The market is full of hype around “agent” breakthroughs — from OpenAI’s browser-capable models to automation stacks like N8N. But most of these products still sit closer to the consumer and SMB side of the spectrum.

Over the past few months, I’ve been chasing the most “AI-friendly” zones — places where AI feels less like a demo and more like real leverage. Manufacturing, e-commerce, law, consulting — I’ve tried them all. It’s been a quick A/B test across industries, ICPs, product angles, and sales motions.

The biggest friction in enterprise AI adoption tends to show up at three critical stages:

1. The Business Owner Level

They normally have some rough idea about what AI can do, but not sure the exact “can-do” of AI/ agent system, and how it can integrate into their business/ existed systems.

Every industry have a general different level of understanding, I’ve found e-commerce is more open, and manufacturing & transporting are the most conservative.

For AI solution providers, the ideal way to engage them is by showing ready-to-use examples — concrete use cases that help customers imagine how AI could fit their world. It’s a classic B2B challenge: as Steve Jobs said, “people don’t know what they want until you show it to them.”

But here’s the paradox for startups — we rarely have dozens of polished use cases. Most of the time, we enter the conversation with one or two demos and use them as a thinking tool to help customers uncover their own needs. That’s how discovery actually happens.

Palantir used this model early on. They had one powerful demo, and used it to map customer pain points across entirely different industries — guiding clients to see how the same technology could solve their specific problems.

The process works best when the decision-maker already has a conceptual understanding of AI. For example, one of my clients — a 20-year-old company — had a CEO who wanted to explore how AI could drive efficiency. He already had a clear mental model of what AI can and can’t do, which made the initial conversation productive.

Still, even with an aligned CEO, the next hurdles usually appear with the department heads or IT teams — the ones who must turn vision into action.

2. Middle-level Managers Level

Department heads fear disrupting existing corporate systems. They often ask the same question: “Can this integrate with our existing SaaS tools or workflows?”

I see this concern on two levels.

First, they don’t want to abandon what they’ve already built. They want improvement, not replacement. If our AI agents completely overtake their existing systems, it could change how the team operates — even how the department is managed. Suddenly, the CEO might question the value of what the manager and their team spent years building. That creates insecurity.

Second, we’re not just competing on product features — we’re entering a political game against existing SaaS vendors, internal budgets, and the finance team’s rules. It’s a system of power — not just tools.

3. The Front Line Level

The final hurdle is the front-line employee. Employees want job security, so they resist change.

I won’t hesitate to say it: AI will change how people work. When you try to integrate AI agents into a company’s daily workflow, you face a wall of friction from the front-line workers.

For example, in one of my cases, the CEO was eager to adopt AI. He wanted it to analyze data and help him make better decisions. But when we tried to work with his IT team to integrate data from their existing systems, the project stalled. The IT team found an endless list of “challenges” and excuses.

The project has been stuck for weeks, and the reason is obvious: the employees know this agent might replace their jobs, or at least threaten the responsibilities they cover.

These are the three main walls I’ve hit selling B2B AI products in 2025. As a startup, we can’t afford to run through walls. We have to find the doors.

Our path forward has to be based on these facts:

• We don’t have the resources to educate the market. We must find customers who already understand AI’s potential.
• We must find the “spot” existing SaaS can’t cover. Or, we must provide a 10x better solution, (which, let’s be honest, is rare and hard to prove).
• Don’t fight over existing value—create new value. Make people feel stronger because of AI, not threatened by it.

So, the smarter path for an AI provider isn’t to threaten the current system, but to expand it. Find the “no man’s land”—the untapped areas where AI can enhance existing workflows. Reframe our value around that.

Managers are far more willing to adopt solutions that empower their teams without blowing up their entire structure.

The post Selling AI in 2025: What I’ve Gotten Wrong appeared first on Felix Kang.

• Product is the problem you solve.
• Team is the talent you assemble to solve it.
• Resources are the fuel—capital, brand, and distribution.
• Culture is the operating system that governs how everything runs.

Recently, I’ve gained a deeper appreciation for the role of management systems and culture.

I’ve come to see its evolution in three distinct stages, best measured by how the team experiences its leader.

Stage 1: The System — The CEO is Feared (or at least, closely watched)

In the beginning, the founder’s will is everything. You have a unique insight, and you drive progress through sheer force and tight control. The product is the only thing that matters.

But once you find product-market fit, the company has to grow beyond you. This is the awkward stage of scaling. You introduce formal management: departments, levels, OKRs, KPIs. It’s a heavy, clunky scaffolding, but it’s necessary to stop things from breaking.

In this phase, the CEO’s job is to be the system’s enforcer. You’re in the trenches, driving execution, and making unpopular decisions to instill discipline. The team might not love you for it, but they follow the rules because you’re watching. Leadership is direct, visible, and top-down. It works, but it doesn’t scale.

Stage 2: The Habits — The CEO is Respected

The management system doesn’t change much between Stage 1 and 2. What changes is the team’s behavior. The rules on paper start to become rituals in practice.

