Have you heard these exciting AI news? - November 14, 2025 AI Updates Weekly

Agents Write Code for Task Execution [00:09]

• A new trend sees AI agents writing code to execute tasks, rather than being directly instructed, a method demonstrated by Entropic.
• This approach dramatically reduces token usage; for instance, 150,000 tokens can be replaced by 2,000 by shifting execution to code.
• This method is described as being 98.7% more cost-effective and results in more precise, reliable, and faster execution.
• Entropic provides sandboxes and templates for implementing this code execution approach, with code written in TypeScript and Python representing tools.

This becomes a new trend.

Coding Leaderboards and Multimodal Models [00:41]

• Leaderboards updated on November 9th show Claude models dominating coding tasks.
• While some paid models like Quen Max are high, Claude holds top positions.
• Gemini 2.5 Pro leads in chat, with GPT 4.5 and other GPT models following.
• World Labs has introduced Marble, a multimodal frontier world model capable of generating fully navigable 3D environments from various input types, including text, images, and videos.

So their idea is kind of similar to Yan Lern's idea that uh people uh like regular human brain operates in the three-dimensional world but large language models only operate in the world of descriptions.

OpenAI GPT 5.1 and Job Platform [05:34]

• OpenAI has released GPT 5.1 for paying users, offering "instant" and "thinking" versions.
• The "instant" model is described as warmer, more conversational, playful, empathetic, and approachable.
• The "thinking" model is designed to make better decisions, including token allocation for responses, prioritizing speed for simple questions and more thought for complex ones.
• OpenAI is also reportedly developing its own version of LinkedIn, a jobs platform, expected mid-next year.

So people enjoy working with it.

Google's Nested Learning and AI Data Understanding [07:19]

• Google's "nested learning" is designed to solve the catastrophic forgetting problem, where training on new information causes models to lose previously learned knowledge.
• Unlike human brains, AI models have knowledge locked into their context window or pre-training data, hindering continuous learning.
• Nested learning treats model architecture and training algorithms as the same concept, viewing models as systems of smaller, nested optimization problems with different learning speeds.
• Google's "HOPE" system integrates a continuum memory system with a spectrum of memory updating frequencies, from milliseconds to months, demonstrating superior performance on benchmarks.
• Both Gemini and Anthropic now support JSON schema, enabling models to return results in specified formats, which is crucial for integrating LLMs into production software systems.

So this is research paper but a very interesting approach.

Advanced RAG Systems and AI Image Generation [10:18]

• World Model RAG (Retrieval-Augmented Generation) from UC offers state-of-the-art performance, with up to 20% improvement over existing methods.
• It uses a neuro-inspired framework with "operator" and "reconciler" components, creating local semantic graphs from text chunks to reconstruct complete episodic summaries.
• Google's "Nana Banana 2" is an upcoming ultra-light AI image generation model delivering high-quality visuals at exceptional speed, featuring native 2K output, improved text rendering, and multilingual support.
• Nana Banana's multi-step workflow involves planning, generation, review, and error correction, positioning it as a potential "Photoshop killer."

The architecture represents a fundamental shift from bag of chunks retrieval to persistent structured memory.

AI Coding Agents and Economic Shifts [12:33]

• Verdant AI offers a multimodal orchestration platform combining models like GPT5 and Claude, featuring long context, persistent memory, and automated verification.
• Tools are emerging to analyze data directly within Excel and Google Sheets using natural language prompts.
• Yann LeCun is reportedly planning to leave Meta, a significant event for the AI research community.
• Microsoft's CEO, Mustafa Suleyman, published a manifesto for "humanist super intelligence," emphasizing AI designed to work with humans within limits.
• China's ByteDance has launched "Duba seed code," an affordable AI coding agent priced at approximately $130 for the first month and $5 thereafter, showing state-of-the-art performance on coding benchmarks.
• Lovable, a European company, has gained about 8 million users and is valued at $6 billion, with over half of Fortune 500 firms using their platform for coding tasks.

This is a disaster for Meta.

AI's Impact on Jobs and the Market [17:31]

• Meta has open-sourced an Automatic Speech Recognition (ASR) model for 1,500 languages, available on Hugging Face and GitHub.
• The "Magnificent Seven" companies (Microsoft, Nvidia, Google, Apple, Tesla, Amazon, Meta) are driving a significant portion of the S&P 500's growth.
• Microsoft is facing a bottleneck with GPUs, having more than they have electricity to power them, highlighting electricity as a new constraint.
• Nvidia is investing in startups like Core, Figure AI, and OpenAI, which then purchase Nvidia GPUs, creating a symbiotic business relationship.
• Research suggests LLMs are injective and invertible, meaning they preserve complete input information in their hidden representations, with implications for transparency and interpretability.

Microsoft has more GPUs than electricity to run them.

AI Project Success and Browser Automation [20:16]

• Studies present conflicting data on AI project success rates: one claims 95% fail to deliver ROI, while another shows 75% of companies see positive returns.
• Success is linked to augmenting existing processes with ready-to-use AI models, rather than attempting large-scale business transformations.
• Retriever.ai, a browser extension that automates operations within the browser, has released version two, showing significant improvements.
• Google Collab is now available as a VS Code extension, allowing users to connect to their Collab accounts directly from their VS Code editor for enhanced convenience.

So if you want to report success do augmentation.

AI Sandboxes and Valuations [22:30]

• LangChain has introduced sandboxes for deep agents, partnering with cloud providers to offer virtual environments for code execution, simplifying the process for developers.
• Corser's AI editor has raised $2.3 billion at a $29.3 billion valuation, demonstrating rapid growth since its 2023 founding.
• Anthropic plans to invest $50 billion in building US AI data centers in Texas and New York.
• Emad Mostaque predicts that within three years, cognitive labor will become economically worthless, with humans becoming the "dumbest people in the team."

Cognitive labor becomes economically worthless.

AI for Niche Markets and Economic Models [25:32]

• For mathematical calculations, libraries like SymPy offer a free and effective alternative to LLMs, though combining AI and SymPy can yield the best results.
• Availability of LLMs in Moscow includes Perplexity, Chinese models, Hugging Face, LM Studio, and OpenRouter, with crypto payment options due to restrictions on regular credit cards.
• Russian banks like Zerbank offer their own AI models (GigaChat), and Yandex has a new model with a 32,000 token context window.
• AI is making money in small niches by making Software as a Service (SaaS) 10 times more affordable, creating business opportunities for small and medium-sized businesses.
• The recommended approach is to start an AI agency, using no-code platforms to build custom solutions, as 95% of businesses have yet to implement AI.

So it really becomes a business opportunity.

AI Orchestrators and the Gold Rush [31:48]

• The concept of "AI orchestrator" is emerging, referring to individuals who can direct AI systems, move up the value chain to strategic thinking, and build AI-native skills.
• The current AI landscape is compared to a gold rush, with opportunities for AI agent builders, content agencies, and generative SEO.
• Companies like Verizon laid off 15,000 people in November, highlighting the ongoing job market shifts driven by AI.

One person with AI agents can now do work of 10 people.