In partnership with

Good morning. It’s Monday, December 8th.

On this day in tech history: In 1951, Christopher Strachey’s checkers program pulled an all-nighter on the Ferranti Mark I, becoming one of the first computer systems to play a complete board game without human intervention. Running on punch cards and vacuum tubes, it used a primitive minimax-style approach, but the implications were anything but small. It proved early computers could handle real strategic decision-making, a quiet but crucial precursor to modern reinforcement learning.

In today’s email:

• OpenAI’s GPT-5.2

• Nano Banana 2 Flash

• Grok 4.20 Makes 20% Stock Market Gains On It’s Own

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI’s GPT-5.2 ‘code red’ response to Google is coming this week 

OpenAI is accelerating the release of GPT-5.2 to as early as December 9 in a direct “code red” response from Sam Altman, targeting advanced reasoning, stability, and speed. The update is designed to close gaps exposed by Google’s Gemini 3 while focusing on ChatGPT’s core reliability rather than flashy new features.

At the same time, OpenAI has been forced to dial back in-app suggestions that looked like ads. Prompts encouraging users to connect with Target or explore apps like Peloton sparked backlash, even as leadership insisted no live ad tests were running. The feature has been disabled and new user controls are being developed.

ChatGPT still leads globally, but growth has slowed to 6% over the last four months. Gemini, powered by its Nano Banana image generator and deep Android integration, has seen daily engagement double, while rivals like Perplexity and Claude post triple-digit adoption.

Nano Banana 2 Flash is Google’s new cost-friendly AI powerhouse

Google is reportedly preparing to launch Nano Banana 2 Flash, a cost-optimized version of the Nano Banana 2 Pro, delivering comparable performance at lower operational costs. Internally dubbed “Mayo,” this tiered model is designed for high-volume or cost-sensitive deployments, integrating within the existing Gemini stack for enterprises and high-volume deployments.

Gemini Live is also coming to desktop web, bringing mobile features like voice, screen sharing, and live translation to browsers. Early signs suggest it could leverage Gemini 2.5, Gemini 3 Pro, or the Flash variant, enabling real-time multimodal processing and translation for productivity, learning, and collaboration.

Gemini 3 Pro sets a new benchmark in multimodal vision AI, excelling in document derendering, spatial reasoning, screen comprehension, and video analysis at 10 FPS. With a 1 million-token context window, Gemini 3 Pro enables deep reasoning across text, images, video, audio, and code. Alongside, Google has also released 15 practical examples demonstrating Gemini’s capabilities.

On memory, Google introduced Titans and the MIRAS framework. Titans selectively retains key information in sequences over two million tokens, while MIRAS enables attention-free models like Moneta and Memora. These advances point toward AI that learns continuously across text, video, and genomics.

Antigravity users on Pro and Ultra tiers now get higher rate limits while retaining full access to Gemini 3 Pro and workflow tools. Gemini adoption could lift Alphabet shares to $400, while OpenAI reassesses spending, and Cloudflare data shows Google’s 3–5× data advantage over rivals.

DeepMind CEO Demis Hassabis expects 2026 to see multimodal models, interactive video worlds, and autonomous agents capable of sophisticated cross-device tasks, all trends that are fully reflected in Gemini’s roadmap. Read more.

Grok 4.20 crushes Alpha Arena, turns $10K Into $12K in two weeks

Season 1.5 of Alpha Arena concluded with Grok 4.20 emerging as the clear leader. Over 14 trading days, the model generated a 12.11% return, turning $10,000 into $12,193, while GPT-5.1 and Gemini 3 posted losses of 3.4% and 5.7%, respectively.

Grok 4.20’s advantage came from its Situational Awareness mode, which allowed it to monitor competitor positions and PnL, then act preemptively. Notable trades included a high-leverage long in Palantir timed with a surge in retail sentiment. The model’s real-time edge was fueled by direct access to the X (Twitter) Firehose, processing roughly 68 million tweets per day to generate ultra-short-term signals, compared with competitors relying on delayed news or financial filings.

Looking ahead, Elon Musk signaled Grok 5, hinting at a multimodal framework integrating market data, social sentiment, and macro signals, alongside a Grok Trader API for institutions. Read more.

• Meta links its AI chatbot to CNN, Fox, USA Today, and Le Monde for live news

• STARFlow-V shows Apple can make stable generative video without relying on diffusion

• Limitless joins Meta as the company pivots fully to AI-enabled wearables

• U.S. data centers take years to build, China can do hospitals in a weekend, says Nvidia CEO

• Essential AI, founded by ‘Attention Is All You Need’ paper co-author, releases Rnj-1 with SWE-bench scores near GPT-4o

• 70% of creative professionals conceal AI use despite efficiency gains, new Anthropic research finds

• Meta pushes Phoenix mixed-reality glasses launch to 2027 to get the device right

• Field study finds visual AI ads increase click-through rates up to 19% while disclosure reduces impact

• Watch: AI -powered LaserWeeder G2 at work, removing weeds without any chemical use

• Google’s Gemini 3 and chip strategy position it to surpass OpenAI, says AI pioneer Geoffrey Hinton

• Watch: Jet-powered Volonaut airbike shows off real-world speeder bike flight demo

• Perplexity launches BrowseSafe to block 91 percent of prompt injection attacks on AI browser agents

• Chinese researchers unveil GeoVista, an open-source AI that can pinpoint photo locations nearly as accurately as Gemini 2.5 Flash

• GITAI showcases robots autonomously assembling 5-meter tower for future Moon and Mars habitats

• Poetiq tops ARC-AGI-2 leaderboard with 54% accuracy and half the cost

• Watch: China introduces humanoid traffic robot to manage busy intersections in cities

• Cohere CEO says company is on “right side” of AI bubble, contrasting Google’s warning

• Microsoft rolls out GPT-5.1, reminders, and projects to boost Copilot productivity

📄 Light-X: Generative 4D Video Rendering with Camera and Illumination Control

📄 NeuralRemaster: Phase-Preserving Diffusion for Structure-Aligned Generation  ​​

📄 SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs 

📄 Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

📄 PosterCopilot: Toward Layout Reasoning and Controllable Editing for Professional Graphic Design  

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!