Culture emerges in the small, consistent actions that need no top-down command:

• Every meeting ends with clear action items and owners.
• Product mocks are debated based on customer delight, not just short-term metrics.

But here’s the trap: many assume culture comes simply from hiring for perfect alignment. In reality, very few candidates are born fully compatible with company values. More often, people bring different strengths and weaknesses:

• Someone executes with precision but lack first-principles thinking.
• Someone thinks deeply but dislike tedious execution.
• Human nature is hard to change. No amount of lectures, rules, or KPIs can rewrite someone’s core traits.

Your job isn’t to change their nature. It’s to shape the environment. You create a feedback loop—rewarding aligned behaviors, course-correcting misaligned ones—until the desired habits become muscle memory. The team follows the norms not because the CEO is watching, but because the team is watching.

Stage 3: The Values — The CEO is Invisible

The final stage is the hardest to reach. Here, the CEO’s presence is barely felt in day-to-day operations. The company runs on a deeply embedded set of values that function like an immune system, automatically rejecting behaviors that don’t fit and nurturing those that do.

It takes years to reach this stage.

A vivid example comes from Anker, the global leader in charging accessories (even Trump has used their power banks). At one point, Anker had over 80 versions of power banks list on their store. Quantity diluted efficiency, and despite repeated warnings, the pattern persisted.

The CEO,Steven Yang wanted an Apple-like focus: fewer, better products. But commands, OKRs, and KPIs all failed. For an individual product manager, shipping more SKUs was a rational way to de-risk their career and look busy.

To flip the logic, the CEO introduced a penalty mechanism:

• Each category had a products cap. If a team exceeded it, their bonuses were cut. Too many SKUs, and the bonus dropped to zero.

The initial friction was immense. Many PMs quit.

But the ones who stayed believed in the philosophy of “less is more.” The culture shifted almost overnight. The wasteful 80-product problem vanished because the system was now wired to reward focus. The CEO didn’t need to approve every product decision anymore; the culture did it for him.

It is a process from stage 1 to stage 2, and he claimed that Anker yet enter stage 3.

When we look across these three stages, a pattern emerges:

• Stage 1 depends on systems and management.
• Stage 2 depends on habits and shared practices.
• Stage 3 depends on values embedded so deeply that they function as an invisible operating system.

This is the paradox of building: the more invisible the CEO becomes, the more powerful the company becomes. The less you rely on top-down control, the more the organization becomes self-sustaining.

Culture isn’t the soft stuff you deal with after you’ve figured out product and growth. It’s the hardest system you will ever build—and the only one that truly scales. In an era where AI can automate workflows, a dense culture is what automates decisions. One offers efficiency; the other builds an enduring company.

The post How to Manage a Company When the CEO Disappears: Culture Does appeared first on Felix Kang.

Think of it as an efficiency multiplier, not a ground-up revolution.

Our feeds on X (Twitter) have been saturated with news of AI funding, flashy new features, and “auto-agents” promising to revolutionize everything. But let’s be honest, most of these dazzling demos or consumer-focused tools don’t translate to real-world enterprise application. They might solve small, specific problems for end-users, but they often fall flat when confronted with the complex, multifaceted needs of large organizations.

The Real AI Frontier Isn’t in Tech

The unmet needs within enterprises are staggering, and they’re not confined to the tech sector. The intelligent upgrade of operations is a must-do for every businesses in next 10 to 15 years. Yet, the majority of AI developers remains focused on the tech fields. The truly massive opportunities, however, remain largely untapped in sectors like manufacturing, healthcare, retail, and agriculture.

Take Tyson Foods, for example. The largest chicken company in the U.S. is investing a colossal $1.3 billion by 2025 in automation, with a strong emphasis on AI analytics, robotics, and intelligent logistics. What’s more, they’re committing $50 million in 2025 alone to upskilling and educating their workforce.

This isn’t an isolated incident. Many traditional industries are deep in the trenches of significant digital investments, and they possess a powerful, growing appetite for intelligent solutions. As these non-tech sectors truly embrace AI, an enormous, dormant market will be unleashed, driving both a boom in practical AI product development and a dramatic surge in AI-related employment.

Where Enterprises Really Are with AI

Here’s the often-overlooked reality of current enterprise digitalization: these companies aren’t shopping for another “cool” feature.

The far more common scenario is they’ve already sunk millions into CRM or ERP systems, yet only 20% of the functionality ever gets used. The other 80%? Often too complex or simply unknown to employees. Here’s where AI truly shines: it’s the solution to unlock that massive, unused potential.

When I talk to decision-makers, it’s clear their understanding of AI is still largely conceptual. They think chatbots, basic Q&A, and rudimentary knowledge bases. Even with exposure to advanced use cases, when faced with their own operational bottlenecks, they’re often paralyzed. There’s a glaring lack of in-house AI expertise to guide them, and the perceived integration risk for agents is sky-high. So, they default to throwing more human capital at the problem.

Our Playbook for Enterprise AI Adoption

How do we cut through the noise and get enterprises to truly understand and embrace AI services?

