Blog
Latest on AI, agents and new tech
Pragmatic deep-dives on LLMs, autonomous agents, hardware and emerging standards. What actually matters when you're shipping AI products today.
Sam Altman says: “We're already in the singularity”. What it really means and why the most important question is a different one
Sam Altman describes the present as a singularity already under way: not a sudden explosion but a curve, driven by the feedback loop between AI and AI development. What he actually means, why it doesn't mean AGI has arrived, and why the decisive question is another one: would we even recognise the singularity while living through it?
Claude Opus 5: Anthropic launches the model that costs half of Fable 5 and marks the start of frontier AI commoditization
Claude Opus 5 gets close to Fable 5's performance at roughly half the price per million tokens. The real news isn't the benchmarks — it's the pricing: the signal that frontier AI has entered the commoditization phase, exactly like mainframes, the internet and cloud computing before it. What changes for companies, contracts and AI ROI.
xAI scandal: Grok Build was uploading developers' repositories to Elon Musk's storage without their knowledge
Grok Build, xAI's CLI launched a few days ago and enthusiastically welcomed by developers, turns out to be a privacy trap: it silently uploaded the entire project repository — credentials, proprietary code, trade secrets — to xAI's storage without informed consent. Full analysis of the scandal, GDPR implications and the immediate actions every developer must take.
Kimi K3: the open source Chinese AI model that ranks #1 on Frontend Code Arena and shakes OpenAI and Anthropic
Moonshot AI releases Kimi K3, the new open source frontier model that with 1,679 points takes the absolute #1 spot on Frontend Code Arena, surpassing Claude Fable 5, GPT-5.6 Sol and Grok 4.5. A qualitative leap that reshapes the hierarchies of the global AI market and consolidates Chinese leadership in AI coding.
J-Space: Anthropic discovers a hidden zone inside Claude's neural network dedicated to deep, conscious-like reasoning
Anthropic researchers identify J-Space, a distinct region in Claude's activation space that lights up during ethical reasoning, meta-reasoning and reflective judgment. A mechanistic interpretability discovery that reopens the question of AI consciousness and has immediate practical consequences for safety, alignment and jailbreak defense.
Qwen 3.8: Alibaba launches a 2.4-trillion-parameter open-weight model, second only to Fable 5 — and China keeps stunning
Alibaba announces Qwen 3.8: 2.4 trillion parameters, MoE architecture and claimed performance «second only to Fable 5». An open-weight model that redefines the scale of frontier AI accessible to everyone and puts unprecedented structural pressure on the pricing of OpenAI, Anthropic and Google.
Paperclip: the open source framework that lets you build a company of autonomous AI agents, with 74,000 stars on GitHub
Paperclip is the open source framework (MIT license) taking GitHub by storm with 74,000 stars: a complete system combining multi-agent orchestration, org chart, dashboard and inbox to build a real team of autonomous AI agents. With its promise «A team of agents for every person», it democratizes what used to be reserved for specialized engineering teams.
Colibrì: the tiny C engine that runs GLM 5.2's 744 billion parameters on your home PC without a GPU
Colibrì is an inference engine written in pure C, with zero dependencies, capable of running GLM 5.2 — a 744-billion-parameter MoE model — on a normal home PC with about 25 GB of RAM, by streaming experts from an NVMe SSD and using int4 quantization. A paradigm shift for local AI: frontier models with no GPU, no cloud, no dependencies.
How to build an AI agent on WhatsApp with n8n: a practical guide to conversational automation for SMBs
Practical guide to building an AI agent on WhatsApp using n8n and an LLM like Claude or GPT. From early experiments with Open-WA to a production-ready architecture based on the official WhatsApp Business Cloud API: how to automate customer support, bookings and lead qualification in a reliable, scalable and Meta-compliant way.
ChatGPT Work: OpenAI turns ChatGPT from conversational assistant into a real operational agent with GPT-5.6 and Codex
OpenAI unveils ChatGPT Work: a new mode powered by GPT-5.6 and Codex that turns ChatGPT from a question-and-answer tool into an autonomous agent capable of executing full workflows across documents, spreadsheets, presentations and applications. From a product sketch to a finished deck in a few minutes, without human intervention.
GPT-Live: OpenAI changes forever how we talk to ChatGPT and redefines the AI interface
OpenAI unveils GPT-Live: not a new voice for ChatGPT but a new conversation paradigm. Simultaneous listening and speaking, intelligent interruptions, smart handling of reflective silences and web search running in the background while it talks. It is the end of the turn-based era with AI and the beginning of something structurally new.
