💰 NVIDIA Bets $5B on Intel Comeback
Intel jumped 23% after NVIDIA invested $5B to co-develop chips, marking its best day in nearly four decades. The move follows U.S. and SoftBank investments aimed at reviving Intel’s fortunes. While manufacturing isn’t part of the deal, deeper collaboration could shift the U.S. chip landscape.
💸 DeepSeek Trained AI Model for Just $294K
Chinese firm DeepSeek claims its reasoning-focused R1 model was trained for only $294K—far less than U.S. rivals. The model ran for 80 hours on 512 NVIDIA H800 chips, with A100s used in early prep. The revelation may stir debate over China's AI cost edge amid scrutiny of its chip access.
🚀 Huawei Unveils AI Supercluster to Rival NVIDIA
Huawei says its Atlas 950 SuperCluster, set to launch in 2026, will surpass NVIDIA and xAI systems in compute power using over 500,000 Ascend chips. The move aligns with China’s push for tech self-reliance. Analysts warn of exaggeration, but Huawei’s AI ambitions remain formidable.
✨ Chrome Gets Smarter with Gemini AI Overhaul
Google is turning Chrome into an AI-first browser with Gemini tools that summarize tabs, enhance search, spot scams, and change passwords. A coming update adds agentic assistants to handle tasks like booking appointments. The rollout begins soon, but speed and cost remain question marks.
🤔 FRIDAY FACTS
Proteins That Don’t Exist in Nature
AI is now designing proteins that don’t exist in nature — built entirely from scratch, no evolutionary blueprint required. And one of these AI-engineered enzymes could help solve one of our planet’s biggest environmental problems.
Stick around to learn more! 👇
📽️ VIDEO
Ex-OpenAI CTO Reveals Plan to Fix LLMs’ Biggest Problem
Mira Murati’s Thinking Machines Labs tackles LLM non-determinism with fixes for reproducibility, boosting trust, debugging, and benchmark stability.
♿ ACCESSIBILITY
UK Study Finds Neurodiverse Workers Gain Most from AI, Not the Usual Productivity Crowd
A UK government study of Microsoft 365 Copilot users found that neurodiverse employees, including those with ADHD and dyslexia, reported significantly greater satisfaction and usefulness from AI tools than their neurotypical peers. While overall productivity gains were mixed, the AI assistant proved a game-changer for accessibility—offering embedded support that helped users write reports, manage tasks, and decode social cues with more confidence.
The study suggests AI’s most profound impact might not be in speeding up tasks we already do well, but in enabling participation for those previously underserved by traditional tools. As one user put it, “It’s leveled the playing field.” → Read the full article here.
🏭 INFRASTRUCTURE
Inside Microsoft’s AI Megafactory
Microsoft has unveiled its largest and most advanced AI datacenter yet—Fairwater, a 315-acre complex in Mt. Pleasant, Wisconsin purpose-built to train and deploy massive AI models at unprecedented scale. Housing hundreds of thousands of NVIDIA GPUs and delivering 10x the performance of the fastest existing supercomputers, Fairwater is optimized for AI from the silicon up: with two-story rack layouts, closed-loop liquid cooling, and integrated exabyte-scale storage. → Read the full article here.
🧩 REASONING
DeepSeek-R1 Ditches Human Labels to Teach AI How to Reason
A team of researchers has introduced DeepSeek-R1, a large language model trained not through human-annotated reasoning examples but via reinforcement learning driven solely by answer correctness. Unlike traditional models that mimic human reasoning, DeepSeek-R1 evolved its own reasoning behaviors—like self-reflection and verification—without being explicitly taught.
It outperforms supervised counterparts on complex STEM and coding tasks, demonstrating that LLMs can independently develop advanced problem-solving strategies. The project’s distilled models are open-sourced, potentially reshaping how the AI community builds future reasoning systems—less by example, more by evolution. → Read the full paper here.
🤖 MODELS
Meta’s MobileLLM-R1 Signals Pivot Toward Tiny, Task-Specific AI for the Enterprise
Meta has released MobileLLM-R1, a sub-billion parameter reasoning model optimized for on-device math, coding, and scientific tasks—marking a major step in the industry’s shift toward small language models (SLMs). Compact yet potent, MobileLLM-R1 outperforms peers like Alibaba’s Qwen3 on key benchmarks, but comes with a major caveat: a strict non-commercial license.
It joins a growing wave of lightweight, privacy-friendly, and cost-predictable AI tools from Google, NVIDIA, and Liquid AI, tailored for enterprise deployment. The monolithic “god model” is giving way to a modular, microservices-style AI architecture. → Read the full article here.
🛰️ NEWS
What Else is Happening
🦾 Figure Trains Robots with Human Video: Project Go-Big aims to build the largest humanoid dataset ever, teaching Helix to navigate and act by watching people, not simulations.
🎧 AirPods Do Live Translation: Apple’s new AI-powered earbuds translate speech in real time, making cross-language chats as easy as talking to Siri. (Paywall)
🔗 Google Lets You Share Gemini Gems: Custom AI assistants called Gems can now be shared like Google Docs, making it easier to collaborate and avoid rebuilding the same bots twice.
💵 Microsoft & NVIDIA Pledge $45B to UK AI: In a historic move, the tech giants will fund data centers and AI infrastructure—but critics warn of rising energy use and environmental fallout.
👥 Notion Launches AI Agents: New agents can analyze data, update pages, and automate tasks across tools—handling complex workflows with context from your Notion workspace.
🧰 TOOLBOX
Creative Image Generation, Smarter Productivity, and Curated Digital Inspiration
🖼️ CraiyonGenerate unique images from text prompts using free AI-powered tools. | ✅ BeeDoneBoost focus with smart to-dos, habits, and automated routines. | 🗂️ GloriaMundoExplore curated digital projects that inspire devs and creatives. |
🤔 FRIDAY FACTS
AI-Engineered Enzymes Are Learning to Eat Plastic
Using cutting-edge models like RFdiffusion to generate novel protein shapes, ProteinMPNN to design sequences, and AlphaFold to validate structures, scientists are now creating synthetic proteins with custom-tailored functions. One of the most compelling applications? Engineering enzymes to break down plastic.
Specifically, polyethylene terephthalate (PET) — the tough plastic in soda bottles and clothing fibers that can persist for centuries. AI-guided engineering has produced enzymes, like FAST-PETase, that degrade PET in days at moderate temperatures, without harsh chemicals.
Most of today’s breakthroughs come from AI-engineered versions of natural enzymes. But researchers are also beginning to design truly de novo enzymes — proteins built atom-by-atom with no natural ancestor — opening the door to powerful new recycling strategies. If successful, these efforts could dramatically reduce plastic waste and help address one of the world’s most stubborn environmental challenges.
🗒️ FEEDBACK
Help Us Get Better
What did you think of today's newsletter? |
That’s a Wrap!
📢 Want to advertise with Forward Future? Reach 550K+ AI enthusiasts, tech leaders, and decision-makers. Let’s talk—just reply to this email.
Thanks for reading today’s newsletter—see you next time!
Matthew Berman & The Forward Future Team
🧑🚀 🧑🚀 🧑🚀 🧑🚀