We’ve heavily invested in user cognitive alignment. Before they even get hands-on with our product, we ensure they have a crystal-clear understanding of how our AI ‘workers’ function, why integration is frictionless, and why our solutions are inherently robust and reliable.

This strategic focus is the core of our sales methodology: “Cognition Build – Pain Point Identification – Solution Mapping.” This guided approach significantly accelerates sales cycles and empowers confident decision-making:

• Direct Visual Comparisons: We create side-by-side videos illustrating a specific workflow: the human-driven steps versus the AI-powered steps. It’s incredibly impactful and easy to digest.

• Collaborative Pain Point Mapping: We sit down with clients, helping them articulate their standard operating procedures (SOPs) and identify critical inefficiencies. This empowers them to discover their own bottlenecks.

• Targeted AI Interventions: Once they’ve recognized their inefficiencies, we present tailored AI services designed to optimize those specific problem areas. It’s about solving the problems they know they have.

A Prime Example: AI Enhancing Sales Workflows

Let’s look at a concrete example of AI seamlessly integrating with existing workflows: AI sales services, like what Clay is doing. They’re not trying to replace sales reps. Instead, they’ve streamlined lead qualification, leaving the final outreach and negotiation to the human touch.

Clay developed an AI research agent called “Claygent,” essentially an AI + SDR Agent. This tool lets users build customized data sources and rich workflows tailored to their needs, helping businesses scour the web for prospective client information.

The workflow with Clay is incredibly simple, just three steps:

1. Retrieve and acquire data.
2. Verify and provide sources.
3. Output the retrieved results in a specified format.

This doesn’t disrupt how sales teams operate; it follows their existing workflow. Sales personnel don’t need to relearn anything when using it, making it easier to measure value through quantifiable service effects. This makes it a no-brainer for enterprises to adopt.

Consider the cost: a single sales lead, if generated by a human sales rep (or customer service), might cost around $37.50. With an AI sales agent, that cost drops to a mere $0.69. No workflow changes, no team adjustments, and instant, tangible results. This is the sweet spot that enterprises are most keen to invest in right now.

The post The Enterprise AI Sweet Spot: Unlocking 80% of Untapped Efficiency appeared first on Felix Kang.

Russell argues that the glorification of hard work is a “slave morality” and that modern society no longer needs slaves. He divides work into two types:

1. Physical labor—arduous and poorly paid.
2. Managing or persuading others—comfortable and well-paid.

The latter category, which includes politics and executive roles, thrives on persuasion rather than deep expertise. Success here depends on marketing oneself, not mastering a domain.

Russell claims there’s no justification for denying most people leisure. The idea that humans wouldn’t know how to fill their time if they worked only four hours a day reflects poorly on modern civilization. Historically, people knew how to enjoy leisure—until efficiency became a cult.

Russell makes an interesting claim: if we’d kept the efficiency methods from World War II, we could all be working four-hour days by now. Instead, we went back to the old system where some people work too much and others can’t find work at all.

But there’s a flaw in this argument. Those efficiency methods didn’t appear out of nowhere. They were created by people working hard under pressure. This is how innovation always happens. Look at tech today—companies like OpenAI, Anthropic, Google, DeepSeek release major breakthroughs every few months precisely because people are pushing hard, not because they’re working less.

The real question isn’t “how little can we work?” but “how can we make work meaningful?”

I definitely agree that modern society should provide more leisure. And modern urban people have forgotten how to use leisure well. Today’s urban leisure has become passive – dining out, watching movies, attending sports games. You rarely see spontaneous folk dances anymore except in remote villages. Yet that same creative impulse still lives in us.

With true leisure time (not just time to recover from exhaustion), we might create new art forms – maybe not traditional dances, but experimental games, boundary-pushing music, or something crazy (in positive way)

For the past two years, I’ve worked 10-12 hour/day. I lose a lot leisure time. There is no doubt that my life is not balanced, but the process has given me deep industry knowledge and unexpected innovations.

What if I have 8 hours leisure time every day? I might find my true interests, spend my effort and energy into creative pursuits. In my case, it can be literature or Guitar (I haven’t played for almost a year, it is right next to me bed)

Just like how Paul graham discuss his attitude about the work:

Don’t let “work” mean something other people tell you to do. If you do manage to do great work one day, it will probably be on a project of your own. It may be within some bigger project, but you’ll be driving your part of it. choose a field, learn enough to get to the frontier, notice gaps, explore promising ones. This is how practically everyone who’s done great work has done it, from painters to physicists.

The post Work, Leisure, and the Paradox of Innovation appeared first on Felix Kang.

In 2024, we’re witnessing unprecedented changes in AI that are reshaping how we work and create.

