AI-generated Users
Highlights
Insights
A recent McKinsey survey suggests that
a. Executives are underestimating employee AI usage—by a factor of three.
b. Leaders must think beyond automation and pursue AI as a business model innovation tool, not just a cost-cutter.
c. Sales, marketing, software engineering, and R&D account for 75% of AI’s economic potential—but employees in these roles have mixed enthusiasm. Leaders should prioritize AI use cases that make employees’ lives easier, not just drive efficiency.
d. Half of employees want formal AI training, but a fifth report receiving little to no support. Companies that invest in internal AI education programs will see faster AI adoption and ROI.
News
So many new AI models have been launched recently that it feels overwhelming. However, the most important news might be something seemingly trivial: Microsoft has made Think Deeper, a feature based on OpenAI o1, available to all Copilot users. This means your company likely has access to advanced step-by-step AI reasoning capabilities for tackling complex issues.
Implications: You can now explore ways to integrate this sophisticated problem-solving ability into your workflow or operations. The potential is immense, but extensive experimentation and risk management will be necessary.
Innovation Insights
1. Superagency in the workplace: Empowering people to unlock AI’s full potential (McKinsey)
While leaders assume AI adoption is slow among employees, the research shows they are three times more likely than leaders realize to be using AI for significant portions of their work. Despite the widespread investment in AI (92% of companies plan to increase AI spending), almost no companies believe they have fully integrated AI into their workflows. Employees aged 35-44 are the most enthusiastic about AI and are already acting as AI adoption champions, fielding questions from their teams and encouraging experimentation. Industries with the highest potential for AI-driven economic impact, such as consumer goods and retail, are investing less in AI compared to sectors like healthcare and telecom. While nearly all companies are experimenting with AI, only 19% report a revenue increase of more than 5%, and only 1% have achieved full AI maturity, signaling a major gap between investment and realized value.
2. The Case Against AI-Generated Users (IDEO)
Design research today is caught between pressures on time and budget and the allure of AI-generated, synthetic participants. The authors caution that while generative AI might seem like a shortcut, it can’t replicate the depth, nuance, and unexpected insights that real human users provide in open-ended interviews and observations. They point out that AI-driven “users” lack emotional cues, authentic context, and the organic tangents that often lead to breakthrough ideas. Instead, the authors advocate a range of resourceful methods—from family-and-friend feedback to intercepts and expert interviews—to gather genuinely meaningful data. Although AI can help shape research plans or spark fresh angles, real people remain essential for uncovering the complexity and richness of human experiences.
3. Problem-Finding AI Agents and Exponential Serendipity (Gianni Giacomelli)
AI problem-finding agents are poised to become a game-changer in 2025, shifting from merely solving problems to identifying them in the first place. These agents will use structured idea analysis, knowledge graphs, and automated knowledge curation to discover the “why” and “what” before determining the “how,” helping organizations surface meaningful and previously unseen challenges. By deliberately colliding ideas and leveraging AI-enabled ideation, machines can find intersections between different knowledge spaces, whether in innovation ecosystems, consumer behavior trends, healthcare gaps, or business bottlenecks. This ability to identify emerging needs and overlooked opportunities can revolutionize industries by allowing companies, policymakers, and researchers to focus on problems that are both desirable and solvable. To successfully implement AI problem-finding agents, organizations should start small with human-centered, scalable proof-of-concept projects, ensuring these tools enhance decision-making and strategic foresight rather than just providing answers.
4. What DeepSeek Means for Open-Source AI (IEEE Spectrum)
DeepSeek has emerged as a major player in open-source AI with its latest models, DeepSeek-V3 and DeepSeek-R1, which rival top-tier models from OpenAI and Anthropic while being significantly cheaper to train. Despite U.S. export restrictions on Nvidia’s high-end chips, DeepSeek managed to develop V3 using the less powerful H800 chip and innovative training techniques like the “DualPipe” algorithm and mixture-of-experts (MoE) architecture, keeping costs under $6 million. Its reasoning model, DeepSeek-R1, adopts a novel training method that combines a small supervised fine-tuning dataset with reinforcement learning, making it both cost-effective and highly capable. The models have gained widespread adoption, with over 700 derivatives on HuggingFace and even running on low-power devices like the Raspberry Pi. While DeepSeek’s models are open for use and modification, the lack of transparency regarding training data has sparked debate in the open-source community, leading to efforts like HuggingFace’s Open-R1 to create a fully open alternative.
5. The GenAI Focus Shifts to Innovation at Colgate-Palmolive (MIT Sloan Management Review)
Colgate-Palmolive has taken generative AI beyond efficiency improvements, using it as a core driver of innovation by synthesizing consumer insights, identifying unmet needs, and generating new product concepts. One of the most surprising insights is their use of digital consumer twins—AI-generated consumer profiles that predict real-market reactions, allowing for rapid testing of multiple product ideas without traditional focus groups. They have also successfully combined AI-augmented idea generation with digital twin validation, finding that AI-supported concepts perform as well as, or better than, those created by humans alone. Unlike many companies that use AI for personal productivity, Colgate-Palmolive has integrated it deeply into its business strategy, even establishing an AI Hub to democratize AI tools for employees while ensuring responsible use. By tracking AI’s measurable impact through KPIs and feedback, they are setting a precedent for how generative AI can be leveraged to enhance creativity, innovation, and decision-making in large-scale consumer product development.
