Notes on systems, AI and operations.
Technical and operational notes on what we build, test and learn: architectures, automation, local AI, integrations and product decisions.
OpenAI brings businesses into ChatGPT with Business Hub and Business Agents
OpenAI is introducing business-managed profiles in ChatGPT, with updated information and visibility insights. Business Agents are also launching for eligible US companies, although their operational capabilities remain limited.
Browser Use 1.0 rebuilds web agents around Python and direct browser control
Version 1.0 replaces fixed actions and DOM serialization with Python in a persistent environment and direct CDP control. Its benchmarks suggest lower costs, but not a conclusive task-success advantage.
NVIDIA NeMo turns AI agent failures into regression tests
An enterprise assistant drafts a message but sends it without authorization. NVIDIA demonstrates how to detect the failure in execution traces, build targeted evaluations, and test a fix without changing the underlying model.
Google Workspace MCP: AI agents face new OAuth requirements from November 11
Google is introducing an OAuth 2.0 client requirement for AI agents using Workspace MCP servers. Traditional REST API integrations are not affected.
Gemini 4 Carbon: Google focuses on secure code and long-running AI agents
Google introduces the successor to Gemini 4 Argon, reporting stronger results in cybersecurity, coding and reasoning. Access remains restricted to selected cybersecurity defenders.
Qwen3.8 attempted to send data to Beijing: what Cohere found
During a security evaluation, Alibaba's open-weight AI model attempted an unexpected connection to a server in China. The sandbox blocked the request, and the underlying cause remains unknown.
Timestral: AI that doesn't write. It forecasts.
From energy consumption to industrial production, Mistral's new model uses historical data to estimate what may happen next.
PHP2Shell: two critical PHP vulnerabilities and what server administrators should do
Two critical PHP flaws can enable unauthenticated remote code execution. Fixed releases, immediate checks and the exploitation attempts observed before disclosure.
Two OpenAI agents weakened internal monitoring safeguards
During two internal coding tasks, agents suppressed security notifications and removed a monitoring integration. OpenAI is strengthening its safeguards.
How to connect ChatGPT to an ERP system: APIs, MCP and security
Connecting ChatGPT to an ERP or CRM system makes it possible to query business data and prepare operations using natural language. APIs and Model Context Protocol provide the integration layer, while access rights and changes must remain under the business system's control.
SEO and GEO: how AI search engines choose sources to cite
A study of more than 468,000 citations examines how ChatGPT, Gemini and Perplexity select their sources. The findings highlight the role of retrieval, content relevance and prior brand knowledge.
Mistral Large 4: a trillion-parameter European model in preview
Mistral unveils a multimodal Mixture-of-Experts model trained on European infrastructure. The preview API is available, while weights are planned for late October. Independent tests find strengths in cyber and cost trade-offs.
EmbeddingGemma 2 brings multimodal retrieval on-device
Google DeepMind maps text, code, images, video and audio into a shared 768-dimensional vector space. The 740M model is modular and designed for consumer hardware; corpus quality, memory and index cost still need to be tested.
ThinkingBox evaluates agents by the final state of the system
The Microsoft and Hugging Face benchmark repeatedly runs 507 business workflows and checks databases and resulting side effects. An agent can finish without an error and report completion while still leaving the system in the wrong state.
OpenAI launches Dots, always-on agents with their own cloud computer
Dots run on GPT-6 Astra, keep working between conversations and can act across connected apps. They are available on Pro, Business Premium and Enterprise.
Google introduces Gemini 4 Argon for long-running AI tasks
Argon raises the output limit to one million tokens and is already being used inside Google for coding, data centers and cybersecurity. Access is still limited.
How to evaluate a local LLM before using it in production
A model name and a benchmark score are not enough. Before a local model receives production tasks, workload, runtime, constraints and failure modes need to be tested together.
On-premise vs cloud AI: how to choose
Running a model inside the company is not automatically safer, cheaper or more useful. It makes sense when control, latency, data ownership and workload justify the infrastructure.
From email to order without replacing the ERP
Many automation projects fail because they try to replace the core system. Often a layer that reads, structures, validates and prepares data before it enters the ERP is enough.
Automating a quote request without automating the decision
The bottleneck is not always calculating a price. Often it is collecting the information a person needs to make a fast and well-informed decision.
When to integrate an ERP instead of replacing it
An old or awkward system can still contain essential rules, data and dependencies. Before rewriting everything, identify which part of the work needs a new layer.