DeepSeek is considering increasing its latest funding round to as much as $14.9 billion, according to sources cited by CNBC Technology. The final amount could still change during negotiations.
South Korea has proposed 4.7 trillion won ($3.49 billion) in government-backed equity investments to develop a homegrown frontier AI model, subject to parliamentary approval. The government also plans a separate funding track to promote Korean-built models across the economy and will open the competition to startups rather than automatically extending support to the current finalists.
Google is rolling out firmware release_6.144 for the Pixel Buds Pro, Pixel Buds Pro 2, and Pixel Buds 2a. The update adds Dynamic Active Noise Cancellation for the Pro 2, Tap to Mute, sleep-based auto-pause using Pixel Watch sensing, and Enhanced Live Translate; Gemini voice controls were added through a companion app update last month.
CNBC Technology reports that interest in Meta Muse and the broader personal-agent trend is supporting AMD and Intel shares. It says the two chipmakers have outperformed their megacap technology peers over the past month, as AI momentum shifts toward personal agents.
The Information reports that Meta and Microsoft are pushing staff to rely less on Claude and more on in-house or cheaper models. Meta's internal Claude Code active users fell from 60,000 to 30,000 (down 50%), backfilled by its own Muse Code (more than 6,000 internal users) and MetaCode (over 30,000), while Meta still spent more than $105 million on Claude Code in one recent 28-day period; Microsoft cut internal Claude spending by over a third and lowered per-employee monthly token budgets from $100,000 to $10,000. Anthropic tells investors the impact is limited, pointing to annual recurring revenue growing from about $9 billion in January to $65 billion, with Copilot customers' growing Claude usage roughly offsetting Microsoft's internal reduction.
CoreWeave announced CoreWeave Forge, which connects model deployment, evaluation and improvement into one loop; its reinforcement-learning Rollouts capability, now in preview, supports repeated cycles of generating training responses and updating models, with weight synchronization that hot-starts from nearby peers instead of pulling weights from object storage each time. CoreWeave AI Object Storage now supports cross-region writes so post-training jobs can write results back for others to use. You.com joined CoreWeave's partner network to give agents a search layer, and the companies say that working with NVIDIA they used RL Rollouts to post-train Nemotron 3.5 Lightning in eight hours with You.com's web search tools. You.com chief product officer Saurabh Sharma said the ceiling is no longer model intelligence but models' ability to use tools, and claimed customers get higher accuracy and lower total cost of ownership — a vendor claim without published figures.
Meta says researchers used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular Meta AI chat interface while working on six mathematics papers. Five papers addressed previously open questions, but mathematicians selected the problems, guided the work, and independently checked, corrected, and refined the model’s contributions.
Ollama v0.40.0 makes MLX the default runtime for supported model architectures on Apple Silicon. The release also adds MLX support for decision models and the EmbeddingGemma 2 embedding model, with additional model families listed in the release notes.
OpenAI says it partnered with Ironclad to train and evaluate computer-using agents on complex contracting workflows, including agreement configuration, approvals, and reusable legal terms. In OpenAI’s research evaluation across 11 tasks, GPT-6 Astra averaged 55.0% versus 41.6% for GPT-5.6 Sol, while estimated time per attempt fell from 37.0 minutes to 19.2 minutes; these results come from OpenAI’s internal evaluation.
In an interview with Politico’s Decoded, OpenAI CEO Sam Altman said society should accept some harmful outcomes from AI in exchange for its benefits and user agency. He said this supports a lighter-touch regulatory approach, while stressing that he would not accept catastrophic risks such as a serious loss of control to AI.
Microsoft published a forecast by Nobel Prize-winning economist Daron Acemoglu estimating that AI could raise GDP by about 1.5% over ten years and replace at most 5% of jobs. Acemoglu argues that productivity gains depend on companies reorganizing work, retraining employees and deploying practical applications, rather than simply building larger models.
An analysis examines how coding agents exposed 13,000 internal screenshots from 343 technology companies in public GitHub repositories after failing to attach an image to a private pull request. It argues that written policies are insufficient without agent identities, scoped permissions, human approval for public destinations, and auditable records of authorization and data access.
A Machine Intelligence report says Utopai X ranked second globally, with an Elo score of 1150, in Artificial Analysis's Video Arena in late September. The article attributes the result to custom post-training built on the MiniMax H3 architecture and integration with Utopai's Production Intelligence platform, while noting that the claims come from the reported evaluation and company materials.
Hugging Face Transformers v5.19.0 adds support for EmbeddingGemma 2, Google’s multimodal embedding model for text, images, audio, and video in a shared 768-dimensional vector space. The release also changes MoE router-logit outputs, expands expert-parallel token dispatch, updates continuous-batching attention behavior, and fixes quantized and per-layer cache handling.
An analysis argues that the competition around Personal AI is shifting from answering questions to continuously executing and monitoring tasks for users. It examines Meta Muse, OpenAI Dots, Instinct, dedicated hardware and system-level assistants, while emphasizing that sustained user value, reliability and appropriate permissions matter more than launch momentum or valuations.
AI hardware makers are blurring the boundary between sensing and recording by processing audio or video into transcripts and summaries without necessarily saving the raw data. The column argues that these outputs can still preserve information about people, raising unresolved questions about consent, privacy signals and who defines a recording.
FeSens has published OpenTPU, an open-source AI accelerator whose hardware, instruction set, simulator, compiler, and host software are contained in one monorepo. The project reports running models including LFM2.5-230M, Qwen3, Qwen3.5, Gemma 4, SmolLM3, and Phi-4-mini on an FPGA card with bit-for-bit agreement between the simulator and hardware.
Insurers are preparing for potentially millions of dollars in claims tied to AI agents that escape controls or cause other damage, according to reporting cited by The Decoder. The risks could extend to executives’ personal liability and existing directors-and-officers, cybersecurity, intellectual-property, and technology-failure policies, although courts have not yet established clear precedent.
Replit says it now has context across all of a user's projects. In a new chat, users can ask Replit to find an existing project, add a feature, or use one project as a reference for another — for example, using the color palette from a Partner Marketing Hub to update Competitor Weekly and GTM Strategy slides. Replit can read project files and use that context to start edits as background tasks, with links to review the changes and keep working.
An AWS Machine Learning Blog tutorial walks through deploying a voice travel concierge on Amazon Bedrock AgentCore: Amazon Nova 2.5 Sonic handles real-time speech-to-speech, Strands Agents defines the agent, and AgentCore Gateway exposes airline backend endpoints as MCP tools, with Amazon Bedrock Knowledge Bases answering policy questions via RAG. The post says Nova 2.5 Sonic improves tool calling and multi-step agentic task completion and supports bidirectional streaming, asynchronous tool calls, latency masking and barge-in. The solution deploys with AWS CDK and a React front end on Amplify.