P.INC AI DISPATCH

So far, the development of artificial intelligence has been confined to a narrow demographic of people, made by engineers for engineers, resulting unsurprisingly in products that excel at coding but often fail to resonate with a broader audience. marketing itself has become unrelatable, with divisive statements designed for virality. This has led to a sense of alienation, with many feeling that AI is something being done to them, not for them. A growing movement toward human-centered AI is bringing creatives, designers, and cultural leaders into the fold. Both Starmer and Jensen have recently emphasized that future programming will be more “human-centered”, with vibe coding (using prompts to generate code) becoming the standard. This trend is exemplified in Anthropic’s expansion of its team to incubate “experimental products” in a new lab led by Instagram co-founder Mike Krieger and Charlie Smith, or previously CMO of Loewe, joining tech company Nothing. These are moves that signal a commitment to building AI that is not only powerful technically, but also intuitive, accessible, and culturally relevant. We are living in a time where technology is in many ways positioned upstream of culture in terms of its influence on politics, infrastructure, and funding. It’s perhaps crucial this tendency gets reversed by steering AI from the perspective of creative, artists, and philosophers who think deeply about what a meaningful society might look like.

The dominant tech aesthetic of 2025 was the “AI manifesto” website, a multitude of beige and off-white canvases populated by narrow serif headers and self-consciously serious pronouncements. This trend, adopted even by the White House, is designed to project trustworthiness in a world increasingly wary of AI. Perhaps this is where Pantone got its idea for color of the year. However, this visual homogeneity is beginning to feel less like sincere communication and more like an intentional façade for an industry users still have a hard time trusting. As we enter 2026, there is a palpable desire for a more varied and expressive internet. The consensus is that it’s time to move beyond the minimalist performance of seriousness and embrace the weird, wonderful, and more delightful web.

The long-standing venture capital playbook of scaling software-as-a-service (SaaS) startups on a subscription model is being rewritten. As AI agents become more capable, the value proposition of per-seat pricing is diminishing, with one IDC report predicting that 70% of vendors will abandon the model by 2028. Instead, investors are pivoting toward companies that combine AI with hardware and infrastructure. In fact, a16z just raised over $15B across five funds with explicit allocation to AI infrastructure and bets on “American Dynamism”, which prioritizes companies that operate at the intersection of “atoms and bits” and that build products at an infrastructural scale with a focus on long-term societal impact. This marks a significant shift from software-only solutions to a more holistic view of technological progress.

Last year marked the rise of slop with a flood of low quality, mostly unsolicited AI-generated content that has so thoroughly blurred the boundary between real and artificial that the default posture online has shifted to skepticism of everything we see. Every image is guilty until proven human. Feeds are now saturated with everything from AI-generated music to deepfake celebrity images. Nonetheless, in a decade that some critics believe has a shortage of newness, slop is strangely also one of the few frontiers being pushed, as articulated by Pitchfork’s overview of brainrot music. Many say the big thing this year will be IRL experiences and designing AI to curate more genuine forms of interaction. One recent example is AI dating start-up Known which uses voice AI to help people get on in-person dates. The fact that audio is becoming so central also signals that nobody wants to stare at a screen anymore. The Cut coined friction-maxxing as a resolution, which represents a conscious rejection of digital optimization in favor of more meaningful, real-life engagement.

The traditional model of institutional forecasting and data collection, long dominated by think tanks, polling, and credentialed experts, is facing a crisis of confidence. Platforms like Polymarket and Kalshi, which saw their valuations soar in 2025, are evolving from niche betting forums into sophisticated, real-time information infrastructures. We’re entering an era where markets create outcomes rather than merely reflecting them. This is a shift from authority-based systems to incentive-based systems. We are also exiting the era of consequence-free opinions where commentators could make bold predictions, move on when wrong, and never face corrections. Prediction markets now impose a type of capitalism where incorrect beliefs get priced out. As venture capital firms like a16z champion this structural reframing, the scope of what can be forecasted and traded is expanding rapidly to include everything from macroeconomic data to the outcomes of political votes, fundamentally altering how we produce and consume knowledge about the future.

The race to build ever-larger, do-everything foundation models appears to be giving way to a more focused strategy. While giants like OpenAI have pursued sprawling product roadmaps (with Atlas, Sora, and a GPT-5 that did not live up to the hype), the real momentum in 2026 will probably be with companies that are focused on specialization. Anthropic is a prime example of this: it does not even have an image model, yet its revenue is projected to reach $70B by 2028 by focusing on building an excellent product for coders. Some are also talking about the era of the “deep wrapper” or applications built on top of foundation models that are model-agnostic and hard to displace because they own a specific workflow, dataset, distribution, and governance layer. Harvey is one such “deep wrapper”, an AI workspace for legal professionals used by half of the top 100 US law firms that has achieved an $8 billion valuation by catering to the unique needs of top law firms. Additionally, the remarkable efficiency of Chinese models like DeepSeek V3, which was trained for a fraction of the cost of its competitors, demonstrates that a deep understanding of a specific domain can be a more powerful moat than sheer model size or throwing more compute and more chips at a diluted, general product.

