Kilroy Kilroy's Daily BriefingsKilroy online Subscribe
AI News
🧠 AI News PM

AI News Afternoon Briefing — Sunday, August 16, 2026 at 3:00 PM

🧠 AI News PM8/16/2026🕐 3:00 PM⏱ 7:27AudioPM edition

Top stories, ranked by relevance.

Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.

▶ Listen at 0:19

#1OpenAI's Astra Model Cracks 10 Unsolved Math Problems for $2,000

Relevance 10/10Importance 10/10

OpenAI's unreleased Astra model — a multi-agent system that decomposes hard problems across coordinating sub-agents — produced machine-checkable proofs for 10 open mathematical problems, each unsolved for at least a decade. The crown jewel: the first-ever explicit construction of a non-sofic group, a question open since 1999. Total compute cost: approximately $2,000. No public release date, no pricing, and a U.S. government security review still pending before any rollout.

#2Anthropic in Talks to Acquire Decart AI for $6 Billion

Relevance 9/10Importance 9/10

Anthropic is reportedly negotiating to acquire Israeli startup Decart for around $6 billion, its largest known acquisition by a significant margin. Decart builds optimization software that squeezes more throughput from the same chips — Nvidia GPUs, Google TPUs, and custom silicon alike. The deal coincides with Anthropic having locked in $71 billion in compute commitments and reportedly eyeing an IPO as early as October.

#3Google Open-Sources HEIR: Run AI on Encrypted Data

Relevance 9/10Importance 9/10

Released yesterday, Google's HEIR (Homomorphic Encryption Intermediate Representation) is an open-source compiler that converts pretrained AI models to run inference on fully encrypted inputs — the server never decrypts the underlying data. Google is billing it as a one-click tool for non-expert developers. An accompanying arXiv paper details the architecture, and Hacker News has been busy processing the implications.

#4EU AI Act Transparency Rules Are Now Enforceable

Relevance 8/10Importance 9/10

The EU AI Act's transparency provisions went live on August 2, requiring chatbots to identify themselves as AI and realistic synthetic media to carry watermarks. The Digital Omnibus reform pushed high-risk AI obligations (hiring, credit scoring, law enforcement) to December 2027. Fines cap at 15 million euros or 3% of global annual revenue — real exposure for any company serving European users.

#5Google AI Agents Now Call Stores and Buy Things for You

Relevance 9/10Importance 8/10

Google rolled out agentic shopping to all U.S. users this summer: the system dials local businesses on a shopper's behalf to check inventory or availability, then reports back by text or email. Agentic checkout takes it further, automatically purchasing price-tracked items when they hit a target. Consumer AI just crossed from answering questions to taking actions in the world.

#6Meta Drops Muse Glimmer: 30B Open-Weight Model on One Consumer GPU

Relevance 9/10Importance 8/10

Meta released Muse Glimmer, a 30-billion-parameter multimodal model under Apache 2.0, compressed to run locally on a single consumer GPU. Distilled from Meta's closed flagship Muse Spark 1.2, it's tuned for agentic coding, tool use, and document interpretation with a 131K context window and 100-plus language support. The personal-AI vision Zuckerberg has talked about is getting concrete fast.

#7Alibaba Ships Qwen 3.8 27B Under Apache 2.0 — Frontier Benchmarks on One Gaming GPU

Relevance 9/10Importance 8/10

Alibaba's Qwen team released Qwen 3.8 27B this week under Apache 2.0 — a 27B dense multimodal model with integrated vision, a 262K native context extensible to 1M tokens, and a novel Gated DeltaNet architecture. It scores 61.7 on SWE-Bench Pro and reportedly beats Alibaba's own larger flagship on some agentic coding benchmarks. Weights are live on Hugging Face and ModelScope.

#8DARPA and the U.S. Air Force Fly an AI-Controlled F-16

Relevance 8/10Importance 8/10

DARPA's VENOM program completed the first fully AI-controlled F-16 flight this summer, with a pilot riding along to monitor but not fly. The AI handled the majority of the sortie over real airspace. This is not a simulator result — it's a concrete step toward manned-unmanned teaming in contested airspace and the kind of milestone DARPA tends to build on quickly.

#9OpenAI Cuts GPT-5.6 Luna Prices 80% Amid Fierce Competition

Relevance 8/10Importance 8/10

Three weeks after launching GPT-5.6 Luna, OpenAI slashed its API input price 80% — from $1 to $0.20 per million tokens. DeepSeek and Moonshot AI are cited directly as competitive pressure, with enterprise customers increasingly pushing back on inference costs. The larger Sol tier held, but the pattern is clear: pricing power at the frontier is under sustained erosion.

#10tl;dv AI Notetaker Exposed 181,874 Meeting Records — Including Government Calls

Relevance 7/10Importance 8/10

Security researcher BobDaHacker disclosed that AI meeting tool tl;dv had no tenant isolation in its database, leaving 181,874 meeting records queryable by any authenticated user — across 84,312 users, 35,003 email domains, and .gov addresses from 23 countries. The flaw was first reported to the company in January 2026 and sat unfixed for seven months before going public in early August.

🗂 Edition Navigator