Sources
- AI at Meta / @AIatMeta
- Amazon Web Services / @awscloud
- Anthropic / @AnthropicAI
- Cursor / @cursor_ai
- Google / @Google
- Google Cloud Tech / @GoogleCloudTech
- Google DeepMind / @GoogleDeepMind
- Grok / @grok
- Hugging Face / @huggingface
- Microsoft / @Microsoft
- OpenAI / @OpenAI
- OpenClaw🦞 / @openclaw
- Sequoia Capital / @sequoia
- Tesla / @Tesla
- Twitter / @a16z
- Waymo / @Waymo
- xAI / @xai
- Y Combinator / @ycombinator
Company@X — 2026-07-10#
Signal of the Day#
OpenAI announced that its newly available GPT-5.6 Sol Ultra model successfully produced a proof for the 50-year-old Cycle Double Cover Conjecture in under one hour. This breakthrough was achieved by utilizing 64 subagents, marking a massive leap forward in deploying multi-agent reasoning for complex mathematical research.
Key Announcements#
OpenAI · Source OpenAI launched GPT-5.6 Luna, a highly optimized model that outperforms the highest reasoning setting of GPT-5.5 at a 25x lower cost. Simultaneously, the company rolled back UI changes and reset usage limits across Codex and ChatGPT Work following developer pushback regarding unexpectedly expensive default settings and multi-agent regressions.
SpaceXAI · Source SpaceXAI introduced Grok 4.5, a frontier intelligence model trained explicitly with Cursor for coding and agentic workflows. Perplexity confirmed that Grok 4.5 scored highest among all frontier models in computer orchestrator evaluations at roughly half the cost of Opus 4.8. Box is already utilizing the new model via MCP to autonomously navigate secure folders, extract terms, and identify debt covenant conflicts in dense legal agreements.
Google · Source A joint six-day sprint involving over 100 AI agents and human developers successfully optimized Gemma 4 inference to be 5x faster on a single NVIDIA A10G GPU. This novel human-agent collaboration achieved a peak lossless speed of 315 TPS. Additionally, Google Cloud introduced the Open Knowledge Format (OKF), an open specification that standardizes the LLM-wiki pattern into a portable, interoperable format.
AWS · Source AWS launched Amazon Quick, its AI workspace assistant, in the London Region with native deep research and automated workflow capabilities. The deployment is explicitly built for strict data sovereignty, utilizing EU CRIS to guarantee that inference requests remain entirely within European AWS infrastructure for regulated financial and healthcare sectors.
Hugging Face · Source To accelerate the development of computer-use agents for creative tasks, Hugging Face open-sourced a massive image editing trajectory dataset. The dataset maps first-person designer reasoning to executable actions, including Photoshop MCP tool calls, menu paths, and specific coordinate clicks, aimed at solving the “choppy edits” problem in current visual agents.
Cursor · Source Cursor introduced “side chats,” which are durable agent conversations that developers can @-mention to pull context directly into main coding threads without interruption. The AI editor also added local search indexing for rapid querying across thousands of past agent transcripts. This launch comes as the broader AI-native editor space continues to surge, with opencode recently crossing VS Code in monthly web traffic.
Also Noted#
- OpenAI (Source): The company doubled its Bio Bug Bounty to $50K and shifted it to an ongoing private program, challenging researchers to find universal jailbreaks against its frontier biosafety safeguards.
- llama.cpp (Source): Following a direct contribution from NVIDIA, the framework added DFlash support to its speculative decoding techniques to further accelerate local model inference.
- Google AI Studio (Source): Developers can now claim permanent, custom URLs to give their deployed AI applications a personalized domain.
- Exa (Source): The AI search engine announced a strategic partnership with Cerebras to merge its high-speed search with ultra-fast hardware inference.