1. Knowledge Compounds Faster Than Ever: We’re in a “compressed century” where the next decade will pack in the technological progress that would have normally taken 100 years. From AI breakthroughs to medical innovation, software & hardware application evolvement… Keeping up with new technologies isn’t just beneficial – it’s essential for thriving in this rapidly evolving landscape.
2. Focus on Industry-Specific AI Solutions: Instead of chasing the next “0 to 1” breakthrough, focus on vertical AI opportunities. A new blue ocean is emerging as AI becomes essential at both business and personal levels – soon individuals will need specialized AI tools just like businesses need SaaS today. This dual market of business and personal AI solutions could be 10x larger than traditional SaaS.
3. Don’t Settle for Niche Success: While focusing on vertical AI opportunities is essential, don’t let success in a specific niche lull you into complacency. Instead, focus on building “technical compound interest”—a foundation of skills, knowledge, and iterative progress that multiplies over time. AI has significantly lowered the cost and time required for experimentation, enabling you to explore new possibilities and future-proof your efforts.
4. Redefining Roles in the AI Era: Tech giants are reshaping their teams, hiring fewer traditional engineers as routine programming tasks are increasingly handled by AI software engineers like Devin . AI is disrupting existing job functions while creating a demand for individuals who can think broadly and connect dots across disciplines. This shift presents a unique opportunity for everyone—engineers, product managers, marketers, and sales professionals alike—to reinvent themselves and explore new breakthroughs across all roles.
5. Early Movers Win: We’re at a technological turning point. Those who adapt and move early will reap the rewards.

Building and Growing with Clarity

Building something meaningful takes both strategy and patience.

1. Honor the Role of Time: Foundational elements like building a cohesive team and cultivating deep technical expertise cannot be rushed. Respect the time it takes to develop these pillars of success.
2. Start Small, Then Scale: Begin with small, innovative steps to capture the market opportunities. Once you achieve PMF, scale up decisively.
3. Clarity Drives Motivation: A clear and tangible goal inspires your team far more effectively than complex incentive structures. Use product metrics and authentic user feedback to guide your efforts.
4. Rethink Brand Building: Avoid draining resources on endless events and generic content. Instead, leverage trending topics, social momentum, and timely opportunities to amplify your brand’s reach.

Personal Growth in a Time of Transformation

The Essence of Growth Remains Unchanged. Even as AI and technology advance, personal growth continues to stem from exploring possibilities, nurturing balance, and breaking through self-imposed barriers.

1. Break Self-Imposed Limits: Don’t let your job title or background restrict your exploration. Whether exploring AI, science, or crypto, move beyond surface-level buzzwords and dive into the fundamental principles that drive these fields. True growth comes from understanding the essence of new domains.
2. Find Internal Balance: Happiness and fulfillment shouldn’t hinge on external validation. Practices like meditation have helped me better understand myself and my place in this changing world.
3. Foster Cross-Disciplinary Thinking: Some of the most groundbreaking ideas arise at the intersection of disciplines. Stay curious, embrace diverse perspectives, and connect seemingly unrelated dots to spark innovation.

2025 will be a year to take those insights further, to refine what’s been started, and to keep moving forward with clarity and purpose. There’s a lot to look forward to.

The post Navigating The Inflection Point appeared first on Felix Kang.

There are not much new design or game mechanism behind them, what developers do is just remixes of existing genres, like FPS + Infinite runner; RPG + Card; SLG + RPG – and differentiate through themes, art styles, and character designs…

This trend signals a fundamental shift. The gaming industry has evolved into a standardized manufacturing sector, splitting into a stark 1/99 divide: 1% hand-crafted innovative titles, and 99% “fast-moving consumer goods”.

This mirrors what happened in the video content market. Platforms like TikTok and YouTube now dominate users’ attention with vast quantities of template-based, standardized content. This standardization and demand for rapid production creates the perfect opportunity for AI to revolutionize game development.

The standardized and fast production demand, giving AI the perfect application to solve it. Every types of games can be disassembled into components, and reconstructed. Majority of the game dev process can be automated by AI systems, e.g. numerical system, level design, art design, 3D modelling, animations, or story writing, which can give the game infinite story development.

However, this AI gaming revolution likely won’t begin on traditional platforms like Steam or PlayStation. Instead, expect it to emerge from hyper-casual mobile games, social media mini-games, and Web3 gaming platforms. Here’s why:

1. Current AI Capabilities Match Casual Gaming Needs
   
   While AI-generated content hasn’t reached AAA game standards, it’s perfectly suited for casual games where component reuse is common and quality expectations are different. AI excels at creating variations on existing patterns—exactly what the casual market demands.
2. AI Excels at Iteration, Not Innovation
   
   Similar to how LLMs train on existing human knowledge, AI game development tools will excel at remixing and recombining existing game design elements rather than creating entirely new paradigms. This aligns perfectly with the casual gaming market’s need for familiar mechanics with fresh twists.
3. Speed and Standardization Drive Casual Gaming
   
   Casual games have short lifecycles, requiring constant updates and fresh content to maintain user engagement. AI-powered development pipelines can deliver rapid iterations and new features to drive user retention and monetization.
4. Gaming as a Feature
   
   Consider trending crypto games like Catizen and Hamster on Telegram. These products prioritize tokenomics and social features over complex gameplay. AI-generated game elements can provide adequate entertainment while developers focus on core economic and community features.

