## News / Update
OpenAI’s scale and momentum dominated headlines: ChatGPT now serves roughly 800 million weekly users, with the API handling about 6 billion tokens per minute and roughly 259 trillion tokens each month. To meet demand, OpenAI announced a multiyear partnership with AMD to deploy about 6 GW of Instinct GPUs, while also acknowledging compute is the major bottleneck. NVIDIA crossed a $4 trillion valuation milestone, underscoring the hardware race powering AI. CoreWeave is acquiring Monolith AI to bolster industrial AI R&D, and OpenAI and Stripe proposed the Agentic Commerce Protocol, an open standard for secure agent-driven transactions. Perplexity opened its first major EU office in Berlin, Medal spun out an AI lab after declining an acquisition, and Google DeepMind introduced CodeMender, an agent that already patched dozens of open-source security flaws. Anthropic open-sourced Petri for auditing risky model behaviors. In robotics, Figure humanoids have been running full workdays for months on BMW’s production line, and Amazon’s team showcased robots learning from long human motion sequences. Additional ecosystem updates included VS Code Dev Days expanding globally, Meta FAIR recruiting 2026 PhD interns, Google’s limited-time student access to its AI Pro Plan, Baseten and Nebius scaling video-generation infrastructure, and Common Crawl integrating IBM GneissWeb annotations for higher-quality training data.
## New Tools
OpenAI rolled out a comprehensive agent and app-building stack: the ChatGPT Apps SDK to build native, interactive apps inside ChatGPT; AgentKit and ChatKit for faster agent development; and newly published Agent Builder docs. Codex went GA with a new SDK, Slack integration, and admin analytics, plus GitHub Actions support to run Codex in CI/CD. A lighter, cheaper speech-to-speech model, gpt-realtime-mini, broadened access to voice AI. Across the open-source and dev tooling space, mcp-ui support landed on Hugging Face’s MCP Server, LanceDB 2.1 stabilized its file format with compression while preserving fast random access, and Jina’s reranker v3 shipped in GGUF and MLX for better search/ranking. Infrastructure providers highlighted production-ready stacks for video-gen workloads, signaling continued investment in high-throughput, low-latency inference.
## LLMs
Open models surged: GLM-4.6 jumped to the top of open-model leaderboards and trended on Hugging Face, with developers praising the GLM 4.5/4.6 family for capable, low-cost agent workflows. Efficient design was another trend: IBM’s Granite 4.0 H Tiny ran impressively on an iPhone 17 Pro, and ServiceNow’s Apriel-1.5-15B-Thinker matched larger models’ reasoning on a single GPU without RL. Tencent’s Hunyuan Vision 1.5 tied for the #3 spot on vision leaderboards, while a compact 0.6B model demonstrated strong parallel tool-calling, beating larger peers on that task. Anthropic’s Claude 4.5 showed powerful memorization yet weaker reasoning under estimation, and ARCS V3 posted consistent 90–98% on ARC-AGI-2, suggesting robustness beyond one-off claims. On the proprietary front, GPT-5 Pro arrived via API and received performance boosts, and GPT-5-codex amassed over 40 trillion tokens processed since August. Data quality advances continued as Common Crawl integrated IBM’s GneissWeb annotations for better filtering and curation.
## Features
ChatGPT evolved into a more capable hub with native, interactive apps users can talk to directly and a new Apps SDK that opens distribution to a massive audience. OpenAI expanded its API lineup with GPT-5 Pro, Sora 2 (and Pro), Veo-3, and a low-cost speech-to-speech model, cutting some media prices by up to 80%. The GPT-5 API prioritized speed with around 40% faster requests for premium users, and a new service dashboard added real-time visibility into reliability and usage. Ecosystem partners layered on enhancements: FactoryAI halved GPT-5 token prices for its users, Yupp added GPT-5 Pro access, Codex GA arrived with enterprise features, and Anthropic rebranded the Claude Code SDK to Claude Agent SDK as it powers most agent loops. Creative tools also leveled up: Pika introduced Predictive Video for script-to-scene automation, Synthesia 3.0 added real-time interactivity to avatar videos, and LlamaParse integrated Claude Sonnet 4.5 to upgrade document parsing.
## Tutorials & Guides
Hands-on learning and best practices were a theme: a live session walked through training a small Mixture-of-Experts model; OpenAI’s evals team published a cookbook on building resilient, consistent LLM apps; and the Hugging Face Transformers team shared a detailed blueprint for maintaining a million-line codebase spanning hundreds of model architectures. Practitioners also shared how to trim AI storage costs by up to 65%, interview prep notes from top labs, and a roundup of eight emerging RL trends. DSPy combined with GEPA emerged as a new open approach for prompt optimization with strong results.
## Showcases & Demos
Creative and real-world demonstrations highlighted AI’s expanding reach. Sora paired with GLM-4.6 produced cinematic-quality videos from concepts, while Mattel’s Sora 2 pilot compressed design-to-product cycles to minutes. Vision systems like Moondream spotted hard-to-see subjects in ocean rescue scenarios. On the physical side, Amazon showed robots learning from long human motion sequences, and Figure’s humanoids have been reliably performing production tasks at BMW. Generative media kept advancing, with Grok Imagine’s image-quality upgrade and Sora-driven “dreamscapes” illustrating how rapidly content creation is evolving.
## Discussions & Ideas
Commentary centered on the coming “agent era,” with predictions that 2025 marks a mainstream inflection point as visual builders and standardized protocols mature. Analysts examined the open-model landscape across China, the U.S., and the Llama ecosystem, and raised concerns about benchmarking methods that overemphasize math over real coding or agentic performance. Research perspectives proposed Advisor models to personalize black-box LLMs via RL, argued for RL-based pretraining to seed reasoning, and explored hierarchical memory for better long-horizon thinking. Practitioners warned of large training slowdowns from incomplete FP8 adoption due to late activation outliers. Strategic discussions noted AI demand outpacing compute, geopolitical competition in AI and green tech, and speculative futures like immortal, algorithm-optimized creators. Veterans revisited earlier talks on consciousness and predictive coding, framing how quickly once-fringe ideas enter today’s mainstream debates.