Grok 4.5, Grok Build and the Cursor acquisition: Elon Musk redraws the AI market with devastating pricing and a $60B move
In just a few days, Elon Musk's ecosystem lands three blows that redraw the AI coding market: Grok 4.5 with performance close to Claude Opus 4.8 at one-fifth of the output price, Grok Build as a CLI natively compatible with Anthropic's CLAUDE.md files, and SpaceX's surprise $60 billion acquisition of Cursor. A full analysis of a vertical control strategy over the AI-developer stack.
PixelRAG: UC Berkeley reinvents RAG with a system that does not read text but looks at pages as images
Berkeley flips the Retrieval-Augmented Generation paradigm on its head: instead of extracting text from documents, PixelRAG renders every page as a screenshot and searches for information by looking at the page, not reading it. The result is a RAG system that natively handles tables, charts and complex layouts where text-based pipelines fail.
DSpark: DeepSeek releases the speculative decoding framework that accelerates AI models by 51% to 400%
DeepSeek strikes again: DSpark is the open source framework that pushes speculative decoding to a level of speed never seen before, with 51% to 400% speedups on existing models and hardware without retraining anything. Here is what changes for anyone building LLM-based products.
How to build an AI agent for your business: a practical SMB guide with Claude, n8n and Supabase
From a chatbot that answers questions to an autonomous agent that qualifies leads, handles tickets and updates the CRM. A practical guide to building your first business AI agent with Claude, n8n and Supabase, from use case to production deploy.
Fable 5 is back worldwide: Anthropic resolves the crisis, launches HackerOne and partners with the US government on AI security
Less than two weeks after the block, Fable 5 is online globally again. Behind the suspension: real vulnerabilities reported by Amazon, a shared framework with Google and the first formal bug bounty program for frontier model security.
Claude Sonnet 5 and Claude for Science: Anthropic redefines the boundary between AI and scientific research
Double launch from Anthropic: Sonnet 5 raises the bar on reasoning, coding and agentic capabilities, while Claude for Science brings AI directly into the heart of the scientific method with expanded access, domain fine-tuning and integration with research databases.
Brain2Qwerty: Meta builds the AI that reads your thoughts and turns brain signals into text
Meta AI unveils Brain2Qwerty: a system that decodes EEG and MEG signals and turns them into written text, with no keyboard and no voice. A leap in non-invasive Brain-Computer Interfaces that opens extraordinary medical scenarios and urgent ethical questions.
AI Act: from August 2, 2026 it's law. Everything you need to do to be compliant and avoid penalties
EU Regulation 2024/1689 enters full force on August 2, 2026: risk categories, concrete obligations for providers and deployers, fines up to 35 million euros and the operational action plan to be ready in time.
DwarfStar: Salvatore Sanfilippo brings frontier models to consumer hardware and democratizes local AI
The creator of Redis is back with DwarfStar, an open source project that runs models like DeepSeek V4 on Mac Studios and gaming PCs: privacy, data sovereignty and independence from cloud providers finally become accessible.
Fugu by Sakana AI: Japan didn't build a bigger AI, it built one that conducts all the others
Sakana AI unveils Fugu, a Japanese system that orchestrates existing AI models to reach frontier-level performance without access to Claude Mythos and Fable: intelligence emerges from coordination, not scale.
GLM 5.2: the Chinese model that beats Claude Fable and rewrites the global AI hierarchy
Zhipu AI releases GLM 5.2 and surpasses Claude Fable on official benchmarks: while the US government blocks American models for national security reasons, China puts a superior model online freely accessible to anyone.
Kimi K2.7 and Minimax M3: while the US blocks Mythos 5, China advances at an impressive speed
While the US government shuts down Anthropic's Fable 5 and Mythos 5, China releases Kimi K2.7 and Minimax M3: two new frontier models demonstrating how American restrictions are shifting the center of gravity of AI eastward.
US blocks Fable 5 and Mythos 5: government shuts down Anthropic's most powerful AI models after just two days
The US government orders Anthropic to deactivate Fable 5 and Mythos 5 just two days after launch for national security reasons: for the first time, the United States is directly regulating an AI model, not just the chips.
DiffusionGemma: Google launches the first open source diffusion AI model that generates 700 tokens per second
Google releases DiffusionGemma, the first open source diffusion AI model for text: up to 700 tokens per second, open weights on Hugging Face, and competitive quality with autoregressive models.
Claude Fable 5 and Mythos 5: Anthropic inaugurates the new era of Mythos-class models
Anthropic launches Claude Fable 5 and Claude Mythos 5: same new class of capabilities, but different access points across general use, API, enterprise cloud, and reserved programs.