While some might worry this trend will flood the market with low-quality games, I see it differently. AI is democratizing game creation, similar to how smartphones and editing tools democratized video content. This democratization led to platforms like YouTube and TikTok, which spawned billions of creators and more diverse, high-quality content than the traditional TV/film industry could produce alone.

The future of gaming will likely mirror this pattern, splitting into two distinct categories:

• The top 1%: Elite teams of visionaries creating unprecedented, groundbreaking game experiences

• The other 99%: Enthusiasts and general creators using AI tools to bring their unique ideas to gamified content

This democratization won’t diminish gaming—it will expand it. Just as social media birthed new forms of entertainment beyond traditional media’s imagination, AI-powered game creation tools will enable new kinds of interactive experiences we can’t yet envision.

The real revolution isn’t in AI making games—it’s in AI enabling everyone to become a game creator.

The post Mini Games: AI’s Testing Ground for Mass Game Production appeared first on Felix Kang.

As we enter 2024, PC gaming remains strong. The Steam platform reached an all-time high of 33.7 million concurrent players on Sunday, and a record-breaking 14,531 games were launched on Steam in 2023 – roughly 40 new games per day! Meanwhile, Steam has announced support for AI-generated content in games, stating “We will release the vast majority of AI games.”

Will we see “Free Guy” like characters in the 2024 game market?

Last May at COMPUTEX 2023, Nvidia unveiled its new AI technology Omniverse Avatar Cloud Engine (ACE). This is a real-time AI solution that provides AI models for speech, dialog, and character animation in games – adding real-time interactive capabilities to NPCs. At this year’s CES, they updated the system again. they demonstrated how ACE technology combines speech-to-text recognition and text-to-speech responses with generative AI facial animation and automated character personas to generate computer-created character interactions. BTW, the demo is developed by ConvAI and uses Unreal engine.

NPC Creation in Modern Gaming

In 2024’s upcoming AAA games and current game DLCs, NPCs are still relying heavily on massive human and resource investment for creation.

Games such as The Witcher 3, Red Dead Redemption 2, and Cyberpunk 2077 showcase NPCs that possess remarkable freedom and autonomy, contributing substantially to the narrative’s depth and realism. However, the methodology employed in achieving this still adheres to traditional practices—relying on expansive teams and extensive scripting.

Red Dead Redemption 2 has over 1000 NPC characters across 100+ missions in 6 chapters. Each NPC has their own singer, artist, and voice actor, with nearly 8 years of development and $500 million in costs for the NPCs.

Even the highly anticipated GTA 6, with its colossal $2 billion budget, follows suit. Despite its more expansive open world, the creation of NPCs still necessitates manual intervention and intricate decision tree programming.

Despite this, studies show 52% of players complain current NPCs “just repeat dialogues”, 99% want more intelligent NPCs, and 81% would even pay more for them.

Which AI NPC development tool is better?

There are two major players in the AI NPC development tool market: Inworld and ConvAI. Both support Unreal and Unity engines and have similar functionality – creating 3D avatars, character design, natural conversation, emotions, and reactions. I tested both tools on different engines to set up a functional AI NPC in my 3D world within 2-3 hours. The time required depends on the engine and avatar used.

Unreal Engine:

• Inworld
  • Couldn’t get it running on my Unreal project
• ConvAI
  • Worked well overall. However, the AI voice was inconsistent in tone within conversations.
  • Ready Player Me avatars don’t support gestures, following, or battles. Metahuman avatars can do more but require beefy hardware.
  • The GUI is locked to the game view and can’t be hidden.

I also tried Inworld again in Unity:

• Inworld
  • Quick and easy setup compared to ConvAI.
  • Like Ready Player Me avatars in Unreal, no gesture, follow, or battle support.

In summary, the tools are quite similar – they provide comparable functionality and interaction capabilities. They can stand in for quest givers in 3D games but aren’t robust enough for complex tasks.

I won’t go into full details in this post, but let me know if you want more specifics on any part and I can expand on it in a future post. Comments and let me know.

Challenges Faced by AI NPCs

• Lack of Human-Like Characteristics
  
  AI platforms like GPT-4, Claude 2.1, Inworld, and others boast powerful character generation capabilities. However, players easily discern differences in behavior and emotion compared to human interactions. This discrepancy can disrupt the sense of “real” character engagement, diminishing overall immersion.
• Cost Barriers: The expenses associated with high-quality AI tools and platform APIs pose challenges for small and mid-sized developers, hindering widespread adoption. Economic constraints create barriers, impeding the swift integration of advanced AI technology in the gaming industry.
• Fundamental Model Limitations: The fundamental models underlying these AI tools are not specifically trained for game development purposes. This can complicate seamless integration into game projects. The lack of gaming-specific use cases and optimized configurations poses additional challenges for developers trying to leverage these tools.
• API Stability and Token Limits: Issues such as API stability and maximum token limits further hinder AI integration into the gaming landscape. Even advanced tools like GPT-4 encounter constraints, with token limits restricting the depth and length of NPC conversations. This limitation often results in AI forgetting initial conversation content, particularly detrimental in narratives that rely on extensive and intricate dialogues for in-game storytelling and character development.

