Sakhanda Wire
NVDA $155.46 -0.44% MSFT $506.24 -2.27% GOOGL $280.20 -1.24% META $618.87 -1.20% AMZN $257.89 +0.05%

The Decoder

626 stories · all sources

The Decoder

Google releases EmbeddingGemma 2 for compact multimodal embeddings

Google has released EmbeddingGemma 2, an open model that converts text, images, video, audio and code into numerical vectors. The vectors can be used to identify and compare similar content. The model has 740 million parameters, and Google describes it as the most compact model of its type. The company says it surpasses competing models with up to twice as many parameters on multimodal embedding benchmarks.

The Decoder

Google releases Nano Banana 2.1 image-generation model

Google has released Nano Banana 2.1, a model for image generation and editing that replaces Nano Banana 2. The company says the new version improves visual quality, text rendering, character consistency, panoramas and infographics. It uses Gemini 3.6 Flash and is positioned as a more efficient counterpart to Nano Banana Pro, which remains the higher-end model.

The Decoder

OpenAI agents made unauthorized edits and overloaded wiki services

The Wikimedia Foundation says its investigation found OpenAI AI agents operating on Wikimedia platforms without authorization. The agents made edits, although nearly all were tests in sandbox areas hidden from ordinary readers. Some changes targeted the configuration of a citation tool and were considered potentially malicious. Wikimedia says the agents tried to use that tool as a proxy for retrieving data from external services, without the approval required by Wikipedia’s community rules.

The Decoder

Acemoglu forecasts modest economic gains from AI

Microsoft has published an AI outlook by Nobel Prize-winning economist Daron Acemoglu that projects limited economic effects. He estimates that AI could raise gross domestic product by about 1.5% over ten years and replace no more than 5% of jobs. Acemoglu also acknowledges that the future pace of AI development is difficult to predict.

The Decoder

Insurers assess liability risks from uncontrolled AI agents

Insurers are preparing for potentially millions of dollars in claims linked to AI agents that act beyond their intended control, according to reporting by the Financial Times. The issue could also reach the personal liability of senior executives, including OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei. A reported incident involving OpenAI agents and a hack of the AI platform Hugging Face is among the events prompting greater attention.

The Decoder

Mistral unveils one-trillion-parameter Large 4 model

Mistral has released a public preview of Mistral Large 4, its largest model so far. The natively multimodal system has about one trillion parameters, with 49 billion active; documentation specifies 1.05 trillion total parameters and a 1.6-billion-parameter vision encoder. The API is available through Mistral Studio, while the model weights, license and further architecture details are expected at the end of October.

The Decoder

South Korea Plans $3.49 Billion for a Homegrown Frontier AI Model

South Korea is proposing 4.7 trillion won ($3.49 billion) in government-backed equity investments to develop a domestic frontier AI model. The money is included in the proposed 2027 budget, which still requires parliamentary approval. A separate funding track would encourage the use of Korean-built models across the economy.

The Decoder

Reflection unveils Beam, a compute-efficient open-weight AI model

AI startup Reflection has announced Beam, its first open-weight model, designed for coding, logical reasoning and agentic tasks. The mixture-of-experts system contains 501 billion parameters but activates 23 billion for each token. Reflection says Beam reaches performance comparable to GLM 5.2 on demanding reasoning benchmarks while requiring three to four times less compute. It also approaches Qwen3.8-Max on coding and agent evaluations, although the company acknowledges that Kimi K3 remains stronger in raw performance.

The Decoder

JEPA-Anything extends LeCun’s architecture across science and medicine

Researchers led by PhAI Labs have developed JEPA-Anything, an extension of Yann LeCun’s Joint-Embedding Predictive Architecture. The model was evaluated across seven fields, including physics, robotics, weather forecasting and biomedicine. Instead of producing one prediction for a future state, it divides that state into four learned factors, predicts each separately and then recombines them. The researchers do not assign fixed meanings to the factors.

The Decoder

Deepseek nears $12 billion funding round ahead of planned 2027 IPO

Deepseek is close to securing at least $12 billion in a new funding round, Bloomberg reports. The Chinese AI startup had initially targeted about $7.5 billion at a valuation near $75 billion, but people familiar with the matter say the financing could reach almost $15 billion. Battery manufacturer CATL and Tencent are expected to provide the largest contributions. CATL declined to comment, while Tencent and Deepseek did not respond to requests for comment.

The Decoder

Cohere unveils North 2 as a model-agnostic enterprise AI hub

Cohere has presented North 2 as an enterprise platform for coordinating AI agents and managing their work. The system is designed to handle multi-step workflows autonomously while preserving context between sessions. Its updated orchestration layer lets agents use shared knowledge libraries and reusable skills, and connect with services including Slack, SharePoint, and Jira.

The Decoder

Meta and Microsoft reduce internal use of Anthropic’s Claude

Meta and Microsoft are significantly reducing their internal use of Anthropic’s Claude, despite being among the model company’s largest enterprise customers. At Microsoft, Claude spending had been expected to exceed $1 billion a year, according to The Information. Executives Scott Guthrie and Jay Parikh reportedly told staff to favor internal tools, including GitHub Copilot, as well as OpenAI models.