The AI landscape never slows down and this week, it’s a mix of talent wars, enterprise intelligence, and next-gen creativity.

👥 Meta’s $1.5B Talent Grab — Thinking Machines co-founder Andrew Tulloch joins Meta, marking a major AI talent win as the company doubles down on research and infrastructure.

🎯 Adobe’s B2B AI Agents — New agents like Audience, Journey, and Data Insights automate enterprise marketing, helping brands identify key buyers and drive faster deal closures.

🎵 Gemini 3 Leak — Google’s next-gen model teases music generation, multimodal reasoning, and built-in agentic features, pushing the frontier of creative AI.

From billion-dollar AI hires to smarter marketing copilots and creative multimodal breakthroughs, this week proves the race for AI dominance is hitting a new high note.

Meta Snaps Up Thinking Machines Co-Founder

Andrew Tulloch, co-founder of Mira Murati's AI startup Thinking Machines Lab, has left the company to join Meta amid fierce industry competition for top AI talent. Tulloch had been a foundational figure at Thinking Machines, which launched earlier in 2025, and was reportedly offered a compensation package by Meta reaching $1.5 billion over six years, reflecting Mark Zuckerberg’s aggressive recruitment of AI researchers after a failed attempt to acquire the entire startup. Meta’s hiring spree is accelerating as it ramps up investments in AI infrastructure and aims to outpace rivals in developing advanced AI solutions, while Thinking Machines continues under Murati’s leadership, focusing on enterprise AI despite Tulloch’s exit.​

Adobe Unveils B2B AI Agents for Marketers

Adobe has launched a new suite of AI-powered agents designed specifically for business-to-business (B2B) marketers, expanding on its initial consumer-focused AI agent offerings within the Adobe Experience Platform. These specialized agents help organizations identify key decision makers within corporate buying committees, generate insights on high-value buying groups, and orchestrate cross-channel marketing campaigns to improve demand generation and accelerate deal closures. The agents—including Audience Agent, Journey Agent, and Data Insights Agent—leverage vast stores of customer data to automate complex B2B marketing and sales processes, working collaboratively with human teams to personalize experiences and boost overall efficiency for enterprise clients like Cisco.​

Gemini 3 Leak Reveals Music & Agent Features

Google's Gemini 3 AI model, expected to launch around October 22, 2025, reportedly includes major upgrades such as multimodal reasoning, faster response times, lower costs, and a novel "original music generation" feature. This music generation capability builds on prior Gemini projects that have collaborated with human composers to co-create complex pieces by interpreting input like musician improvisations and suggesting creative motifs. While some skepticism remains around commercial viability due to licensing challenges, Gemini 3 aims to enhance marketing workflows by integrating text, image, audio, and video understanding with potentially embedded agentic behaviors for tasks like web app interaction and content organization. Early testers suggest Gemini 3 could outperform previous versions and competitors in coding and SVG tasks, although final performance, pricing, and delivery timelines remain unconfirmed.​

Hand Picked Video

In this video, we'll look at how you can run DeepSeek-r1 on your Android Device.

Top AI Products from this week

• Strawberry - A browser that brings intuitive AI automation to your daily workflows. Strawberry helps you research, source data and automate repetitive tasks - from data entry to meeting follow-ups.

• n8n AI Workflow Builder - Generate nodes, logic, and structure from text, then shape and ship your vision faster. Rolling out this week to n8n Cloud (Trial, Starter, Pro). Update to v.1.116.0 to try it.

• Mailmodo AI - Run your email marketing in minutes, not hours. With Mailmodo AI, you simply prompt your goals and watch email designs, automations, and reports build themselves. No complex tools, no manual work. Just smarter email marketing with one prompt.

• AI Agent Platform by CometChat - Get a user-facing AI agent or copilot for your website and apps. Use our Agent Builder or connect your own agent. CometChat takes care of everything- agent, UI, guardrails, notifications, and insights. Scalable, compliant, and enterprise-ready from Day 1.

• nanochat - nanochat is a full-stack implementation of an LLM like ChatGPT in a single, clean, minimal, hackable, dependency-lite codebase. Run tokenization, pretraining, finetuning, evaluation, inference, and web UI on a single 8XH100 node.

• Meilisearch Chat - Meilisearch makes it easy to build fast and custom search experiences. Lightning fast: Response at each keystroke. Plug ‘n play: No config required. Ultra relevant: The most advanced full-text search.

This week in AI

• AI Models Mimic Human Purchase Intent - This paper presents Semantic Similarity Rating (SSR), a method using LLMs to simulate human purchase intent in surveys with 90% accuracy by converting textual responses to Likert-scale scores, outperforming direct numeric ratings and improving consumer research scalability and insights.

• Code2Prompt Codebase to AI Prompt - Code2Prompt is an open-source CLI tool and SDK that converts entire codebases into structured prompts for Large Language Models, automating context prep for code analysis, generation, and review.

• Ultravox Real-Time Multimodal Voice AI - Ultravox is an open-weight multimodal LLM that directly processes speech to text without separate ASR, enabling faster, natural voice AI. Built on Llama 3 and Whisper, it supports real-time voice agents and speech translation.

• Sana Fast High-Res Image Generation - Sana is an efficient text-to-image AI by NVIDIA, generating high-quality 2K to 4K images rapidly using linear diffusion transformer, deployable even on laptop GPUs.

• Agent Learning via Early Experience - This paper introduces "early experience" training for language agents using their own interaction data without rewards, improving generalization and bridging imitation and reinforcement learning​

Paper of The Day

SR-Scientist: Scientific Equation Discovery With Agentic AI" presents a framework that evolves LLMs from passive equation proposers to autonomous AI scientists. SR-Scientist writes code for data analysis, implements, evaluates, and optimizes scientific equations through long-horizon interaction with experimental feedback. It uses a code interpreter as a tool and an experience buffer to overcome context length limits. The framework outperforms baselines by 6%-35% across four scientific disciplines, showing robustness to noise, better generalization, and symbolic accuracy. An end-to-end reinforcement learning pipeline further enhances its capabilities.

To read the whole paper 👉️ here.