Plaud Launches New AI Pen and Desktop Meeting Recorder

2026-01-05

Plaud launches NotePin S, a compact AI recorder with 64GB storage, 20-hour battery, and Apple Find My support. Priced at $179, it offers clip, lanyard, and magnetic pin options. Paired with a new desktop app, it enables AI-powered meeting transcription and structured note-taking for virtual collaboration.

Read more

Google Engineer Says Claude Code Completed in One Hour What Her Team Took a Year to Finish

2026-01-04

Google engineer Jaana Dogan revealed Claude Code built a functional distributed agent orchestrator in one hour—matching her team's year-long effort. She praised its rapid progress, noting AI coding has advanced faster than expected: from line completion to full codebase management since 2022. While Google restricts Claude to open-source use, Dogan called its output impressive and motivating. Boris Cherny, creator of Claude Code, shared optimization tips: use verification loops, plan iteratively,

Read more

OpenAI Plans to Launch New Audio Model in Q1

2026-01-04

OpenAI is developing an advanced audio AI model for launch by March, aiming for more natural speech and real-time interaction. Built on a potential new architecture, it may support voice, music, and on-device processing, part of OpenAI’s broader consumer hardware strategy including smart devices and an upcoming audio-first personal device.

Read more

Swiggy Launches Hermes V3: From Text-to-SQL to Conversational AI

2026-01-03

Swiggy launched Hermes V3, a generative AI text-to-SQL assistant powered by LLMs, enabling employees to query data via natural language within Slack. Leveraging vector retrieval, conversational memory, and structured workflows, it improves SQL accuracy from 54% to 93%, supports multi-turn interactions, and includes an interpretability layer for transparency. Integrated with security and metadata systems, it uses Snowflake, LangChain, and vector databases for scalable, governed analytics access.

Read more

DeepSeek Develops mHC AI Architecture to Enhance Model Performance

2026-01-02

DeepSeek introduced Manifold-Constrained Hyper-Connections (mHC), a new AI architecture enhancing gradient stability and efficiency in large language models. By integrating geometric manifolds, mHC improves upon traditional residual and hyper-connections, achieving better performance with only 6.27% hardware overhead. Tested on models from 3B to 27B parameters, mHC outperformed baselines across eight benchmarks, offering scalable, resource-efficient training for next-gen AI.

Read more

SoftBank Completes OpenAI Investment with $22.5B Cash Injection

2025-12-31

SoftBank finalized its $40B investment in OpenAI, becoming the second-largest stakeholder with 11%. The funding supports OpenAI’s Stargate data center project and follows major restructuring. SoftBank sold NVIDIA and T-Mobile shares and leveraged Arm stock for financing. OpenAI's valuation has risen to $500B, with an IPO expected by 2027 at up to $1T.

Read more

Google Improves ART Compilation Speed by 18% Without Compromising Code Quality

2025-12-31

Google's ART team cut Android compilation time by 18% via targeted optimizations in JIT and AOT compilation, improving performance and efficiency without increasing memory use. Using pprof, they identified bottlenecks, streamlined compiler phases, reduced lookup complexity, and optimized data handling. Gains rolled out from June 2025, available on Android 12+.

Read more

Microsoft Copilot: GPT 5.2, Free Access

2025-12-30

Microsoft rolls out GPT-5.2 "Thinking" model as free "Smart Plus" mode in Copilot, enhancing complex task performance across web, Windows 11, and mobile. It excels in coding, document analysis, and office workflows, surpassing GPT-5.1 and matching or exceeding human experts in 70.9% of 44 professional tasks. Achieves 55.6% on SWE-Bench Pro, 80% on SWE-bench Verified, and perfect scores on AIME 2025 and strong results on GPQA Diamond (92.4%) and CharXiv Reasoning (88.7%).

Read more