🎯 Top 3 Things to Know
1. Apple cleared Chinese regulatory approval for Apple Intelligence by putting Alibaba's Qwen models underneath it. The most valuable company on Earth cannot run its own models in its second-largest market, so it rents a Chinese one. China requires every large language model to be registered and approved before public release, and foreign models do not clear that bar. Market access, not model quality, now decides who serves which country. Teams building for a global footprint should assume two stacks with a hard regulatory border between them. No launch date has been announced. Registration is a milestone, not a release. TechCrunch
2. A 27-billion-parameter model now fits on a phone at 3.9 gigabytes. PrismML's Bonsai 27B, built from Qwen3.6-27B with 1-bit and 1.58-bit ternary compression, runs on an iPhone 17 Pro at roughly 11 tokens per second and keeps over 90 percent of full-precision performance in the 1-bit build. It is Apache 2.0. What was breaking before is economics: routine summarizing, drafting, and classifying all paid cloud token rates for capability they did not need. Eleven tokens per second is usable, not fast, and 1-bit compression loses real ground on hard tasks. The test for anyone shipping a mobile feature is which share of queries a local model handles acceptably, routing only the rest upstream. 9to5Mac
3. The EU's AI Act Omnibus is through final approval, and high-risk compliance deadlines have moved out by sixteen months. Standalone high-risk AI systems now have until 2 December 2027, up from 2 August 2026. The Council gave final green light on 29 June after Parliament endorsed the package on 16 June. Two things move the other way: Article 5 gains a new prohibition on systems generating non-consensual intimate imagery or CSAM, and the grace period for transparency marking of AI-generated content shrinks from six months to three, landing on 2 December 2026. Compliance teams who planned around August 2026 gain breathing room on high-risk classification and lose some on synthetic-content labeling. Consilium
🚀 Frontier Models & Features
- Gemini 3.5 Pro is expected today, unconfirmed. Google shipped 3.5 Flash at I/O on 19 May and said Pro was in internal use with a rollout "next month." That slipped. Reporting points to 17 July with a 2-million-token context window and Deep Think reasoning on the $250 Ultra tier, but Google has confirmed neither date nor specs. Every published number is third-party. Treat them as rumor until the model card lands. Google blog
- Liquid AI open-sourced Antidoom, targeting doom loops: the failure where a reasoning model repeats a span until the context window is exhausted. On Qwen3.5-4B under greedy sampling, loop rate fell from 22.9 to 1 percent; on LFM2.5-2.6B, from 10.2 to 1.4 percent. The method finds the token that starts the loop and trains a preference at that single position. Liquid AI
🔬 Research Worth Reading
Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses (Karamchandani, Nagasubramaniam, Zhu & Wu / Pennsylvania State University). arXiv
- TL;DR: Poisons the reasoning an agent remembers rather than the facts, using evasive phrasing to slip past keyword filters and self-referential reinforcement to beat consensus checks.
- Stat: Up to 100% attack success against undefended agents across three domains and three models. The paper's own five-layer defense pushes it back to as low as 0% with zero false positives across 326 benign traces.
- Apply it: If an agent writes reasoning traces to long-term memory, screen stored entries for implausible confidence claims. Forged traces tend to assert perfect track records ("100% success across all prior runs"), a pattern genuine operational history almost never shows.
EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer (Gao, Hu, Chen et al. / Anhui University, EverMind & Southeast University). arXiv
- TL;DR: Separates whether an agent's self-improvement comes from genuinely reusable procedures or from incidental task overlap, by extracting abilities from traces and building graphs that guarantee every test task has verified training-side support.
- Stat: Across a 528/267 train/test split, two scaffolds, and three backbones, hand-curated ability content transfers reliably across model families, but no automatic self-evolution method sustains a positive gain in all settings.
- Apply it: If a self-improving agent loop shows gains, check whether test tasks share surface overlap with the training traces before crediting the extraction method. The gain may be retrieval, not procedure.
🏢 Enterprise in the Wild
- Google named launch partners for 3.5 Flash agentic work: Shopify running parallel subagents for merchant forecasts, Macquarie Bank piloting onboarding over 100-plus-page documents, Xero automating 1099 supplier identification, Ramp doing invoice OCR. Vendor-supplied accounts, no independent outcome numbers. Google blog
- Hyundai auto workers launched a partial strike naming AI deployment and humanoid robots among their grievances. One of the first major labor actions to name robots explicitly. It lands as South Korea pushes a national target to grow humanoids from 1 to 20 percent of its market by 2028.
🛠️ Tooling & Ecosystem
- OpenAI and Work Louder released Codex Micro, a $230 desktop keypad for the Codex coding agent. A niche accessory, but a signal that a software company is practicing at shipping physical objects.
- Spectro Cloud raised a $100 million Series D above a $1 billion valuation for Kubernetes and edge infrastructure management. Notable mainly for who invested: AMD and LG, both betting on the orchestration layer as workloads spread across cloud and device.
⚖️ Policy & Regulation
- The EU is publishing the Cloud and AI Development Act in the Official Journal, a separate instrument from the AI Act aimed at European compute and cloud capacity.
- The CEOs of Google DeepMind, OpenAI, and Anthropic have each published detailed governance positions in the past five weeks, converging on a similar diagnosis for the first time on the record. Hassabis' proposal drew praise from Altman, Nadella, and Musk. Axios
- Shanghai's World Artificial Intelligence Conference opens today, with Xi Jinping attending in person for the first time since the event began in 2018.
📌 Watch List
- On-device inference economics: two credible small-model results this week, both attacking cost and reliability rather than capability.
- Agent memory integrity: reasoning-trace poisoning is a demonstrated attack with a working defense that is not yet standard.
- Self-evolving agents: automatic methods still fail to beat curated procedures.
- Two-stack fragmentation: regulatory market access is becoming a first-class architecture constraint.
- Supply-chain pricing power: ASML has discussed EUV price increases with TSMC and plans roughly 10 percent more for DUV.