In partnership with
Dear Readers,
Today we have something very special for you:
NVIDIA is kicking off the Cosmos Cookoff, a global virtual challenge inviting developers and researchers to build real-world physical AI solutions using NVIDIA Cosmos open world foundation models. With $11,000+ in prizes, including an NVIDIA DGX Spark™ and RTX™ 5090 GPU, the event to build reasoning AI models and agents, runs from Jan 29 to Feb 26. Backed by expert mentorship, livestreams, and hands-on recipes for robotics, autonomous vehicles, and video analytics! It’s a high-impact chance to benchmark skills, ship a real project, and stand out in the fast-moving physical AI space.
And as if that weren't cool enough, we are pleased to tell you that on February 10th, our Editor-in-Chief Kim Isenberg, who now has a community of over 100,000 followers on X, will be live-stream interviewing top NVIDIA researcher Ming-Yu about Cosmos!
Don’t miss out! Add this to your calendar: February 10, 9 a.m. PT —Livestream: Intro to NVIDIA Cosmos | With Ming-Yu ft. Superintelligence
In Today’s Issue:
🛡️ Internal sources report Satya Nadella is personally stress-testing Anthropic’s Cowork agents
⚡ A new "Quantization-Aware Distillation" (QAD) technique delivers 4x higher FLOPS on Blackwell GPUs
👁️ Google DeepMind debuts Agentic Vision
🚀 Ex-OpenAI reasoning lead Jerry Tworek founds Core Automation
✨ And more AI goodness…
All the best,
🤖 Microsoft Scrambles as Anthropic Advances
Anthropic’s new Cowork AI - capable of automating tasks across apps like Excel, PowerPoint, and Slack - has sparked urgency inside Microsoft, with leaders fearing it outpaces 365 Copilot in real-world productivity. CEO Satya Nadella is personally testing rival AI agents and pushing teams to accelerate development, even leveraging models from Anthropic itself, as the battle to own the “digital co-worker” heats up—especially on Windows, where Microsoft sees a strategic edge.
🚀 NVIDIA Supercharges NVFP4 Inference
NVIDIA just released NVFP4 for Nemotron 3 Nano, delivering up to 4× FLOPS over BF16 on Blackwell GPUs while preserving near-BF16 accuracy through quantization-aware distillation (QAD). By quantizing both weights and activations, the 30B-A3B MoE model achieves ~1.7× memory savings over FP8, making powerful local inference viable even on consumer GPUs like the RTX 5090. This single-stage QAD approach simplifies deployment and avoids the instability of traditional post-training pipelines—big efficiency wins with minimal accuracy tradeoffs.
🤖 Gemini Gains Agentic Vision Power
Gemini 3 Flash introduces Agentic Vision, turning image understanding into an active, step-by-step investigation by combining visual reasoning with live code execution. This Think–Act–Observe loop lets the model zoom, crop, annotate, and calculate directly on images—delivering a consistent 5–10% boost across vision benchmarks and reducing hallucinations in complex visual tasks. For developers, it unlocks more reliable inspection, visual math, and grounded answers across products built with Google DeepMind tech.
Why are weather forecasts so hard to get right? Microsoft Azure and NVIDIA
$1 Billion to Reinvent AI Training
The Takeaway
👉 Ex-OpenAI reasoning lead Jerry Tworek launches Core Automation, targeting $500M–$1B in funding to develop fundamentally new AI training methods
👉 The startup bets on "continual learning"—AI that updates from real-world deployment rather than static training—a capability no current frontier model possesses
👉 Tworek plans to rethink core architecture: moving beyond transformers, questioning gradient descent, and building models needing 100x less training data
👉 Core Automation joins a growing wave of "neolabs" (Humans&, Thinking Machines Lab) challenging incumbents like OpenAI and Anthropic on foundational reserach
The mastermind behind OpenAI's reasoning revolution just left to build something radically different. Jerry Tworek, the architect of groundbreaking models like o1 and o3, has founded Core Automation - and he's chasing up to $1 billion to reinvent how AI learns. Tworek wants to build AI that actually learns on the fly from real-world experience. Today's models can't do this. They're trained once and deployed frozen. Core Automation aims to develop "continual learning" - AI that adapts dynamically, requiring 100 times less data than current systems. Tworek is even questioning fundamental techniques like gradient descent and the transformer architecture itself.
This isn't just another well-funded startup. Tworek led the teams that gave us GitHub Copilot, GPT-4's post-training, and inference-time reasoning. When researchers of this caliber say the industry needs a reset, people listen. The timing is fascinating: Anthropic's Sholto Douglas recently predicted continual learning will be solved "in a satisfying way" in 2026. Is something brewing we don't yet see?
Why it matters: Continual learning represents perhaps the biggest unsolved challenge in AI - models that grow smarter through experience rather than static training. If Tworek succeeds, it could fundamentally change how AI systems are built and deployed across every industry.
Sources:
🔗 https://www.theinformation.com/articles/ex-openai-researchers-startup-targets-1-billion-funding-develop-new-type-ai?rc=bfliih
What Will Your Retirement Look Like?
Planning for retirement raises many questions. Have you considered how much it will cost, and how you’ll generate the income you’ll need to pay for it? For many, these questions can feel overwhelming, but answering them is a crucial step forward for a comfortable future.
Start by understanding your goals, estimating your expenses and identifying potential income streams. The Definitive Guide to Retirement Income can help you navigate these essential questions. If you have $1,000,000 or more saved for retirement, download your free guide today to learn how to build a clear and effective retirement income plan. Discover ways to align your portfolio with your long-term goals, so you can reach the future you deserve.
Triple Therapy Beats Pancreatic Cancer For The First Time
After six years of relentless research, Spanish scientist Mariano Barbacid and his team at CNIO have achieved something extraordinary: completely eliminating pancreatic tumors in mice - with no recurrence and minimal side effects. This is huge.
The triple combination therapy uses daraxonrasib (a KRAS inhibitor), afatinib (an EGFR/HER2 kinase inhibitor), and SD36 (a selective STAT3 PROTAC) to simultaneously attack three survival mechanisms that tumors use to resist treatment. This approach prevents the resistance that typically appears when targeting only one pathway - a problem that has plagued pancreatic cancer treatment for decades.
KRAS mutations appear in roughly 90% of pancreatic cancers, making this one of the most data-rich targets for computational drug discovery and AI-assisted research. The multi-target strategy mirrors how machine learning models increasingly tackle complex problems: not with single solutions, but through ensemble approaches.
Human clinical trials could begin within three years if regulatory processes proceed smoothly. Could AI accelerate this timeline even further?
Pancreatic cancer has a five-year survival rate below 10%. This breakthrough proves that even the most treatment-resistant cancers can potentially be defeated through strategic combination therapies - a principle that could reshape how we approach other deadly diseases.
Equipment policies break when you hire globally
Deel’s latest policy template on IT Equipment Policies can help HR teams stay organized when handling requests across time zones (and even languages). This free template gives you:
Clear provisioning rules across all countries
Security protocols that prevent compliance gaps
Return processes that actually work remotely
This free equipment provisioning policy will enable you to adjust to any state or country you hire from instead of producing a new policy every time. That means less complexity and more time for greater priorities.
