OpenAI launches GPT-Live in ChatGPT Voice

#1 📝 Anthropic News

Our position on open-weights models - Anthropic CEO Dario Amodei states Anthropic has never advocated banning open-weights models and argues that non-dangerous open-weights are a public good while U.S. bans would not address his primary national-security concerns: authoritarian regimes developing superior AI for military/repression and the misuse of powerful models for cyber/biological attacks and alignment failures. Instead Anthropic calls for three policies: stop selling powerful chips and chipmaking equipment to China (citing scaling laws), crack down on industrial-scale distillation (which can bring the Chinese frontier within a few months of the U.S.), and mandate pre-release safety testing for all sufficiently capable models, open or closed, on a global basis.

#2 𝕏

OpenAI launched GPT-Live in ChatGPT Voice, bringing real-time voice interactions to Education, Business, and Enterprise plans globally.

#3 📝 Simon Willison

moonshotai/Kimi-K3 - Moonshot released weights for their 2.8 trillion parameter Kimi K3 (1.56TB on Hugging Face). The K3 license tightens commercial restrictions compared to K2, requiring separate agreements for large Model-as-a-Service businesses, and OpenRouter is already offering K3 via multiple providers at similar pricing.

#4 𝕏

Mustafa Suleyman launches MAI-Cyber-1-Flash with MDASH multi-agent security harness, hitting 96% on the CyberGym benchmark—12 points above Mythos—at half the cost.

#5 𝕏

Hugging Face is co-founding the Open Secure AI Alliance with NVIDIA and other industry leaders to share research, tools, and real-world expertise for identifying software vulnerabilities and bolstering AI security.

Also covered by: @Cognition

#6 𝕏

NVIDIA AI contributed open models, weights, data and research to the Open Secure AI Alliance, including its NVIDIA Labs Object-Oriented Agents (NOOA) open agent harness. It also released accompanying frameworks and benchmarks to accelerate secure AI development.

Also covered by: @Cognition

#7 𝕏

NVIDIA AI launched six new Agent Harness capabilities—task specification, agent composition, tool registry, memory management, evaluation, and feedback loops—to streamline multi-agent orchestration and boost GPU-scale LLM performance.

#8 𝕏

There's An AI For That emphasizes that agents are only as good as the context you hand them, launching HydraDB—a graph layer unifying memory, NVMe, and object storage to store rich context and reveal why an agent acted.

#9 𝕏

LlamaIndex 🦙 launched create-llama-worker, an npm CLI that scaffolds a ready-to-go Cloudflare Worker for edge parsing, classification, and extraction. Just run npm create @llamaindex/llama-worker — no boilerplate, no setup headaches.

#10 𝕏

Philipp Schmid rolled out EvoCode, a new evaluation spanning 26 tasks and 227 sequential rounds in one persistent container with cumulative tests after each turn.

#11 𝕏

Jason Zhou built an “agent-context-audit” skill leveraging @trq212’s Claude Code system prompts to apply his audit learnings and declutter his CLAUDE.md by ~50%.

#12 𝕏

Peter Yang shows how Jason used Codex to analyze a week of his Slack messages and build a “/write-like-me” skill that drafts replies in his voice, tailored for coworkers, executives, and external users.

#13 𝕏

Lenny Rachitsky spotlights how Anthropic’s Head of Product Dianne Penn uses evals-as-PRDs, token-sweating, and a Labs incubation model to sharpen frontier AI like Golden Gate Claude.

#14 𝕏

Guillermo Rauch highlights Kimi’s paper showing container-level isolation can’t stop agents from causing host kernel panics. He recommends Firecracker microVMs (as used in Vercel Sandbox) as a safe execution boundary.

#15 𝕏

Cognition launched a trustworthiness evaluation testing whether open LLMs repeat propaganda, comply with problematic requests, or generate insecure code based on who they’re serving.

#16 𝕏

Garry Tan demonstrates that GBrain is now state-of-the-art for agentic retrieval without any LLM rewriting, with full evaluation results in the gbrain-evals GitHub repo.

#17 𝕏

Yann LeCun warns that despite AGI hype, we still lack level-5 autonomous cars, cat-like dexterity, or AI that can learn to drive in hours like a teenager. He argues current AI falls far short of true general intelligence.

#18 𝕏

Andrew Ng endorses Jensen Huang’s Nvidia letter, calling for open models and robust defense harnesses after the OpenAI–Hugging Face hack. He warns that closed models aren’t safer but represent regulatory capture.

#19 𝕏

OpenAI finds that GPT-4 Enterprise use across 20+ companies automates routine tasks and spawns roles like prompt engineers and AI supervisors. This shift frees up employees for more creative and strategic work.

#20 𝕏

Madhu Guru argues that the best product reviews simulate market reactions to your ideas—compressing months of learnings into an hour with a room full of experts who deeply understand the space and hold strong, often correct, opinions.

#21 𝕏

Garry Tan argues that maintaining open weights and competition in the model layer is crucial to prevent a “god model” monoculture that would homogenize societal values and virtues.

#22 📝 Simon Willison

An opinionated guide to which AI to use to do stuff - Ethan Mollick's guide has evolved from focusing on chat models to emphasizing agentic systems that can perform extended work. The post explains differences between ChatGPT and Claude modes (Work/Codex and Cowork/Code) and notes that ChatGPT mobile's Work mode can permit internet access, changing capabilities significantly.

#23 𝕏

Lenny Rachitsky says you need frontier products to showcase frontier AI models and let users actually experience their magic.

#24 𝕏

Santiago Hyperagent launches cloud-based agents—zero setup or servers required—that learn new API skills, support browser actions, code execution, image/video generation and integrate with hundreds of tools.

#25 𝕏

Rowan Cheung unveils a mobile manipulator with a wheeled base extending 3'–5'9" and two articulated arms reaching 80" to work beds, counters, and closets, priced at $449/month or $7,999 upfront—half the cost of 1X’s Neo.