Good morning. It’s Friday, January 2nd.

On this day in tech history: In 1980, Digital Equipment Corp deployed XCON (eXpert CONfigurer), an expert system that configured VAX computers, saving millions annually. This niche commercial success proved rule-based AI's real-world value amid hype, configuring 80% of orders flawlessly. It used 10,000+ rules, foreshadowing knowledge graphs in today's LLMs.

In today’s email:

• The Screen Is Dying – And OpenAI Is Building What Comes Next

• Deepseek Publishes Fundamental Breakthrough in Transformer Architecture

• Alibaba Bets on Openness with Qwen-Image-2512, a Rival to Google’s Nano Banana Pro 

• 5 New AI Tools

• Latest AI Research Papers

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

The Screen Is Dying – And OpenAI Is Building What Comes Next

The biggest roadblock for AI isn't compute or data. It's the screen-heavy, clunky interfaces that still demand our full attention and OpenAI is trying to fix that. The company is now restructuring around one core problem: their audio models aren't yet good enough for real-time agents or audio-first hardware.

Per reporting from The Information, OpenAI has merged engineering, product, and research teams over the past two months to tackle persistent gaps in accuracy, latency, and responsiveness, areas where voice still lags behind text.

The merged team is creating a new audio system for fast, two-way conversation, replacing rigid back-and-forths and paving the way for hands-free AI devices. Former Character.AI researcher Kundan Kumar is leading the effort, with an internal release target of Q1 2026.

This work is directly tied to OpenAI’s hardware roadmap following its $6.5 billion acquisition of Jony Ive–backed startup io. Greg Brockman has already framed 2026 around enterprise agents and scientific acceleration with voice is emerging as the control layer for both.

But dominance has a price tag. OpenAI is averaging $1.5M per employee in stock compensation, equity projected to eat 46% of 2025 revenue. Compensation costs are set to balloon another $3B annually through 2030. But in the escalating talent arms race against Meta's nine-figure packages, retention isn't optional – it's the strategic moat. Read more.

Deepseek Publishes Fundamental Breakthrough in Transformer Architecture

DeepSeek is kicking off 2026 by rewriting the rules of large-model training. The Chinese AI start-up, led by founder Liang Wenfeng, rolled out its Manifold-Constrained Hyper-Connections (mHC), a tweak on residual connections and ByteDance’s Hyper-Connections that stabilizes gradients across massive networks.

mHC uses manifolds which are complex, multi-dimensional structures, to keep gradient flow robust, even in LLMs with 3B, 9B, and 27B parameters. Tests show faster learning, stronger performance across benchmarks, and hardware overhead of just 6.27%, making high-scale training far more practical.

In short: you get deep, stable models without breaking memory or compute budgets.

Publishing the work on arXiv, DeepSeek is signaling exactly how its next major model will be built, blending transparency with serious technical edge. Read more.

Alibaba Bets on Openness with Qwen-Image-2512, a Rival to Google’s Nano Banana Pro 

Alibaba is making a clear bet on where enterprise AI is headed. Its newly released Qwen-Image-2512 positions itself as an open-source counterweight to Google’s Gemini 3 Pro Image, which recently reset expectations for text-heavy, production-ready visuals but locked those gains behind a proprietary cloud stack.

Qwen’s pitch is actually simple. Match the capabilities that actually matter with clean text rendering, layout control, and realism, while removing vendor lock-in. Released under an Apache 2.0 license, Qwen-Image-2512 can be self-hosted, modified, and deployed commercially, giving enterprises control over cost, data residency, and localization.

The model improves human realism, material textures, and multilingual text accuracy, and ranks as the top open-source image model in Alibaba’s human evaluations. For teams that want convenience, Alibaba also offers managed inference at $0.075 per image. Read more.

• DeepMind’s Grefenstette charts progress in autonomous, aligned AI agents

• Chinese AI Startups Beat Silicon Valley to the Public Markets in a Massive IPO Wave

• How Meta engineered ad transparency to buy time with regulators

• 2026 set to be the year of mega-IPOs as SpaceX, OpenAI, and Anthropic prepare $3 trillion market debut

• IBM: 2026 marks the rise of super agents and human-in-the-loop enterprise AI

• Union of Concerned Scientists: ‘US and China run parallel AI races with wildly different goals’

• The next leap for AI could come from learning inside language-driven virtual worlds

• Moonshot AI secures $500 million to scale computing power and build the next Kimi model

• AI forced a white-collar worker into manual labor

• A step-by-step look at how ChatGPT, Gemini, and other LLMs generate text

• China rolls out super AI science network to challenge Trump's Genesis Mission

• Watch: China’s new drone system tackles fires autonomously

• 40B model from iQuest leads the pack in 2026 launches, now on Hugging Face

• Epoch AI is using satellites to reveal where America’s massive AI compute lives

• Watch: Drone-based solar cleaning

• Europe's banks prepare to cut 200,000 jobs as AI automates back-office work

• The Top AI Stories of 2025 by IEEE Spectrum

• Four AI research trends enterprises need to track in 2026

• xAI introduced Grok for Business and Grok Enterprise plans. Grok for Business comes at $30 per seat

• Creative malware turns robots.txt into a frontline against uncontrolled AI scraping

• Open-source computer vision just got faster with OpenCV 4.13

• A humanoid robot weighs in on the AI boom without calling it a bubble

• Xi touts China’s AI, chip wins in triumphant New Year’s speech

• China unveils the world’s first modular AI service space powered by a humanoid robot

• Instagram’s Adam Mosseri says in a world of flawless AI images, raw imperfection becomes the new proof

• TSMC begins mass production of 2nm chips as AI demand pushes semiconductor limits

• Investors predict AI is coming for labor in 2026

• Brookfield launches its own cloud arm to lease AI-ready data centers and control the full value chain

• Emerge’s Top 10 WTF AI Moments of 2025 

📄 DreamOmni3: Scribble-based Editing and Generation

📄 Yume-1.5: A Text-Controlled Interactive World Generation Model ​​

📄 Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion

📄 Evaluating Parameter Efficient Methods for RLVR

📄 UltraShape 1.0: High-Fidelity 3D Shape Generation via Scalable Geometric Refinement

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!