Apple teams up (secretly) with Google on Siri overhaul; OpenAI preps GPT-5 updates

AI Weekly Update - November 3, 2025

Get bigger weekly updates! For a limited time, you can grab a month-long trial of our paid newsletter which includes an extended weekly updates (11! additional stories this week) and some other benefits to keep up with the rapid AI news cycle.

last week’s top stories

🍎 Apple tapping Google’s Gemini to reboot Siri. In a major shift, Apple will pay about $1 billion annually for access to Google’s 1.2 trillion-parameter Gemini AI model. The Gemini AI will serve as a stopgap brain for a rebuilt Siri, vastly exceeding Siri’s current capabilities until Apple’s in-house system is ready. This custom Gemini integration aims to enable more complex, multi-step queries on Siri starting next year, without bringing Google’s search into Apple devices. Read more

🧠 OpenAI’s GPT-5.1 family set for imminent rollout. OpenAI is finalizing GPT-5.1, an upgraded suite including a base model, a reasoning-optimized variant, and a Pro tier. Sources say the models are already on Azure, aligning with OpenAI’s 3-4 month release cadence (GPT-5 launched in August). While not a dramatic leap over GPT-5.0, version 5.1 seems to focus on upgrading speed and performance. Read more

🤖 Google’s Gemini leak reveals Nano Banana 2 image AI. A new leak confirms Google is preparing two advanced models: Gemini 3 Pro for coding and Nano Banana 2 for realistic image generation. “Nano Banana 2”, codenamed GEMPIX2, was spotted on Google’s AI platform, indicating it could ship as early as December with high-resolution (4K) image outputs. Together with Gemini 3’s million-token context window, these models signal Google’s push to rival OpenAI’s latest systems. Read more

📝 Leaked 52-page memo details OpenAI co-founder’s case against Altman. Newly surfaced deposition testimony reveals that OpenAI’s then–chief scientist Ilya Sutskever compiled a 52-page memo in 2023 accusing CEO Sam Altman of a “consistent pattern of lying” and manipulation. The memo, shared only with OpenAI’s board, catalogued instances of Altman undermining and pitting executives against each other, and even alleged omissions like failing to inform the board of major initiatives. Sutskever and fellow director Mira Murati had secretly gathered evidence for months, using the memo to justify Altman’s brief ouster. Read more

💬 Snapchat inks $400M AI search deal with Perplexity. Snap has struck a partnership to integrate Perplexity’s conversational answer engine into Snapchat’s chat interface. Unlike typical deals, Perplexity will pay Snap $400 million (cash and equity) over one year for access to Snapchat’s 940 million users. Starting early 2026, Snapchatters can ask questions to the My AI chatbot and get answers sourced by Perplexity’s large language model instead of standard web search. Read more

🇨🇳 China orders state projects to drop US AI chips. In an aggressive self-reliance move, Beijing has banned foreign AI chips from new government-funded data centers. Regulators instructed that any state-subsidized facility still under construction must remove Nvidia, AMD, or Intel chips and use domestic alternatives (like Huawei’s) instead. Projects beyond 30% completion are reviewed case-by-case. The policy (one of China’s boldest efforts to purge foreign tech) comes amid a pause in US–China trade hostilities, yet it will likely dash Nvidia’s hopes of selling even scaled-down AI GPUs into China. Read more

🏆 Open-source model ‘Kimi K2’ shines in benchmarks. Moonshot AI’s new open-source model Kimi K2 Thinking has vaulted to the top of leaderboards, even outrunning OpenAI’s proprietary GPT-5 on complex tasks. Released on November 6 with 1 trillion parameters (in a sparse Mixture-of-Experts design), K2 Thinking excels at long-horizon reasoning and code generation. It posted state-of-the-art scores: 60.2% on the BrowseComp web reasoning test and 71.3% on SWE-Bench coding evals. Notably, K2’s agentic reasoning (up to 300 tool-use steps) surpassed Claude 4.5 and GPT-5 on several benchmarks. Read more

🎄 Coca-Cola doubles down on AI-generated holiday ads. Despite mixed reactions last year, Coca-Cola is expanding its use of generative AI in its iconic “Holidays are Coming” campaign. The company’s global head of generative AI, Pratik Thakar, said the 2024 AI-created Christmas ad (which drew creative industry backlash) was actually one of Coke’s best-performing holiday spots ever. This success spurred an even more ambitious AI-powered ad for 2025, now featuring improved AI cinematography, physics, and rendering quality to address last year’s visual glitches. Read more

---

🧪 AI Research of the Week

Densing law of LLMs
From Tsinghua University and OpenBMB

Jake’s Take: This paper tries to pinpoint how much measurable skill you get per parameter (learned weight) within an AI model. The authors estimate this “capability density” by turning AI benchmark scores into an effective parameter count, then comparing that to the actual size across 51 base models. They find the best models double capability per parameter about every 3.5 months. This means teams may be able to eventually match todays model’s accuracy with smaller models, lower latency, and lower cost.

I think it’s worthwhile for foundation companies to start tracking capability density as a core metric, investing in data quality, routing, and compression instead of chasing size, and pressure-test the trend on multilingual, retrieval, and safety tasks before moving budgets.

---

and then, even more news…

🏥 Microsoft forms ‘Superintelligence’ research team. Microsoft has created a new MAI Superintelligence unit, led by ex-DeepMind co-founder Mustafa Suleyman, to target AI systems more capable than humans in specific domains. The team’s first mission is a “medical superintelligence” for diagnosis within 2–3 years. Unlike efforts chasing general AI, Suleyman says this “humanist superintelligence” approach will tackle defined problems (like early disease detection or battery design) with superhuman accuracy but minimal existential risk. Microsoft is pouring substantial funding and recruiting top talent (including hiring chief scientist Karén Simonyan) to staff the group. Read more