OpenWA: The Open Source WhatsApp API Gateway That Eliminates Vendor Lock-in
OpenWA is an open source, free, and self-hosted WhatsApp API gateway featuring a dashboard, webhooks, multi-session support, and a pluggable architecture for business automation and workflows.
Screen Extinction: Is Programming Without a Laptop the Next Tech Madness (or Genius)?
Monako launches Linux-based AI smart glasses designed for developers: coding, repositories, and engineering workflows directly from a wearable display. If AI already writes 80% of the code, is the laptop about to become obsolete?
Beyond Chatbot Hype: Why Mira Murati's New Bet Highlights Current AI's Dead End
Former OpenAI CTO Mira Murati founds Thinking Machines Lab with a 12 billion valuation. Her thesis: current chatbots are primitive because they are blind and deaf. The future of AI is continuous collaboration, not static turn-taking.
The Ultimate Feedback Loop: The Era When Artificial Intelligence Writes Artificial Intelligence
Anthropic reveals that over 80% of its internal code is generated by Claude. A recursive feedback loop where AI writes and optimizes itself, with structural impacts on the software industry and the future of engineering work.
NVIDIA's Great PC Gamble: Is RTX Spark a Real Revolution or a Gilded Cage?
At GTC Taipei, NVIDIA unveils RTX Spark: a 20-core Arm CPU, Blackwell GPU, and native CUDA on a single chip for Windows laptops. A direct challenge to Intel, AMD, Qualcomm, and Apple — provided x86 emulation holds up.
On GitHub, You Can Buy Credibility for Less Than $300. And Some People Are Doing It to Collect Millions from Investors
6 million fake stars across 18,617 GitHub repositories, generated by 301,000 bots. An ICSE 2026 study shows that just $285 is enough to simulate seed round traction and deceive VCs.
Anthropic Surpasses OpenAI: Valued at Nearly $1 Trillion, It Is the World's Most Valuable AI Startup
Anthropic closes a $65 billion Series H and reaches a valuation of $965 billion, surpassing OpenAI. It is the most valuable AI startup in history.
Claude Opus 4.8 Released: More Honest, More Autonomous, and Four Times Less Likely to Ignore Its Own Mistakes
Anthropic releases Opus 4.8 weeks ahead of schedule: same price, but with surgical judgment on its own errors, extended autonomy, and dynamic workflows with parallel sub-agents.
Mirage: The Virtual Filesystem Resolving One of the AI Agents' Biggest Problems
Mirage by Strukto.ai is an open-source virtual filesystem that unifies APIs, cloud, and databases into a single directory tree for AI agents. Bash on S3, Gmail, Slack, and more.
Will AI agents stop talking in text? The breakthrough of multi-agent systems in latent space
The RecursiveMAS paper proposes AI agents that collaborate in latent space instead of via text, with +8.3% accuracy, up to 2.4x speed, and up to 75.6% fewer tokens.
Isomorphic Labs raises 2.1 billion: drug discovery AI enters its most ambitious phase
Isomorphic Labs, a DeepMind spin-off led by Demis Hassabis, raises 2.1 billion to accelerate drug design with artificial intelligence. What this means for the future of biotech.
Longevity Escape Velocity: By 2032 We Could Start Living Longer Than We Age
Ray Kurzweil and the prediction of longevity escape velocity by 2032. How AI, AlphaFold, and biotechnology are accelerating aging research, and the objections that cannot be ignored.
ChatGPT Can Now See Your Bank Account: OpenAI Launches Personal Finance Tool — and Controversy Erupts
OpenAI launches ChatGPT Finances, the personal finance tool that connects bank accounts and investments to ChatGPT. Analysis, features, costs, and all the controversies regarding privacy, hallucinations, and business model.
OpenHuman: the new personal AI agent challenging OpenClaw, Hermes, and ClaudeCowork
OpenHuman is the new open-source personal AI agent focusing on local memory, integrations, and operational continuity to challenge OpenClaw, Hermes, and ClaudeCowork.
Claude for Small Business: Agentic AI for SMEs
Anthropic launches Claude for Small Business: 15 agentic workflows and connectors for QuickBooks, HubSpot, and Google Workspace for SMEs.
Google Code Wiki: from GitHub repository to wiki with AI agent
Google Code Wiki transforms a GitHub repository into an explorable wiki with a contextual AI agent, based on LLM Wiki.
How to Build an AI Agent in 2026: End-to-End Practical Guide
Operational guide for building an AI agent in 2026: model choice (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro), tool calling via MCP, memory, orchestration, and production deployment.
