OpenAI reported a 2.5x increase in usage of its agentic products (Codex, ChatGPT Work features) over the past week, signaling rapid real-world adoption of AI agents in production workflows.
Codex hit roughly 6 million active users, growing about 1 million in a single day and overtaking Claude's coding usage. Swyx called it 'IPO altering stuff' for OpenAI, linking to Latent Space analysis.
Chinese LLM developer DeepSeek is in discussions to raise $1.5 billion at roughly a $71 billion valuation, with an IPO targeted for 2027 or earlier. Existing investors include Tencent and the Beijing National AI Industry Investment Fund.
Think Through a Bottleneck enforces strict isolation between induction and deduction stages in frozen LLMs, yielding up to 14-point gains on ARC-AGI-2 and nearly doubling Verilog synthesis accuracy on ChipBench.
Direct-OPD transfers RL policy shifts from weaker to stronger models using implicit reward signals, boosting Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in 4 hours on 8 A100 GPUs.
ABot-N1 decouples cognition from control using a slow-fast architecture, achieving SOTA urban-scale navigation with POI arrival boosted from 35% to 77.3% and 95.4%/92.9% SR in complex indoor/outdoor scenes.
ABot-AgentOS provides a deliberative agent layer with Universal Multi-modal Graph Memory and self-evolution, achieving 88.7 on LoCoMo and 60.4 on OpenEQA EM-EQA with failure-driven improvement.
StudioRecon decouples background and humans using video diffusion densification and deformable Gaussians, achieving SOTA novel view synthesis with +1.5 to +5.0 dB PSNR gains over prior methods.
Zhipu AI released GLM-5.2, a new open-weights model featuring IndexShare in its DSA layers and a stronger Multi-Token Prediction head reusing IndexShare and KVShare techniques. Available on Hugging Face.
PrismML released Bonsai 27B, based on Qwen3.6, using ternary {-1,0,+1} weights with FP16 scaling (1.71 bits/weight, 5.9 GB). Retains roughly 95% of full-precision performance on reasoning, tool-calling, vision, and agentic tasks with 262K context. A 1-bit variant at 3.9 GB fits iPhones.
LMSYS published optimizations for serving GLM-5.2 on Blackwell hardware, achieving 500+ tokens/s per user at batch size 1. Key wins include 18-34% better single-user interactivity and a new TopK-V2 kernel 2.33x faster at 80K context.
NVIDIA published a technical blog on Ralph loops — iterative agentic patterns using filesystem as persistent memory — with LangChain Deep Agents Harness for Nemotron 3 Ultra. Tuning the harness rather than the model improves agent benchmarks without fine-tuning.
NVIDIA published a detailed analysis showing Blackwell NVL72 delivers the highest performance per watt for AI factories, positioning tokens per watt and cost per token as the key metrics for maximizing AI infrastructure revenue and profit margins.
New York enacted a one-year statewide moratorium on data centers requiring at least 50 megawatts, citing concerns over electricity prices, water consumption, and pressure on local infrastructure, with more than 12 GW of new loads awaiting grid connection.
AI startup Reflection AI signed a $1B+ compute agreement with Nebius for access to NVIDIA's newest processors, following a separate $150M/month deal with SpaceX through 2029, highlighting how securing compute capacity has become as critical as model development.
NVIDIA slashed approved Asian buyers for advanced AI chips by more than half, implementing a stricter whitelist and enhanced compliance reviews including on-site data center visits, targeting entities in Singapore, Malaysia, and Japan suspected of facilitating diversion to China.
Elon Musk's xAI installed 59 natural-gas turbines for its Colossus 2 data center project in Tennessee without obtaining required federal clean-air permits, with predominantly Black communities near the site raising health and environmental justice concerns.
AI team platform from the creator of Pythagora (33K GitHub stars) that assembles agent teams around business ideas to handle websites, outreach, content, and operations autonomously while keeping the user in control.
Open-source (AGPLv3) visual workbench for AI coding agents with CRDT-based branching session trees, Miller column navigation, JavaScript plugin system, and multi-client support across desktop, browser, and phone.
AI employee platform for e-commerce where users set high-level goals like revenue targets and a coordinated AI team plans, assigns specialist agents, executes, and reports updates via Slack or Discord.
Governance tool that installs approved policies as code onto AI agent harnesses for real-time runtime control with allow/block/escalate/advice outcomes and detailed auditing, initially focused on coding agents like Claude Code.
Programming language for AI-generated code with visible effects in function signatures, content-addressed definitions, and the Warp testing framework with replay and handler substitution.
Roughly $200M at a $2B valuation with Lightspeed in discussions to lead. The startup would focus on AI models for drug repurposing and discovery, with multiple OpenAI researchers expected to join.
AI-native compliance platform automating workflows for 500+ financial institutions raised $27M in a combined seed and Series A round led by CRV with Y Combinator and Pathlight Ventures.
Voice- and audio-forward AI dating service providing curated introductions raised $18M in seed funding backed by Match Group, FirstMark Capital, and Pace Capital.
Source: