Lumis Daily Briefing — Jun 09, 2026 — Apple ditches in-house AI, goes all-in on Google Gemini
Apple Rebuilds Siri on Google Gemini in Stunning Pivot
Apple has abandoned its own AI model stack and replatformed Siri on Google Gemini, a seismic shift that hands Google a strategic foothold inside every Apple device and signals Apple's in-house AI efforts have failed to keep pace. This reshapes the competitive landscape for on-device AI and the Apple-Google relationship overnight.
Xiaomi's MiMo Hits 1,000 Tokens/sec on a 1-Trillion-Parameter Model
MiMo-v2.5-Pro-UltraSpeed delivers 1T-parameter inference at 1,000 tokens per second, a throughput benchmark that demolishes previous records and makes frontier-scale models viable for real-time applications. If reproducible, this resets cost and latency expectations across the industry.
Signal Calls Out UK Surveillance Bill as Existential Threat to Privacy
Signal's formal statement argues the UK's latest legislative push would functionally ban end-to-end encryption by mandating backdoor access, threatening secure communications for millions globally. The policy could force Signal and similar services to exit the UK market entirely, setting a dangerous precedent for other democracies.
xAI Pivots to Data Centre Rentals, Raising Lab-Credibility Questions
Analysis shows xAI is generating significant revenue by leasing its Colossus supercluster capacity to third parties, behaving more like a real-estate investment trust than a frontier AI lab. This raises hard questions about whether xAI's compute is actually being used to advance Grok or simply to monetise infrastructure.
Popular Thesis: AI Capability Gains Are Genuinely Slowing Down
A widely-read analysis argues that benchmark improvements and qualitative capability jumps have materially decelerated in 2025-2026, challenging the assumption of continued exponential progress. If the slowdown is real, it has direct implications for AI investment theses and product roadmaps built on near-term AGI timelines.
Researchers Find Attack Selection Critically Undermines AI Safety Evals
A new arXiv paper demonstrates that the choice of adversarial attacks used in agentic AI control evaluations dramatically skews safety scores, meaning published safety benchmarks may be systematically overoptimistic. This is a direct challenge to the evaluation frameworks regulators and labs rely on to certify safe deployment.
Thermo Fisher Antibody Data Manipulation Alleged at Scale
A detailed investigation claims widespread manipulation of antibody validation data across Thermo Fisher's product catalogue, a finding that could invalidate years of life-science research that relied on those reagents. The implications for reproducibility and regulatory submissions in biotech are severe.
Get tomorrow's briefing delivered at 07:00 UTC.
Lumis synthesizes the top AI and tech developments into a sharp 3-item briefing — personalized to your sources, delivered before your day starts.
7 days free. Cancel anytime. No credit card needed.