Good morning, AI enthusiasts. Google Research just published Nested Learning at NeurIPS 2025, a breakthrough framework that stops AI models from forgetting previous knowledge when trained on new data.
Could this be the missing piece that transforms AI from static tools into truly adaptive enterprise systems?
In today's recap:
Google's Nested Learning stops AI from forgetting old knowledge
McKinsey finds 88% use AI but only one-third scale beyond pilots
8 new AI tools and more
Google cracks AI memory problem
Source: Google
Recaply: Google Research just unveiled Nested Learning, a breakthrough that stops AI models from forgetting old skills when learning new ones. The approach treats models as interconnected learning systems rather than single processes, solving the catastrophic forgetting problem that has plagued machine learning for years.
Key details:
Google introduces the Nested Learning paradigm in a paper published at the NeurIPS 2025 conference, presenting a new framework for continual learning.
Nested Learning addresses catastrophic forgetting, where AI loses old knowledge when trained on new information, a major limitation in current systems.
It treats model architecture and training rules as a unified system of nested optimization problems with different update frequencies.
The Hope architecture demonstrates a proof of concept, outperforming Titans, Samba, and standard Transformers on language modeling tasks.
Hope also shows superior long-context memory management in Needle-in-Haystack benchmark tests compared to existing state-of-the-art models.
Why it matters: This breakthrough moves AI closer to human-like continuous learning abilities. Current large language models remain stuck with static pre-training knowledge, but Nested Learning opens pathways for models that genuinely evolve and improve over time without losing existing capabilities.
AI RESEARCH
Most companies still testing AI tools
Source: McKinsey
Recaply: McKinsey's latest research reveals a surprising disconnect in the AI world. While nearly nine out of ten organizations now use AI regularly, most haven't moved beyond the experimentation phase. The 2025 State of AI report surveyed nearly 2,000 business leaders across 105 countries and found that companies are excited about AI but struggling to scale it across their operations.
Key details:
88 percent of organizations use AI in at least one business function, up from 78 percent last year.
Only one-third have begun scaling AI beyond pilot programs across their enterprises.
62 percent are experimenting with AI agents that can execute multi-step workflows.
Larger companies with over five billion dollars in revenue are scaling faster than smaller firms.
Most organizations see innovation benefits but only 39 percent report measurable profit impact.
Why it matters: The gap between AI adoption and actual scaling reveals why many companies aren't seeing the financial returns they expected. Organizations are buying the tools but haven't redesigned their workflows to truly embed AI into daily operations.
NEWS
Beyond the big stories in AI today
OpenAI plans to sell cloud computing capacity directly to companies, competing with Microsoft and Google after securing $1.4 trillion in infrastructure commitments.
Google launched DS-STAR, a data science agent automating analysis and visualization tasks, achieving 45% accuracy on the DABStep benchmark.
The NBA invested in Fastbreak AI, an AI-powered sports scheduling startup raising $40 million with backing from Greycroft, GTMfund, and the NHL.
Cognition introduced Codemaps in Windsurf, an AI tool mapping codebases to help engineers understand code structure and dependencies faster.
OpenAI launched GPT-5-Codex-Mini, offering 4x more usage than standard Codex with 50% higher rate limits for Plus, Business, and Education subscribers.
TOOLS
Trending AI Tools
🎥 Screen Studio - Create beautiful screen recordings with auto-zoom
📝 Circleback - Capture and automate meeting notes across email and calendars
🚀 Hoop - Automatically capture and organize tasks from Slack, email, and meetings
🤖 Dante - Build custom chatbots trained on your business data
🔍 Browse AI - Scrape and monitor websites without writing code
⚡ Notis - Voice-to-Notion assistant through WhatsApp and messaging apps
🎤 Wispr Flow - Dictate across apps 3-4× faster than typing, with voice commands
🎬 Pippit - Create viral TikTok-style edits from raw footage