In a landmark decision that reshaped the AI landscape, Apple announced it has chosen Google’s Gemini to power the next generation of its Siri voice assistant and broader Apple Intelligence features. The multi-year partnership represents a major victory for Google, validating its steady, research-driven approach to AI development and giving it a central role in one of the world’s largest consumer ecosystems. For OpenAI, the decision marks a significant setback, highlighting the increasing competition and the challenges of maintaining a lead in a rapidly evolving market. Apple’s choice underscores a prioritization of credibility, speed, and on-device processing, areas where Google has demonstrated a significant edge. Many say Google logged its best year since 2009, surpassing Apple in market cap and releasing Gemini 3 which has one of the fastest growing user bases.

The gold rush to create “AI wrappers” (an app that sits between the user and the AI model) is officially over. With most such companies projected to fail in 2026, the market is signaling a clear preference for more physical products. This has led to a renewed focus on hardware as a critical moat. Pickle’s AI-powered glasses and LEGO’s Smart Bricks are for instance novel form factors for personal computing that have driven much of the CES conversation earlier this month. This trend suggests that the future of AI will be deeply embedded in the physical devices we use every day, creating a more ambient and intuitive user experience of AI.

The speculative frenzy that defined the memecoin boom of recent years has given way to a harsh reality. In 2025, the memecoin market crashed by over 60%, and more than 11.6 million crypto projects failed. Meanwhile, stablecoins have crossed their proof-of-concept phase and have processed an estimated $33 trillion in transactions in 2025 (rivaling major traditional payment networks). As many AI agents begin to transact autonomously, payments are becoming native to the internet itself, which blockchain is actually well designed for–instant, programmable, fully automated, and transparent.

The association of AI with productivity, optimization, and automation is facing a cultural backlash. Users are starting to realize that generative AI tools often automate away the joy of creativity and of imagination, often reducing the process of creating something to a one-shot experience. This has led to a fascinating shift in how AI is being designed and marketed by tech companies. Instead of generic, productivity-focused chatbots, we are seeing the rise of AI personas—models with distinct personalities, memories, and even belief systems. Companies are discovering that users form stronger connections with AI that feels less like a tool and more like a character, a trend exemplified by the popularity of Anthropic’s Claude, whose perceived personality has become a key differentiator in the market (see which Harry Potter house AI models have been sorted into).

The chatbot as the primary AI interface is being increasingly revisited. As products move away from screens, companies are exploring alternative formats that prioritize play, creativity, and exploration over pure utility. Inspired by game design and child development, this emerging paradigm focuses on creating AI companions and tools that foster learning and the discovery of novel forms of interaction. This is happening in parallel to other industries where we are starting to notice the rise of products, usually reserved for adults, now designed for children–from kid perfume to kid makeup. Companies like Tolan’s alien companion and artist Ian Cheng’s systems-based AI friend are starting to become more popular. Some are also claiming that the next big skill in the AI industry will not be prompt engineering, but game design with recent research showing that skills like visual thinking actually improves model reasoning. As the FT recently noted, intelligence is about learning and much of this has to do with exploring new forms of play and new definitions of cognition.

In November 2025, Hugging Face CEO called it: we’re not in an AI bubble, we’re just in an LLM bubble. The prevailing assumption that bigger is better in the world of large language models is being challenged by a new wave of research focused on efficiency and architectural innovation. DeepSeek’s new paper shows for instance how model efficiency depends on good design, not scale or funding. Another Chinese company, Z.ai produced a model (GLM-4.7) that is 7 times cheaper than Claude while beating it on benchmarks. Another start-up called Poetiq beat Google and OpenAI’s top models with a model that cost $30 per task (half of the price of bigger models). We are seeing a shift in interest toward small, more local language models, as well as what leaders in AI are calling world models that are capable of continuous understanding and semantic interpretation.

The grand, abstract pursuit of AGI is being superseded by a more pragmatic focus on developing specialized techniques that deliver useful, tangible results. The future, it seems, is not a single, monolithic AGI but a diverse ecosystem of bespoke, custom-tailored AI solutions. This is evident in the rise of Mixture-of-Experts and the development of tools like Anthropic’s new tool Anthropic’s new tool Bloom, which allows anyone to create benchmarks for evaluating the behavioral characteristics of AI models. The focus is shifting from broad, generalized benchmarks to more nuanced assessments of model behavior. This more targeted approach is enabling the creation of more domain-specific AI, from generative interfaces bespoke to each user to the development of unique “cognitive signatures” that represent how each user thinks and remembers.

While 2025 was hailed as the the year of reasoning in AI, the next frontier in AI cognition lies in the development of recursive models. These models, which are capable of breaking down complex prompts into smaller, manageable parts and then recursively calling on themselves to generate solutions, represent a significant leap forward in AI’s ability to tackle complex, multi-step problems. Early research in this area, including work on training agents through self-play, suggests that recursive models could unlock new levels of creativity, problem-solving, and autonomous learning.

The narrative of an AI race between the US and China is giving way to a more nuanced understanding of the global landscape. The critical question is no longer who will be the first to invent AGI, but who will be most successful in reorganizing their society around AI integration. As Derek Thompson writes, the whole US economy right now is one big bet on artificial intelligence. While Silicon Valley remains focused on invention and frontier research, Beijing is prioritizing deployment and diffusion of AI into every layer of daily life. This has led to a dynamic where the West leads in research, while the East excels in integration. The coming years will not be defined by who builds the smartest model, but by those who can collectively create a deeply integrated AI-native society.