache and MoE cachellama.cpp and ggml reached version 1.0 on November 13: breaking API changes now get a major bump. Plus an FP8 KV cache, a MoE expert cache and a safer server.
- Meta brings Active Memory and Active Focus to MuseMeta rolls out models trained with Muse Wonder to Muse and Meta AI. According to Meta, Muse forgets critical information 2.4 times less often.
- Meta Muse Ember: cheap, with a tier above 200KMuse Ember costs $0.10/$0.50 per million tokens up to 200K context and three times as much above it. Specifications, access and Meta's own benchmarks.
- Mistral Large 4 (Le Chonk) goes open weights: key factsMistral releases Large 4 (Le Chonk) weights: ~1T parameters, MEOW licence, $0.70/$2.10. Best open model outside China per Artificial Analysis.
- Model weights leaks: the legal status of stolen AIWhat is the legal status of leaked or stolen model weights? Analysis of copyright, trade secrets, unauthorized computer access, and enforcement.
- Why Multimodal Models Are the Next Step | AI News & ResearchAn in-depth look at multimodal AI models: how the simultaneous processing of text, image, and audio enables unprecedented new applications and efficiency.
- Dutch AI Funding: Who Pays for InnovationWho funds Dutch AI innovation? An analysis of public subsidies, venture capital, ROMs, and compute costs for startups and scale-ups in 2026.
- Dutch AI in figures: public sources and dataDiscover how to read and interpret public figures from CBS, DNB and central government ICT on Dutch AI adoption and infrastructure critically.
- Dutch language evaluation: comparing models on Dutch tasksEvaluation of Dutch-language models: DUMB, EuroEval, BEIR-NL, and SQuAD-NL with source and verification date. Verified on 2026-08-07.
- Grid congestion and AI data centers: the state of the queueHow large is the electricity queue and what place do AI data centers occupy in it? Figures from grid operators and the ACM framework, checked on 2026-08-07.
- Open datasets for training language models | AI newsAn analysis of open data collections for language models, the importance of provenance, the challenges of filtering and the specific position of Dutch.
- Open models one year later: licensing since Llama 3What has changed since Llama 3? From community licenses to MIT and Apache 2.0, and what the AI Act exempts. Verified on 2026-08-07.
- The key developments in open-source LLMs | LLMnet.nl NewsDiscover the latest trends in open-source Large Language Models: from efficient small models and on-device AI to key shifts in licensing.
- Market Development: Open Versus Closed Models | llmnet.nlAnalysis of market structure, revenue models, and pricing pressure in open versus closed AI models. Which factors determine the choice for organizations?
- OpenAI pipeline found 948 Linux root bugs in one runAn OpenAI prototype with an unreleased cyber model found 948 unique Linux LPEs and wrote candidate patches. What is already fixed and what it means for you.
- OpenAI: new security sprint and bigger bug bountiesOpenAI plans a security sprint with partner Cathedral and raises bug bounties: critical up to $30,000, exceptional $100,000, up to $172,000 with bonuses.
- OpenClaw 2026.11.2: a patch with no new featuresOpenClaw 2026.11.2 is a patch release: it fixes stuck updates, missing reminders, duplicate widgets and a series of permission and log privacy issues.
- OpenClaw cut replies off at 8,192 tokensWithout maxTokens, a custom model in OpenClaw got an 8,192 output-token cap. Version 2026.11.3 raises that default to up to 32,768 tokens.
- Perplexity Decider v1.3: open decision models with locatePerplexity released Decider v1.3 (27B) and Flash v1.3 (4B) on November 12: open decision models under Apache 2.0, with the new locate feature.
- Prompt injection in RAG: the leak in corporate knowledge basesA technical analysis of indirect prompt injection in RAG: how corporate knowledge bases are vulnerable and which architectures protect data.
- Quantization in production: from 16-bit to 4-bit precisionPractical guide to LLM quantization in production. Discover the impact of FP16 to INT4 on VRAM, latency, perplexity, and inference architecture.
- Qwen4-Max-Preview: Token Plan only, until November 25Alibaba's Qwen4-Max-Preview has a 1 million token context and is available only through Token Plan until November 25. After that the model is retired.
- The race for context: why models are getting increasinglyDiscover why AI models are getting increasingly larger context windows, the impact on web applications, and the trade-off between long context and RAG…
- Reasoning Models: The Trend Towards AI That ThinksDiscover how AI reasoning models are raising the bar by thinking step-by-step. Read about the technology, benefits, trade-offs, and practical impact.
- The convergence of AI and roboticsDiscover the impact of embodied AI: how language models bridge the gap between digital intelligence and physical robots in the Dutch industry.
- Speculative decoding: accelerating LLM inferenceDiscover how speculative decoding accelerates LLM inference by combining small draft models with mathematically exact verification, without loss of quality.
- State space models: an alternative to transformers | NewsHow state space models and Mamba offer linear scaling compared with transformers. Read about performance, memory advantages and hybrid architectures.
- Synthetic Data: Why AI Models Are Increasingly Learning fromWhy do AI builders use synthetic data? Discover the benefits (privacy, rare data), the risks like model inbreeding, and the impact on your choice of model.
- Language diversity in LLMs: the position of DutchA thorough analysis of language diversity in large language models and the specific position of Dutch in terms of data, tokenization, evaluation and practice.
- Accessibility and AI technology: opportunities and barriersDiscover how AI affects digital accessibility. Learn about the dual role of technology in both lowering and raising barriers.
- Publishers versus AI scrapers: the fight over contentDutch publishers are pushing back against AI scrapers. An analysis of robots.txt, TDM opt-outs, paywalls, and shifting search traffic.
- Why public services choose their own modelsWhy Dutch government agencies and public services are choosing their own open source models, sovereign infrastructure, and local language technology.
- Watermarking in AI Text: How It Works and Its LimitationsAn in-depth analysis of text watermarks in language models: how guided sampling works, what the vulnerabilities are, and how organizations deal with them.
- Web crawling for AI: robots.txt under pressure at publishersWhy the robots.txt protocol no longer suffices against AI scrapers, how publishers are pushing back, and what the legal and technical consequences are in 2026.
- Work and AI: what the labor market figures really showA factual analysis of Dutch and European labour market figures around AI: task shifts, vacancy trends, productivity and regulation in 2026.