AI Innovations
1. General AI Models and Platforms
OpenAI has introduced ChatGPT Gov, a tailored version of its AI platform designed exclusively for U.S. government agencies, enabling departments to securely utilize AI within a private environment (OpenAI).
Google has quietly launched Gemini 2.0 Pro Experimental, its next-generation AI model designed for improved factual accuracy, coding, and mathematics, available to Gemini Advanced subscribers as an early preview while emphasizing rapid iteration and user feedback (TechCrunch).
Microsoft has made OpenAI’s o1 reasoning model, known as Think Deeper, free for all Copilot users, enabling them to tackle complex queries with step-by-step reasoning without needing a Copilot Pro subscription (The Verge).
Alibaba’s Qwen2.5-1M is an open-source AI model that supports a groundbreaking 1 million-token context length, significantly outperforming its 128K-token predecessors and even competing with GPT-4o-mini, using advanced techniques like Dual Chunk Attention and sparse attention for faster and more efficient processing (GitHub).
Qwen2.5-Max is a large-scale Mixture-of-Experts (MoE) model trained on over 20 trillion tokens, outperforming DeepSeek V3 in key benchmarks and demonstrating competitive results against GPT-4o and Claude-3.5-Sonnet, with its API now available via Alibaba Cloud for advanced AI applications (GitHub).
Mistral Small 3 is a 24B-parameter, latency-optimized open-source AI model that rivals larger models like Llama 3.3 70B and GPT-4o-mini, excelling in fast-response tasks, low-latency function calling, and domain-specific fine-tuning, while being optimized for local deployment and efficient inference (Mistral).
AI2’s Tülu 3 405B, the largest open-weight AI model trained with a fully open post-training recipe, surpasses DeepSeek V3 and GPT-4o in key benchmarks, leveraging Reinforcement Learning with Verifiable Rewards (RLVR) to enhance performance in math, reasoning, and safety, despite significant scaling challenges (Allen AI).
Hugging Face has integrated four serverless inference providers—fal, Replicate, SambaNova, and Together AI—directly into its model pages and SDKs, enabling seamless access to AI inference with customizable provider preferences, flexible billing options, and enhanced ease of deployment for developers (Hugging Face).
2. AI for media
2.1 Image and video
DeepSeek‘s open-source Janus-Pro-7B AI model outperformed OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion in image generation benchmarks, achieving superior visual quality and stability by enhancing training processes, data quality, and model size with 72 million high-quality synthetic images (Reuters).
Alibaba’s Qwen2.5-VL is a powerful vision-language AI model that excels in image recognition, object localization, document parsing, video comprehension, and multimodal reasoning, offering enhanced capabilities in understanding long videos, generating structured outputs, and acting as a visual agent for computer and mobile interactions (GitHub).
Pika 2.1 is a cutting-edge AI-driven video creation tool that enhances animation, motion control, and scene realism with advanced physics simulation, dynamic lighting, seamless style transfer, and extended video length, making it an accessible yet powerful solution for creators, marketers, and animators seeking high-quality, professional-grade content (Pika).
2.2 Music AI
YuE is an open-source AI music generation tool that transforms lyrics into songs, offering a free alternative to Suno and Udio, with customizable styles, multilingual support, and advanced vocal techniques, now available on Hugging Face and GitHub (GitHub).
Riffusion has launched a free AI-powered music platform that generates personalized songs using its Fuzz model, setting itself apart with adaptive learning and accessibility, potentially disrupting the music industry by making AI a collaborative tool for both professionals and casual users (VentureBeat).
3. Other
3.1 AI for science
ESM3 is a multimodal generative AI model trained on evolutionary data that can design functional proteins far from known sequences, with one generated fluorescent protein demonstrating an evolutionary leap equivalent to 500 million years (Science).
3.2 Copyright
The U.S. Copyright Office issued a report concluding that purely AI-generated works are not copyrightable, but works incorporating AI-generated elements can receive protection if they contain sufficient human authorship through prompts, modifications, or expressive input (US Copyright Office).
Other Innovations
1. Mice with two dads
Scientists in China used CRISPR to create mice with DNA from two fathers, producing live pups that survived to adulthood, but the approach remains highly inefficient, raises ethical concerns, and is far from being applicable to humans (MIT Technology Review).
2. Quantum computing
Canadian startup Xanadu has developed a scalable, photon-based quantum computer called Aurora, designed with modular server racks to enable future quantum data centers, though significant challenges in qubit scaling and error correction remain before achieving practical applications (MIT Technology Review).
3. Heart failure
Epicardial engineered heart muscle (EHM) allografts derived from stem cells successfully remuscularized failing hearts in rhesus macaques without adverse effects, demonstrating long-term retention, improved heart function, and vascularization, paving the way for a first-in-human clinical trial in heart failure patients (Nature).
Leave a Reply