Key Elements to “Free Guy”

• Intelligence Beyond Appearance
  
  Just like talented human actors can portray roles well, the realism of AI NPCs depends first on having an underlying model that’s sufficiently smart – with strong general intelligence.
  
  This is mostly determined by pretraining, but game NPC models need broad data plus lots of high-quality dialog, storytelling, etc. tailored to “roleplaying” – not just general text. This enables more human-like conversational ability.

• Cost Reduction
  
  Current industry approaches rely on inefficient API calls across multiple teams. As base model prices fall and indexed knowledge bases for game NPCs emerge, costs can decrease dramatically. Otherwise, high API and token usage will raise game prices or require new monetization like in-game purchases for AI dialog.

• Advancements in Movement and Reactions
  
  While current tools like ConvAI and Inworld excel in handling dialog, actions beyond simple facial expressions often require additional manual programming.
  
  Recent research by Daniel Holden presents promising developments in lifelike simulated actions. When integrated with existing AI decision frameworks, this innovation has the potential to significantly elevate the realism of NPC movements, extending beyond gestures to actions like fighting or hiding.

All in all, In films like Westworld and Free Guy, the idea of NPCs with “consciousness” interacting dynamically with the game world is getting closer, thanks to advances in generative AI. However, these capabilities are mainly in top companies due to tech limitations and funding constraints.

The good news is these challenges are being actively worked on, but we likely won’t see major breakthroughs in 2024.

Comments? Any thoughts to share?
If you liked this post, please follow me on Twitter. You can also find more essays here.

The post After CES 2024, Is “Free Guy” finally here? appeared first on Felix Kang.

1. Opt for Work with a Compounding Effect: Instead of merely trading time for wages, focus on tasks that generate a compounding impact over time.
   
2. Take on tasks aligned with your maximum capabilities： Identify and undertake tasks that match your highest capabilities. If what you’re doing falls below your predefined value, don’t hesitate to adjust, outsource, or delegate to AI.
   
3. Practice Patience in Personal Growth: Understand that personal and professional growth often takes more time than we anticipate. Give yourself the freedom of several months, or even years, to make transformative changes— nothing is easy.
   e.g. OpenAI takes 7 years to launch a product that everyone loves.
   
4. If external progress doesn’t meet expectations, focus on personal growth.
   
5. Prioritize Core Metrics: Concentrate 90% of your efforts on core metrics. This focus will naturally resolve other details, preventing internal friction.
   
6. Lower Expectations for Greater Devotion: Lower your expectations. A smaller set of desires means fewer concerns and obsessions, allowing you to fully immerse yourself in your work.
   
7. No one cares about how hard you work, much like no one cares about what time Buffett wakes up. What truly matters is excelling in your business. Learn to leverage, master AI, and establish your own AI co-pilot to amplify your capabilities. It will amplify your marginal capabilities.
   
8. Life is short: Acknowledge the finite nature of life. Each day presents a unique opportunity for accomplishment. Strive for continual growth, relish the experiences, and take a moment to enjoy the metaphorical blue sky.

The post Something I Want To Share appeared first on Felix Kang.

3D era is coming, everybody will be able to create and share games, film, virtual worlds, and simulations. But, it may not work the way you want it.

Hi Readers

Due to my work, I’ve spent thousands of hours on 3D creations in previous years. Although I don’t publish games, I understand how 3D creations work, from story writing to prototyping and production in a 3D engine. As we all know, 3D creation is unlike photo or video creation; it requires a high learning cost and professional knowledge and is usually done by a team.

To make 3D creation accessible to everyone, we have to solve the problem of the entire pipeline: generating a story, turning the story into game chapters, creating game assets, building and coding the game logic, map, avatar, NPCs, abilities, and more.

AI. Yes, AI has been considered the best way to solve this problem. We have seen some interesting products on the market to solve different problems along the way (I will list them later). But in my opinion, they are heading in the wrong direction.

Coincidentally, I gained inspiration from the office collaboration work platform – Lark. My recent work has transitioned to Lark, an all-in-one work platform. From group messaging, email, documents, video meetings, and calendars, to team wiki knowledge bases and AI assistants, everything can now be accomplished within a single platform. This serves as a compelling example of the integration of AI into the workplace.

This transition prompted me to ponder: AI has already transformed our learning, work, and content creation. Why does 3D content creation still remain complex and somewhat outdated? Could the AI + platform model provide insights and inspiration for the 3D content industry?

So before discussing the future of the 3D engine, let’s take a look at how AI empowers the workspace.

3rd Generation AI workspace

The first-generation workspace is Microsoft Office, where every file is independent, and team members find it hard to collaborate with each other.

The second generation evolved to Google Docs, where team members can access the same file and edit together, but there are issues with file management and knowledge management.

The third generation includes Notion, Lark Doc, etc. The files are based on an integrated knowledge base with robust version control and AI integration.