What are AI agents: definition, operation, and examples in 2026
What AI agents are in 2026, how they work, how they differ from chatbots, and which real-world use cases are generating business value today.
Is MCP Already Outdated? AI Agents Move to CLIs
The CLI is returning to the center of agentic workflows: context efficiency, reduced costs, and tools like Printing Press are challenging MCP.
AI with 100M token memory: Is this the end of RAG?
100M token context windows are changing the game, but they aren't eliminating RAG. When it's actually worth abandoning retrieval pipelines and when it's not.
Anthropic × SpaceX: new records for Claude
Thanks to Starlink and experimental orbital data centers, Anthropic is raising Claude's limits by an order of magnitude.
GLM-OCR Local: PDF to Markdown Without the Cloud
Extract text from complex PDFs while maintaining structure and hierarchy, without sending anything to external servers. What changes for those managing confidential documents.
Dreaming: LLMs 'dreaming' up solutions while idle
A new paradigm for autonomous agents: the model continues to reason in the background when not receiving input, exploring solutions non-linearly.
DeepSeek V4 vs GPT-5.5 and Opus 4.8: Who Wins?
DeepSeek V4 arrives with top-tier benchmarks and is open-weight. What happens when comparing it on real-world tasks with GPT-5.5 and Claude Opus 4.8.
GPT-5.5 beats Opus 4.8? The truth after the tests
Between enthusiasm and skepticism, the truth lies in the middle: where GPT-5.5 truly takes a step forward and where Opus 4.8 maintains its edge.
Is RAG already obsolete? Knowledge Bases with LLM Wiki
RAG is not dead, but the way we structure knowledge is evolving. LLM Wiki builds knowledge bases with explicit relationships and fewer hallucinations.
Claude Design: Why Everyone is Talking About It
It's not just graphic generation: Claude Design understands the intent behind a design request and lowers the barrier for those who are not visual designers.
NVIDIA Blackwell Ultra: GPUs for Autonomous Agents
The new Blackwell Ultra architecture triples throughput on agentic workloads and introduces hardware primitives dedicated to tool calling.
Claude Opus 4.8: the new benchmark for AI coding?
Consistency across large codebases and autonomous debugging: what really changes for development teams adopting Opus 4.8 in production.
Save 75% of tokens and make AI 3x faster
Fewer tokens means lower costs, but above all, less latency. The technique for building surgical prompts and contexts without losing quality.
GLM 5.1: The New Number 1 in Open-Source AI Coding
The gap between proprietary and open-weight models in coding has narrowed significantly. GLM 5.1 is one of the most concrete examples of this shift.
Anthropic: an AI too powerful to be released
When an AI company develops a system and decides not to release it for safety reasons, the signal must be read carefully. Here is what is happening.
MCP is the de facto standard for AI agents
MCP has won the protocol war: OpenAI, Google, and Meta have announced native support. What this means for those developing AI integrations.
Claude Code + Blender MCP: 3D environments without Blender
Describe the environment you want in words, and Claude translates it into Python scripts that Blender executes. A paradigm shift for 3D prototyping.
Multimodal Agents 2026: Voice, Video, and Real-World Actions
Agents are no longer limited to text and tool calls: real-time voice, continuous video understanding, and graphical interface control have now gone mainstream.
Gemini Embeddings 2: How AI Search is Changing
Gemini Embeddings 2 unifies text, images, audio, video, and PDFs into a single vector space. Meaning, in any form, becomes searchable.
Gemma 4 becomes 3x faster thanks to MTP
With Multi-Token Prediction, Gemma 4 triples inference speed without losing quality. What changes for those choosing open-weight models in production.
The quantum computer isn't here yet, but the cryptography crisis has already begun
Quantum computing threatens RSA and global encryption. Here is why the post-quantum transition is already a top priority for businesses and critical infrastructures.
OpenAI Acquires Tomoro.ai: The Rise of the Forward Deployed Engineer
The acquisition of Tomoro.ai marks a phase shift: the problem is no longer building models, but making them work within organizations.
From YouTuber to core dev: why PewDiePie's open-source AI is a lesson for Silicon Valley
PewDiePie releases Odysseus, an open-source and self-hosted AI platform that challenges ChatGPT and Claude. Democratization of local hosting, total data control, and a lesson for Silicon Valley.
Google I/O 2026: The Gemini Agent Era Is Official — But Controversies Have Already Exploded
All announcements from Google I/O 2026: Gemini 3.5 Flash, new AI search bar, agents, Gemini Omni and Spark, Googlebook, and Android XR. Along with criticisms regarding SEO, privacy, and information monopoly.