## News / Update
The AI ecosystem was packed with releases and governance moves. NVIDIA’s GTC drew more than 30,000 attendees and sold out in San Jose, with free virtual access and live keynotes teasing what’s next in accelerated computing. Niantic revealed Pokémon Go has quietly produced a massive AR dataset—over 30 billion images from 143 million players—fueling advances in visual AI. OpenAI introduced the IH-Challenge dataset to train models to respect trust hierarchies and resist prompt attacks, while StepFun made a rare open-data move by releasing its SFT training sets behind Stepfun-Flash. Benchmark integrity tightened as OpenBlocks was removed from Terminal-Bench 2.0 after irregularities. On the research front, several directions advanced: continual RL methods that avoid catastrophic forgetting with large VLA models and LoRA; evidence of “neural thickets” suggesting RL may be less essential for adapting large pretrained models; OpenClaw-RL showing agents can improve directly from usage; LoGeR’s hybrid memory for long-context geometric reconstruction; faster inference via RAT+ attention; and demonstrations that transformers can embed interpreters and reliably run code within their weights. Retrieval also progressed as MixedBreadAI showed late-interaction search can beat embedding-only baselines. Infrastructure got cheaper with Scale Up CPO cutting Mixture-of-Experts inference costs at scale. In document AI, Zhipu’s compact GLM-OCR set a new OmniDocBench V1.5 record (94.62) with a 0.9B-parameter design pairing a CogViT encoder and GLM decoder. Community and industry notes included Sakana AI warning job seekers of Gmail-based recruiter scams, Terence Tao launching the Mathematics Distillation Challenge, and ByteDance pausing Seedance 2.0 amid Hollywood copyright pressure while adding stronger guardrails. Momentum around agents remained high, with the Hermes Agent hackathon topping 70 submissions and a full-stack AI infra hackathon convening major labs. Hardware conversations highlighted a potential shift in inference from GPUs toward emerging LPUs, with new collaborations touted.
## New Tools
Developers gained several new building blocks. GEPA AI’s gskill introduced automated agent training that significantly boosts Claude Code’s repo task completion while cutting time-to-fix, hinting at more autonomous software workflows. A chrome-cdp skill now lets coding agents control Chrome directly via the DevTools protocol, enabling live browser automation without heavyweight frameworks. An experimental framework for trading agents, built on LangChain and Base with wallets and integrations, arrived for rapid prototyping in market environments. In design, MagicPathAI teased a forthcoming platform that aims to tightly couple design and code to streamline large-team handoff.
## LLMs
Model competition intensified. GLM-5-Turbo launched as a high-speed, agent-oriented upgrade to Pony-Alpha-2 and rolled out in hosted environments like Crush with increased quotas—positioning it for production agent workloads. Kimi K2.5’s rapid iteration, backed by custom optimizations on NVIDIA hardware, underscored fierce rivalry among inference providers. New benchmark chatter placed GPT-5.4 at the top for creative writing, long-form output, and emotional intelligence, while Grok-4.20 underperformed due to frequent refusals; a stealth “Hunter-alpha” with a 1M-token window surfaced, rumored to tie into the Qwen-3.5 lineage. An RLM-tuned GLM-5 variant also drew attention for notable capability gains.
## Features
Several products added meaningful capabilities. Gaussian Splatting content can now stream instantly in standard browsers, phones, and headsets with no downloads, turning previously heavy 3D video experiences into frictionless playback. The Transformers library gained FlashAttention-4 support, boosting speed and efficiency for next-gen models. Hermes Agent expanded operational features for production reliability—detecting outages, spinning up investigative subagents, writing runbooks, and cutting time to resolve repeat incidents—while another Hermes-built accessibility bot shipped a cleaner submission experience.
## Tutorials & Guides
New learning resources focused on practical agent-building and research literacy. LangChain launched a course on deploying robust AI agents, addressing nondeterminism, multi-step reasoning, and production reliability. A one-stop gallery of LLM architecture diagrams offered a visual map of model evolution for researchers and enthusiasts. Long-form content rounded out the week: a Post-AGI workshop talk unpacked game-theoretic commitments behind group self-sabotage, and ML Street Talk featured Sakana AI on combining evolutionary algorithms with LLMs for program optimization. A curated weekly research roundup highlighted advances like FlashAttention-4, KARL, Memex(RL), AutoHarness, and SkillNet to help practitioners stay current.
## Showcases & Demos
Hands-on demos spanned defense, robotics, AR, and grassroots automation. Palantir impressed U.S. defense leaders with real-time analysis depth and speed, reflecting why defense AI is commanding attention. At MWC, prototype Android XR glasses running Gemini previewed AI-first smart eyewear. Tsinghua University’s tennis robot rallied at near-human level. On consumer hardware, a 9B-parameter model running on a five-year-old RTX 3060 wrote and debugged a full space shooter, while a local Hermes agent on the same GPU handled project updates and device control. A Hermes-based auto-research prototype showed early promise for scientific discovery, and another Hermes deployment automated Android troubleshooting and GitHub issue filing. Project Aletheia scaled from a Raspberry Pi to a Mac Mini and pivoted to Hermes to accelerate anomaly-detection research. A developer also showcased an autoresearch harness training models on Apple’s Neural Engine in pure Go. Real-world medical anecdotes stood out: one pet owner combined AlphaFold insights and GPT to guide a dog’s cancer treatment with dramatic improvement, and another user credited GPT with finally resolving a years-long health issue.
## Discussions & Ideas
Debate centered on AI’s socioeconomic impact, evolving workflows, and the next hardware and data frontiers. Andrej Karpathy estimated around 40% of U.S. jobs are at meaningful risk from AI, with broader analyses suggesting even more roles are exposed; recent layoffs at big tech fueled speculation that smaller, AI-augmented teams may become the norm. Experts forecast agent-mediated contract negotiations within the decade under legal supervision, signaling faster deal cycles. Media and creative work are in flux, with predictions of hyper-personalized video and the end of traditional creative software paradigms. In healthcare, optimism around AI-plus-genomics and individualized mRNA therapies is tempered by cost and regulatory challenges, while wearables and AI may soon help prevent sudden nocturnal deaths. Technical discourse emphasized synthetic data’s rising role in pretraining, the promise of smaller, cheaper models and aggressive compression, and lessons from distributed systems re-emerging in multi-agent architectures. Security advocates pushed identity-first controls for autonomous agents. Builders debated design tools that abstract away code versus code-centric workflows, and celebrated the momentum behind CLI-driven AI agents—while urging rigorous testing. Geopolitically, observers warned that cheap AI will be harder to contain than past dual-use tech, while rapid Chinese robotics progress could reshape automation leadership. Community culture threads surfaced polarization in online AI debates and broader frustration with “enshittified” platforms. Entrepreneurs argued for building in public to accelerate feedback, and San Francisco’s outsized role as an AI hub appeared stronger than ever.