1. Information find people, not people find information
   
   We’ve been using tools Word, Excel, and PowerPoint for decades. Excel is great for finding data and figuring out problems. But can everyone really use all of Excel’s features? (Think about what the World Excel Championship champion two weeks ago) With AI, we can now easily spot the answer behind the data without needing to be Excel experts. AI + data helps us solve problems by getting information directly, without the need for complicated Excel formulas. That is how AI could change the way of getting information.
   
2. Proactive and Real-time Information Retrieval
   
   Every team has its own way of welcoming new members, telling them about the company’s culture, document rules, decision-making steps, and more. But new folks might forget things, and if the team’s documents are messy or outdated, it can be hard to find what they need. With an AI assistant, all the team’s latest documents are in one place. Newcomers can just ask the AI if they have questions, making it easier to get company info and standards. This makes it simpler to find information and lowers the cost of team communication.
   
3. Personalization – Tailored Workspaces for Everyone
   
   In Word and Excel, everyone sees the same layout, even though we use different features. A good workspace should be like a platform that fits each person’s needs. AI can make the next-gen recommendation engine – not for content but for functions. It can understand what you’re working on and suggest the support and features you need, saving you from a bunch of steps. For example, if a team member asks me for a user profile analysis of last year’s active product users, when I open my workspace, the user behavior data table and analysis template are already there for me.

Will 3D Engines Follow Suit?

There is no all-in-one 3D-dev suite yet

Before the 2010s, most studios created their own internal engines for game development. Nowadays, almost all games are crafted using third-party engines, except some major AAA games that still opt for in-house development (e.g., EA’s Frostbite, Infinity Ward Engine, etc.).

Today, we call them 3D Creation Engines because they’re not just for games; they’re for any virtual simulation. However, these engines mainly tackle issues during the “production stage.” Game development teams still need other tools for different steps, like Google Docs, Slack, Trello, Photoshop, Blender, 3Ds Max, and more.

Potential AI 3D Creation Tools

We have seen some AI tools are trying to solve the 3D creation along the pipeline. Here are some I have spotted

[Story Crafting] Companies like Hidden Door and Storycraft are pioneering AI-driven story-driven games, offering players constant streams of unique, personalized experiences. For instance, game objectives might change based on a player’s Bartle type, broadening the potential audience for each game.

[Game Art] In the realm of art, AI is making waves. Midjourney, DALL-E, Stable Diffusion, and Runway ML are leading the way in generating game art. However, achieving optimal results may require some practice with prompts and familiarity with each model’s temperament.

[3D Models] While text-to-3D modeling is still in its early stages, interesting products are emerging. Tools like DreamFusion can generate 3D models from text inputs, expanding on Google’s 2021 unveiling of Dream Fields. Point-E works similarly to OpenAI’s image generation tool DALL-E; you can describe something, and it generates a 3D model. Meshy aids in turning text or images into 3D or lets you upload a 3D model for AI to add textures.

[In-game Social] Enhancing in-game social and NPC interactions is possible with tools like , convai, and Replica Studios recently showcased a demo featuring a modified version of the Matrix Awakens game, allowing users to converse with NPCs using their own microphone.

I can’t cover them all, but it’s evident that 3D-related AI generation tools are still in their early stages, and we need more time to optimize both the models and user experience. The idea of using over 10 AI tools to create a game is hard to imagine.

AI 3D-dev Suite Will Look Like This

1. All-in-One Game Development Information Flow
   
   Similar to AI workspaces, game development info will be more integrated. From brainstorming ideas to the game story, numbers, game design docs, and later, art guidance. No more scattered files on Slack or Google Docs. AI will bring everything together, preventing info from getting lost in big teams or between departments.
2. AI Generates Over 50% of Framework Content
   
   Whether it’s RPGs, action games, simulations, survival games – the fundamental logic underlying each game is the same. The variations mostly come from art, style, game pace, and other elements. AI models trained on games can generate 50-70% of the basic modules based on the creator’s needs. This allows creators to focus more on crafting the core creative elements.
3. Easier Creation, Simplified Steps
   
   When the complex foundation of game development is automated, 3D creators can focus more on creativity than technical skills. It means we won’t need as many artists or animators but will require more creative planners. This change makes the process less reliant on a lot of manual work, making it more about creativity. Anyone can turn their ideas into playable 3D content, just like quickly editing a short video with ‘CapCut.’

However, Unity and Unreal won’t be replaced, just like Photoshop remains the most powerful tool for image creation. Even with AI, they still cater to professional users and may not be suitable for the broader consumer market, as the dynamics in the consumer market operate under entirely different rules.

But they might not be the go-to 3D creation tools. The rise of UGC platforms and their creation tools will make game creation easier for non-developers.

For instance, back in 2000, Maxis released modding toolkits for The Sims even before the game itself. This allowed players to make personalized in-game assets, making the game bigger and more unique. These toolkits and templates made it easier for players to start creating. Similarly, UGC services like Roblox Studio have made game creation more accessible, creating a cycle where more people are encouraged to make games.

