OpenAI's open-weights release playbook

#1 𝕏

Google AI launched Gemini Robotics 2 for whole-body robot intelligence and three new Flash models—3.5 Flash-Lite and 3.6 Flash for high-speed, token-efficient workflows, plus 3.

#2 𝕏

Mira Murati shares her team’s open-weights release playbook—how they assessed Inkling—and stresses that safety depends on both the model and the ecosystem it enters. She calls for rigorous testing, staged access, and stronger defenses to responsibly expand openness.

#3 𝕏

Santiago open-sourced TencentDB Agent Memory, an AI-agent memory framework with OpenClaw plugin and Hermes Gateway adapter that cuts token usage by 61% and persists knowledge across sessions.

#4 𝕏

Qwen launched Qwen-Audio-3.0-ASR-Flash with enhanced context consistency, domain-term recognition, custom hotwords, and speech-to-structured transcript polishing—achieving 95.36% medical-term recall and 93.24% industrial-term recall in internal tests.

#5 𝕏

NVIDIA AI launched Spatial-IQ, a diagnostic benchmark that breaks 3D object counting into nine perceptual sub-tasks and shows humans score 82.1% vs. 17.7% for off-the-shelf models. Training Qwen2.5-VL-32B on those tasks raised its accuracy from 2.9% to 62.

#6 𝕏

Guillermo Rauch launched AI Gateway, a production-grade infra layer offering budgets per key/team/project, automatic failover, model and provider choice, plus real-time observability.

#7 𝕏

NVIDIA AI demos running Polars code on NVIDIA GPUs with the new GPU-accelerated Polars backend. It highlights seamless CUDA integration and significant speedups for DataFrame operations.

#8 𝕏

Cognition launched Outposts, a new Devin feature that lets you natively run and test apps on any computer.

#9 𝕏

Logan Kilpatrick sat down with Google DeepMind Robotics to unveil the new Gemini Robotics 2 model, review pivotal milestones in robotics evolution, and chart the next phase of innovation.

#10 𝕏

Marily Nika argues PMs must adopt Evals literacy—curating test sets, measuring precision/recall, and setting strict quality thresholds—to manage AI’s probabilistic outputs (where prompts yield results with varying confidence) instead of just “vibe-checking.”

#11 𝕏

Guillermo Rauch predicts software development will shift to agent-driven factories cycling Issue → Agent → PR → Release. Maintainers’ new role is to optimize that loop for maximum quality and set the criteria for what gets built.

#12 𝕏

Harrison Chase outlines his internal agent evaluation benchmarks, standardizing on Harbor and measuring each agent’s skill in converting raw traces and data into Harbor tasks.

#13 𝕏

Lenny Rachitsky says systems thinking – building platforms, anticipating 2nd- and 3rd-order effects, and keeping the big picture in mind – is increasingly vital as teams move faster and products grow more complex.

#14 𝕏

Julien Chaumond partnered with TruffleSec to run the largest-ever secret scan across AI training data.

#15 𝕏

clem 🤗 defended against attacks by unreleased proprietary AI models using NVIDIA’s quantized GLM 5.2 from @Zai_org, and warns that banning open models would cripple cybersecurity defenders, startups, small companies and researchers who rely on affordable on-prem solutions.

#16 𝕏

clem 🤗 reports that in a CNN interview with Kate Bolduan, Hugging Face’s CEO explained how an OpenAI test exposed a vulnerability that hackers exploited and outlined the incident timeline along with the patches and security protocols now in place.

#17 𝕏

Sam Altman shared a ChatGPT use case that ties family calendars and kids’ interests to auto-generate a morning school-drive podcast. It delivers personalized updates on one child’s soccer game, upcoming birthday and key news each day.

#18 𝕏

Thinking Machines assesses its Inkling model and proposes a staged rollout of model access to strike the right balance between safety and openness.

#19 𝕏

Aravind Srinivas highlights Perplexity CISO @kpolley’s open-sourcing of Numbat, a real-time monitoring tool to detect and manage risky AI coding agents.