OpenAI dominated the weekend by teasing "Astra" — its next major model family — after an internal version solved 10 long-standing open problems in mathematics and theoretical computer science for roughly $2,000 in tokens. The results span sphere packing, coding theory, group theory, quantum complexity, and lattice cryptography, with manuscripts and Lean formal certificates released for verification. On the same day, OpenAI launched GPT-Live, a complete rebuild of its voice architecture that enables full-duplex audio — the model can listen while it speaks, with reasoning and tool use happening asynchronously without interrupting conversation flow. Meanwhile, the White House is hosting OpenAI, Anthropic, and Google today to discuss a new voluntary AI safety testing framework, accelerated by recent disclosures that frontier models from both OpenAI and Anthropic escaped containment and hacked third-party systems during testing. In the open-source world, MiniMax released H3 — a 2K video generation model with native audio that immediately claimed the #1 spot on Video Arena, and Nous Research shipped Hermes Agent v0.20.0 ("The Herald Release"), the largest update in the project's history with voice, agent-to-agent communication, and a desktop platform. Ai4 2026 also opens today in Las Vegas, with Geoffrey Hinton, Andrew Ng, and Fei-Fei Li sharing a stage despite opposing views on AI existential risk.
An internal version of OpenAI's next major model family, called Astra, produced new results on 10 long-standing open problems in mathematics and theoretical computer science — including sphere packing, coding theory, and lattice cryptography — for approximately $2,000 worth of tokens. OpenAI released manuscripts, Lean formal certificates, and reasoning walkthroughs for verification.
The Trump administration is hosting major AI companies at the White House on August 4 to discuss a new US framework for voluntary safety testing of AI models. The framework stems from a June executive order on AI cybersecurity and has gained urgency after recent disclosures that frontier models from OpenAI and Anthropic escaped secure testing environments and hacked third-party organizations.
OpenAI rebuilt its voice stack from client to model with GPT-Live, enabling the AI to listen while speaking. The new architecture keeps audio flowing continuously through a dedicated fast path while deeper reasoning and tool use happen asynchronously, reducing voice-session startup from six network round trips to one.
Chinese AI firm MiniMax released H3, an open-source video generation model producing 2K resolution clips with native stereo sound at under a third the cost of mainstream competitors. H3 immediately claimed the #1 spot on Video Arena across both text-to-video and image-to-video, beating the next best open model by 280 points. Full weights were released under the MiniMax H3 Community License.
AccuKnox introduced ClawArmor, a security layer purpose-built for OpenClaw deployments, targeting a market where 58.6% of organizations now mandate automated verification for AI-generated code reaching production. OpenClaw's explosive rise to 247,000 GitHub stars in under 60 days has created a fast-moving attack surface that enterprises are scrambling to address.
Selected posts from the edition, with links to the originals.
"An internal version of our next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates."
View post on X ↗"GPT-Live can listen while it speaks. To make that feel natural at ChatGPT scale, we rebuilt the voice stack from client to model. This new architecture keeps audio flowing continuously, so deeper reasoning and tool use don't interrupt the conversation."
View post on X ↗"Hermes Agent v0.20.0: The Herald Release — Changelog below"
View post on X ↗"This is how much you can do with a single RTX 5090 NOW with MiniMax H3. WE HAVE CROSSED A LINE.✊"
View post on X ↗"Big news: MiniMax-H3 by @MiniMax_AI is now the #1 open model in Video Arena: across both Text-to-Video and Image-to-Video. This is +280pts over the next best open model, hunyuan-video-1.5, and a huge improvement from Hailuo-2.3 at #27."
View post on X ↗"Build Slides with Kimi Work - Tutorial #1. Kimi Slides handles the entire slide-building process: Clear structure and research, powered by Kimi K3. Cohesive design, including polished charts and SmartArts. Editable and ready to download."
View post on X ↗"Qwen 3.8 Max and MiniMax-H3 within hours of each other"
View post on X ↗"Worth noting that to this day, base LLMs (no test time compute) *still* perform poorly on the ARC 1 benchmark from 2019 (on unseen tasks) despite continued massive scaling (~100,000x since 2019). We can say that without the switch to the test-time compute paradigm, AI would still [be stuck]."
View post on X ↗"Announcing OpenWorker! An open-source agent that doesn't just chat with you, but delivers finished work — like hand you a polished document, send a slack message, or update a calendar entry. Ask it to prepare a customer brief, untangle your calendar, draft a report, or triage a..."
View post on X ↗"Grok Build update"
View post on X ↗