AI News Digest

2025-12-29

Dec 29, 2025

It’s been a relatively quiet week in AI—no earth-shattering announcements or dramatic pivots—but that’s precisely what makes it interesting. Beneath the surface calm, we’re seeing the steady drumbeat of the industry’s maturation: technical advancements like improved LLM safety guardrails and interpretability tools, a continuous flow of capital into everything from avatar startups to energy infrastructure, and perhaps most tellingly, the unglamorous but significant march of enterprise adoption, where companies like Salesforce are quietly adding thousands of customers while everyone else obsesses over bubble talk. This is what the AI revolution actually looks like when it moves from hype cycle to infrastructure—less fireworks, more foundation-building.

📰 General News

Microsoft bets on AI to modernize Windows

Microsoft engineer Galen Hunt announced an ambitious goal to eliminate all C and C++ code from Microsoft by 2030, replacing it with Rust using AI-powered translation tools. The team aims for “1 engineer, 1 month, 1 million lines of code” conversion rates. However, Hunt later clarified this is just a research project, not an official company mandate. Microsoft has been gradually adopting Rust since 2023, starting with parts of the Windows kernel, citing better memory safety and security compared to legacy languages.

While everyone talks about an AI bubble, Salesforce quietly added 6,000 enterprise customers in 3 months

While critics debate whether AI is overhyped, Salesforce’s Agentforce platform just added 6,000 enterprise customers in three months, bringing its total to 18,500 companies and $540M in annual recurring revenue. The platform now processes 3 billion automated workflows monthly. Real-world deployments at Williams-Sonoma and Engine are already showing measurable returns, suggesting enterprise AI adoption is accelerating faster than the skeptics realize.

One in a million: celebrating the customers shaping AI’s future

OpenAI hit one million business customers, marking rapid enterprise adoption since launching its business products. The milestone includes major names like PayPal using AI for customer service, Virgin Atlantic for flight operations, BBVA for banking automation, and Moderna for drug discovery research. Cisco, Canva, and thousands of other companies are now building AI into their core workflows. The announcement signals how quickly AI tools have moved from experimental projects to production systems at scale.

💰 BigMoneyDeals

Marissa Mayer’s new startup Dazzle raises $8M led by Forerunner’s Kirsten Green

Marissa Mayer has shut down her struggling photo-sharing startup Sunshine after six years to launch Dazzle, an AI personal assistant company that just raised $8 million at a $35 million valuation. The seed round was led by Forerunner’s Kirsten Green, known for backing Warby Parker and Chime. Mayer, former Yahoo CEO and Google employee #20, admitted Sunshine’s problems were too mundane and the product never gained traction despite raising $20 million. Dazzle will emerge from stealth early next year, with Mayer aiming to build something with the impact of Google Search or Maps.

Lemon Slice nabs $10.5M from YC and Matrix to build out its digital avatar tech

Lemon Slice just raised $10.5M from Y Combinator and Matrix Partners to fix what its founders call the “creepy and stiff” problem plaguing digital avatars. The startup’s new Lemon Slice-2 model creates video avatars from a single image that can livestream at 20fps on a single GPU. Companies can embed these avatars into their sites with one line of code to handle customer service, education, or mental health support. The 20-billion-parameter diffusion model works for both human and non-human characters, setting it apart from competitors like HeyGen and Synthesia.

Alphabet to buy Intersect Power to bypass energy grid bottlenecks

Alphabet is acquiring Intersect Power for $4.75 billion to solve a critical AI infrastructure problem: getting enough electricity to power data centers. Instead of waiting on overwhelmed utility companies, Google will build data centers directly next to wind, solar, and battery facilities. The deal builds on an $800 million investment Alphabet made in Intersect last year and includes future development projects, with the first locations expected online by late 2026.

Tesco signs three-year AI deal centred on customer experience

Tesco has signed a three-year partnership with French AI startup Mistral to embed AI across its operations, from delivery route optimization to personalized Clubcard offers. The UK supermarket giant is establishing an internal AI lab to test tools before wider rollout, focusing on reducing repetitive work for staff and improving customer service. Tesco has doubled its tech team over five years and already uses AI for demand forecasting and supply planning. The deal makes Tesco the first major UK retailer to partner with Mistral, Europe’s only large language model developer.

🔬 Technical

AprielGuard: A Guardrail for Safety and Adversarial Robustness in Modern LLM Systems

ServiceNow released AprielGuard, an 8B parameter safety model designed to protect modern AI agent systems from both traditional risks (toxicity, hate speech, misinformation) and sophisticated attacks like prompt injection, memory poisoning, and multi-agent exploits. Unlike traditional safety filters that only check individual messages, AprielGuard monitors entire agentic workflows including tool calls, reasoning traces, and multi-turn conversations. The model runs in two modes: a fast classification mode for production and a reasoning mode that explains its decisions.

Announcing Gemma Scope 2

Google DeepMind just dropped Gemma Scope 2, a collection of Sparse Autoencoders (SAEs) and transcoders designed to crack open how the Gemma 3 model family actually works under the hood. The tools work with models up to 27B parameters and are built for mechanistic interpretability research, letting researchers peek inside the black box of neural networks. Everything’s available on HuggingFace, plus there are interactive demos on Neuronpedia where you can explore what these models are really learning.

This AI finds simple rules where humans see only chaos

Duke University researchers built an AI that discovers simple mathematical rules governing chaotic systems like weather patterns, electrical circuits, and biological signals. The system reduces thousands of variables into compact equations scientists can actually read and use. In tests across physics, climate science, and neural circuits, it produced models 10 times smaller than previous machine learning methods while maintaining accurate long-term predictions. The approach extends a 1930s mathematical theory by using deep learning to identify hidden patterns in how systems change over time.

Closing Thoughts

This week reminded us that transformative change doesn’t always arrive with fanfare—sometimes it’s the steady hum of technical progress, enterprise deals, and capital allocation that reshapes the landscape. While the headlines may have been quieter than usual, the fundamentals continue their relentless march forward: models getting sharper, checkbooks opening wider, and boardrooms finally moving past the “exploration phase.” In AI, silence often just means everyone’s too busy building to tweet about it.

Stay tuned for next week’s edition, where we’ll presumably cover another round of funding announcements while pretending we’re surprised that throwing billions at the problem keeps yielding results. YAI 👋

Disclaimer: I use AI to help aggregate and process the news. I do my best to cross-check facts and sources (BTW: sources are available on-demand, or you could just google it 😃 ), but misinformation may still slip through. Always do your own research and apply critical thinking—with anything you consume these days, AI-generated or otherwise.