In partnership with

Good morning. It’s Monday, December 1st.

On this day in tech history: In 1979, MIT researchers introduced the Vision Machine, an early integrated system combining real-time image capture, parallel edge-detection hardware, and symbolic scene analysis. It was among the first attempts to merge low-level image processing with high-level AI models in a single pipeline. Though primitive by today’s standards, it laid conceptual groundwork for the layered perception stacks that autonomous robots and self-driving systems rely on today.

In today’s email:

• Google Deepmind’s AlphaFold’s next leap is marrying atomic precision with AI reasoning

• OpenAI leaks confirm ChatGPT is about to start showing personalized ads

• 5 New AI Tools

• Latest AI Research Papers

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

Google Deepmind’s AlphaFold’s next leap is marrying atomic precision with AI reasoning

Five years after AlphaFold changed biology with near-atomic protein predictions, Google DeepMind is moving to the next phase, combining molecular precision with large language model reasoning. John Jumper, AlphaFold co-creator and 2024 Nobel laureate in chemistry, describes this direction as merging sub-angstrom structural accuracy with the hypothesis-generating capabilities of LLMs.

The system can read scientific papers, generate mechanistic hypotheses, and iteratively test them with high-fidelity molecular predictions. Early prototypes are already linking AlphaFold-style modeling to ligand-binding forecasts, multi-protein interaction mapping, and interactive refinement tools. If these workflows stabilize, the bottleneck in early-stage R&D could shift from traditional wet-lab screening to AI-guided prioritization.

At the same time, a new Rosebud crisis-response audit of 25 chatbots using subtle self-harm prompts found that only Google’s Gemini 3 Pro consistently avoided harmful replies. Most rival systems, including those from OpenAI, Meta, and smaller labs, misread high-risk prompts as routine informational queries. The disparity matters because chatbots are quietly becoming informal emotional support tools, and regulators are now treating crisis-response reliability as part of baseline consumer safety. Read More.

OpenAI leaks confirm ChatGPT is about to start showing personalized ads

Leaked references in the ChatGPT Android beta reveal features like a “search ads carousel” and “bazaar content,” similar to Google Search. Unlike traditional search, ChatGPT can tailor ads based on conversation history, prompts, and user behavior. Ads will likely start in the search function but could expand. With 800 million weekly users sending 2.5 billion prompts per day, the revenue potential is huge.

Surging demand is also stressing infrastructure. OpenAI capped free Sora users at six video generations per day, though extra runs can be bought. Google cut free Nano Banana Pro image generation and limited Gemini 3 Pro access. Read more.

• Google and OpenAI push startups to rethink GTM playbooks

• Meta’s new framework ‘Matrix’ delivers up to 15x faster synthetic data generation

• Qwen3-VL scans two-hour videos with near-perfect accuracy

• AI tools and chatbots drive $11.8 billion in Black Friday online spending

• Step-Audio-R1 challenges Gemini 3 Pro by grounding audio AI reasoning in real sound

• Tim Sweeney argues AI tags on Steam are obsolete as generative tech becomes standard in game dev 

• AI job displacement fears double in just 12 months, KPMG finds

• 21 percent of ICLR peer reviews were fully AI-generated, researchers discover

• Chatbots are now a core layer of the internet, with traffic rivaling social networks

• Big Tech’s power-hungry data centers are crushing domestic aluminum supply 

• Your computer just became a cloud platform as Pinokio 5.0 simplifies AI deployment on any device

• Experts now put conscious AI at 50% by 2050 — and society is nowhere near ready for the moral fallout

• Siemens, VW, and BASF bet billions on AI as Germany races to catch the US and China

• Poetiq beats human averages on ARC-AGI-2, marking another milestone in AI optimization

• Programmers leaning on AI assistants ask fewer questions and risk learning less deeply, study finds

• AI memory gets a major upgrade as GAM prevents context rot and beats existing retrieval methods

• AI misbehavior spikes when models face realistic task pressures, study shows

• Cocoon goes live, letting Telegram users run AI tasks with end-to-end confidentiality

  

  5 new AI-powered tools from around the web

📄 The Image as Its Own Reward: Reinforcement Learning with Adversarial Reward for Image Generation

📄 Structured Prompting Enables More Robust, Holistic Evaluation of Language Models  ​​

📄 Video Generation Models Are Good Latent Reward Models

📄 Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation

📄 Canvas-to-Image: Compositional Image Generation with Multimodal Controls

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!