Today’s lead · X
OpenAI says Astra produced 10 machine-checkable math and TCS proofs
Posts and community coverage say an internal version of OpenAI’s next major model family, Astra, generated 10 advances in mathematics and theoretical computer science, with Lean certificates and chain-of-thought walkthroughs. If validated by the math community, it is a notable signal for AI systems as research collaborators rather than only answer engines.
Top signals
3 more
Judge lets Minnesota’s nudify-app ban take effect despite xAI challenge
A U.S. district judge denied xAI’s request for a temporary restraining order against Minnesota’s ban on apps that generate non-consensual sexualized images, allowing the law to take effect while the lawsuit continues.
Dreamina launches Seedance 2.5 with longer AI video generation
Seedance 2.5 is live on Dreamina with native 30-second video generation, targeted editing, multimodal references, and multilingual creation. Enterprise API access through BytePlus is described as coming soon.
Engineers report high-throughput Kimi K3 serving on AMD MI355X
Two posts report a Kimi K3 serving setup on AMD MI355X reaching 952 tokens per second per node and 118 tokens per second single-stream, with claimed throughput and performance-per-dollar advantages versus Nvidia B200 and B300 references.
Tools & repos
3 selected
Port22
Port22 brings coding agents running on a Mac to an iPhone so builders can monitor Claude Code, Codex, and other sessions, see when an agent is stuck, and approve requested actions from the phone.
AgentMicro
AgentMicro is a local-first macOS menu-bar companion for supervising parallel Codex Desktop and CLI tasks without uploading prompts, responses, source code, or task history.
NudgeForMe
NudgeForMe scans sent email threads, identifies missed replies, and drafts follow-ups inside the user’s mailbox so leads, deals, partnerships, and customer conversations do not go cold.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Community discussions
5 threadsClaude Code-built feature passed tests but failed the core customer flow
A production rollback sparked a debate over AI-written tests: the same Claude Code agent implemented a wallet workflow, wrote unit and e2e tests, manually exercised APIs, and still shipped a broken main flow that a customer found.
Voice-agent builders debate whether latency or trust drives user dislike
A voice-agent thread asks whether users object to AI phone agents themselves or to slow turn-taking. The clearest takeaway is that latency matters, but interruption handling, recovery, handoff, and perceived agency matter too.
LocalLLaMA asks why new LLM leaderboards skew toward coding
A LocalLLaMA discussion questions the coding-heavy benchmark landscape and asks for more evaluation coverage across language learning, creative writing, STEM, medical, and biochemistry reasoning.
DeepSeek-V4-Flash run reported on RTX 3090 with CPU MoE offload
A LocalLLaMA user reports running DeepSeek-V4-Flash-0731 UD-IQ3_S through text-generation-webui on an RTX 3090 by replacing bundled llama.cpp binaries and keeping MoE experts in system RAM.
Local model builders debate the lower bound for model size
A LocalLLaMA thread asks whether small models can keep improving indefinitely or whether there is a minimum capacity for broad intelligence, reliability, knowledge, and generalization.
Funding & acquisitions
1 moves
Indian startup funding fell to $662 million in July
Indian startups raised $662 million across 85 deals in July, down sharply from June’s $2 billion spike but above May’s $630 million. AI drew the highest investor interest by deal value and count.
Bengaluru radar
5 events
Loop Café by Elevation Capital
A curated Bengaluru show-and-tell for builders who have working AI loops, from agent fleets to Claude routines and workflow automations. Attendees demo the real loop, explain prompts, context, tooling, token costs, and what broke.
AI Tuesdays: AMA with Marc Manara, OpenAI
A Bengaluru AMA with Marc Manara, Head of Startups at OpenAI, focused on company-building in the agentic era, platform dependence, enduring advantage, and building global AI category leaders from India.
The Algorithmic Battlefield: AI, Autonomy & the Future of Defence
A Bengaluru defence-tech evening on software-defined warfare, autonomy, sensor fusion, robotics, and India’s defence-tech landscape, with a technical deep dive, panel, and open-floor debate.
Lossfunk Research Showcase
Lossfunk’s Bengaluru research showcase brings researchers together for project talks across frontier intelligence science, LLMs for science, agent memory compression, transhumanism, and consciousness.
Glimpses of Singularity: Build products that build themselves
A Bengaluru workshop and live demo for founders and product builders on systems that spec, write, test, fix, and ship, with the human moving from author to director.