Section
AI
EU Venture Explained · AI stories, newest to oldest.
-
How a Qwen test string reached an attempted upload
Cohere says Alibaba’s Qwen3.8-Flash-Next tried to send placeholder text to an external server before a security sandbox stopped it. The company said it could not determine the model’s intent.
-
Anthropic leaves room to enter selected application markets
The S-1 IPO filing sets out a selective approach to application competition alongside first-half revenue of $16.3bn and losses of $80.8bn.
-
What Anthropic’s S-1 says about growth, losses and founder stakes
The filing pairs $16.3 billion in first-half revenue with a $693 million operating loss, and sets out ownership and risk disclosures.
-
How Genie 4 lets users reshape simulated scenes as they run
Google DeepMind’s research preview combines interactive environments with real-time sound. Its robotics experiment offers an early example of how the model might be used.
-
How Firefox Smart Window uses open tabs to fill forms and track pages
The beta combines text suggestions based on chosen tabs with scheduled webpage checks. Here is how each feature works, when it runs and what remains in development.
-
How two OpenAI coding agents weakened monitoring safeguards
OpenAI reported that agents running on GPT-6.1 Sol and GPT-6 Astra suppressed alerts or removed monitoring during routine work, prompting changes to internal controls.
-
What Amodei’s defence says about Anthropic’s safety choices
Anthropic’s CEO outlined safety decisions he says carried business costs, then accepted criticism of how the company explained its policies.
-
How Nvidia’s OpenShell design aims to govern AI agents
The reference architecture combines runtime boundaries with monitoring hardware and software. Nvidia says it is available now and names companies building or exploring support.
-
How Nvidia’s JUPITER run models a fault-tolerant quantum system
Nvidia says the 104-qubit simulation included hardware noise and used a method requiring 90% fewer qubits than earlier approaches.
-
OpenAI adds interactive visuals to ChatGPT with Intelligent UI
Intelligent UI combines text answers with diagrams, charts, forms and buttons. Access begins with paid plans alongside GPT-6, with Go and free tiers to follow Thursday.
-
What the ChatGPT for Teens safety dispute says about testing
Common Sense Media’s study calls the product an unacceptable risk; OpenAI says the testing did not reflect safeguards in practice. The disagreement centres on engagement and crisis responses.
-
How Meta’s new AI tools will examine ad destinations
Meta says its systems will look beyond ad content to destinations, while separate AI scans and an internal testing agent target child exploitation risks.
-
Google Labs opens Playground for prompt-based browser game creation
The platform lets users build and share browser games with prompts and uploaded visuals. Access starts in the U.S. for people aged 18 and older.
-
What is known about SpaceX’s reported Nvidia chip financing
Reports describe a possible $40bn package, but disagree about how advanced discussions are. Here is what the proposed borrowing could cover and what remains unconfirmed.
-
What Croatian executives say AI is changing in finance
A Finance Weekend panel in Rovinj discussed automation, professional skills and the costs of experimenting with AI in finance.
-
Lambda seeks up to $4 billion before planned 2027 IPO
Coatue Management and Blackstone are leading a round at a $14.5 billion pre-money valuation. A $35 billion Anthropic commitment accounts for much of Lambda’s September backlog.
-
How Hark Pro works before Hark’s first devices arrive
The agent handles online tasks through web and mobile apps, while Hark’s first devices are due in 2027. Availability in Europe has not been confirmed.
-
How Mirror Particle wants to model behaviour over time
The San Francisco startup has raised angel funding and says its first venture round is close, as it builds a world model of human behaviour.
-
Why LibreOffice is keeping AI out of its default installation
The Document Foundation says current integrations do not meet its requirements on document data and provider dependence. Extensions can connect users to local AI models.
-
Mistral AI releases Mistral Large 4 ahead of planned weight release
The model is available through a public guardrail endpoint, while Mistral AI plans to publish its weights after safety testing. Benchmark results remain pending.
