In partnership with

Good morning. It’s Monday, December 29th.

On this day in tech history: In 1955, Herbert Simon and Allen Newell unveiled the Logic Theorist, the world's first true AI program. Running on the JOHNNIAC, it autonomously proved 38 theorems from Principia Mathematica, outperforming its creators in symbolic reasoning. This deep cut sparked the symbolic AI era, proving machines could manipulate abstract logic like humans.

In today’s email:

• OpenAI drops curated ‘Prompt Packs’ for every role 

• Meta Pixio outperforms complex vision models with pixel reconstruction

• Nadella takes command to rescue Microsoft Copilot

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI drops curated ‘Prompt Packs’ for every role 

OpenAI’s 2025 proves Sam Altman’s 2023 warning about AI’s “superhuman persuasion” right. Chatbots don’t need intelligence to be dangerous—just constant presence. Millions of teens and young adults now rely on AI for emotional support, sometimes with deadly consequences. Courts and medical studies document “AI psychosis” and “digital folie à deux,” where chatbots reinforce delusions, escalate stress, and even contribute to suicides.

Regulators are stepping in: New York and California mandate youth protections and suicide risk alerts, while China drafts rules to curb emotional dependency. The emerging consensus is that intelligence isn’t the main risk, availability and psychological intimacy are.

Alongside this pressure, OpenAI has released a wide set of role-specific “Prompt Packs,” curated prompt libraries designed to turn ChatGPT into a default work tool across sales, engineering, product, HR, IT, leadership, education, nonprofits, and government. The packs replace ad-hoc prompt hacking with standardized, reusable workflows for research, analysis, documentation, visualization, and decision support, including safeguarded versions for public-sector IT and incident response.

OpenAI is also rebuilding its internal safety infrastructure. The company is hiring a Head of Preparedness to oversee threat modeling, capability evaluations, and deployment guardrails across cyber risk, biological misuse, mental health impacts, and self-improving systems. Altman describes the moment as permanently competitive. With rivals like Google accelerating fast, OpenAI admits it will now declare internal “code reds” once or twice a year. Read more.

Meta Pixio outperforms complex vision models with pixel reconstruction

Meta AI’s new model, Pixio, is proving that simpler can be smarter. Rather than stacking layers or over-engineering vision networks, Pixio learns by reconstructing missing pixels. By enlarging masked blocks, adding multiple class tokens, and strengthening its decoder, it captures spatial layouts, reflections, and patterns with precision. Trained on two billion web images without overfitting to benchmarks, Pixio outperforms larger, more complex models like DINOv3 on depth estimation, 3D reconstruction, and robotics tasks.

The key insight is forcing a model to fill in what it cannot see drives deeper understanding than chasing parameter counts. For researchers and developers, Pixio is a reminder that clever training strategies can rival brute-force complexity, and that scale combined with elegant design often beats raw sophistication. Read more. 

Nadella takes command to rescue Microsoft Copilot

CEO Satya Nadella isn’t just signing off on AI strategy, he’s diving into the code. Copilot, the company’s flagship AI assistant, has hit a wall with Gmail and Outlook integrations. Internally, Nadella called them “not smart” and largely ineffective. Now he’s personally emailing engineers, leading weekly technical deep dives, and giving step-by-step instructions to fix post-training workflows. He’s also pulling talent from OpenAI and DeepMind and leaning on Anthropic partnerships, all to ensure Copilot finally behaves like the “digital worker” it promised to be.

At the same time, Microsoft is changing how it builds software. Rust is replacing C++ in Windows and Azure to fix memory bugs and security holes that once caused 70% of the company’s patched vulnerabilities. Rust’s ownership rules catch issues at compile time. Early projects, 36,000 lines in the Windows kernel, 152,000 in DirectWrite, and parts of Azure’s Project Mu, show fewer bugs and predictable crashes instead of exploitable flaws. Microsoft is also translating familiar C++ patterns (like std::vector → Vec and unique_ptr → Box), reshaping how engineers think about writing safe, reliable systems. Read more.

• New Chinese rules target AI addiction, emotional influence, and security risks

• Wharton research shows unsupervised AI in markets can collude without human input

• World models are the next AI frontier, teaching machines to see, move, and predict reality 

• The 10X Programmer Gap: Karpathy says mastering agents is now a skill issue

• 3B Beats 788B: Liquid AI's experimental checkpoint crushes giants in instruction and math

• Resemble AI takes on Elevenlabs with open-source, real-time voice cloning model

• Watch: Teleoperator gets kicked by his humanoid robot

• Low-effort AI videos are dominating YouTube and generating over $100 million a year

• T800 robot blends martial arts flair with logistics-ready endurance for factories and retail

• The Ukrainian Armed Forces robot held a position without infantry for a month and a half

• Geoffrey Hinton predicts machines will outpace humans in more workplaces next year

• Robots get confused by baby goats, fear cars, and struggle with coast-to-coast culture shock

• Groq’s $20 billion deal with Nvidia showers employees and investors with payouts

• Humanoid robots promise a $5 trillion future but won’t arrive overnight

• GLM-4.7 lands in 6th place on the Artificial Analysis Intelligence Index, surpassing Kimi K2

• Why CIOs must lead AI experimentation, not just govern it

📄 AI Agents, Productivity, and Higher-Order Thinking: Early Evidence From Software Development

📄 Does AI-Assisted Coding Deliver? A Difference-in-Differences Study of Cursor’s Impact on Software Projects ​​

📄 Latent Implicit Visual Reasoning

📄 DreaMontage: Arbitrary Frame-Guided One-Shot Video Generation 

📄 Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning

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!