Therefore, I believe the next AI 3D creation tools won’t be standalone; they’ll come with a user-generated content (UGC) platform, similar to the relationship between TikTok and CapCut. When the general users develop a demand for 3D content creation, AI 3D-dev suites become meaningful.

Imagine a 3D version of TikTok. With a platform where everyone creates game stories, we’ll need an easy-to-use AI 3D tool that turns ideas into playable 3D games in minutes.

Comments? Any thoughts to share?
If you liked this post, please follow me on Twitter. You can also find more essays here.

The post Will AI Kill Unity and Unreal? Maybe the AI Workspace Holds the Answer. appeared first on Felix Kang.

Here’s a concise overview of my observations:

1. Centralized and Efficient Organizational Structure: Rapid information transmission, efficient decision-making, and guaranteed execution efficiency characterize their approach.
2. “Amoeba” Team Organization: Teams compete internally, with the winner gaining more resources and business space—a unique strategy fostering competitiveness.
3. Embracing Simplicity and Efficiency: The commitment to simple execution not only streamlines processes but also contributes to employee satisfaction.

While building a successful company involves numerous factors, these insights serve as a solid starting point. Let’s delve deeper:

1. Centralized and Efficient Organizational Structure

Before delving into how PDD achieves this, it’s crucial to note that different industries may favor different organizational structures. In the case of e-commerce, a mature and low-innovation-required industry, PDD strategically positions itself as a later player in the game, both in China and the US. This positioning requires them to rapidly learn from others and implement innovations swiftly.

Centralized VS Decentralized

A centralized organization implies that the majority of decisions are made by the C-team or central management. At PDD, the central managing team, including figures like Colkin Huang, Abu, Jiazhen, proposes and decides on the majority of company decisions—typically numbering between 10 to 20 key individuals.

On the other hand, a decentralized organization distributes decision-making authority throughout the entire company. Take Bytedance, the ‘App Factory,’ as an example. It launched apps like TikTok, Douyin, Xigua Video, Huoshan Video, etc., in a decentralized manner. This approach allowed for redundancy in projects and human resources, attracting numerous talented individuals to drive various innovations. The success of Tiktok is a testament to the effectiveness of this decentralized model.

Implementing Centralized Management at PDD

The primary advantage of a centralized decision-making structure lies in minimizing time spent on meetings.

Meetings, while standard, often consume valuable time and are considered a less efficient means of decision-making. Instead of relying on discussions alone, PDD opts for in-depth research and simulation to make informed decisions. This approach mitigates the need for excessive meetings, freeing up time for teams to focus on execution, ultimately yielding a higher ROI

Many teams resort to meetings when they don’t have a clear understanding of issues. PDD, however, discovered that most of the meetings only bring less than 10% incremental information, which hardly resulting in fundamental changes to decisions. Recognizing this, PDD emphasizes rapid work implementation and iterative feedback is more beneficial, the 10% incremental inforamtion could also be made up during thie process.

Before Temu entered the US market, their business lead conducted in-depth research on the operation experience of existing e-commerce platforms, the backgrounds of their senior teams, as well as user habits in various market segments, and consumer category demands. This comprehensive research informed the deployment of Temu’s factory supply chain. While the research and decision-making process was time-consuming, once the optimal solution was identified, PDD swiftly executed it, emphasizing the importance of timely implementation.

2. “Amoeba” team organization

What is Amoeba?

Amoeba Management begins with dividing an organization into small units called “amoebas.” Each amoeba leader is responsible for drafting plans and goals for the unit. Amoebas achieve their goals through collaboration and the hard efforts of all their amoeba members. In this system, every employee plays a major role and voluntarily participates in managing the unit, achieving what is known as “Management by All.”

Amoeba in PDD

TEMU has a flat organizational structure, which facilitates rapid information flow and decision-making to adapt to the fast-changing market environment.

PDD’s middle management practices amoeba management logic – continuous learning and innovation, because they need to compete for resources. Under the amoeba management model, the winning winners have a strong fighting spirit, and a stable core team.

The organizational design ensures that the winning team can get more opportunities and room. PDD retains these people through great incentives, while ensuring they maintain their fighting capacity.

How does PDD divide each Amoeba unit?

As the organization expands, the conventional method of allocating business units based on manager’s ability and experiences may not be a good approach.

So what does PDD do instead?

A notable example is the internal competition for the second tab in PDD’s Chinese app. Rather than assigning units by a top-down approach, there were two potential proposals—an engaging community farming game and a short video feature —competed internally between teams.

Pinduoduo made the two teams compete (PK). The farming game team could request DAU and certain amount of content exposure from the company. PDD gave those resources to the farming game team to test if it could bring higher GMV to the product compared to the short video team with the same resources. The business that performed better would be placed in the second tab.

However, this challenge model does not provide unlimited chances for teams to take other teams’ resources. They have to pay a “challenging fee” to the company (usually $1.5M/day) to earn the right to challenge. Also, challengers who fail multiple times will lose company’s trust, making it hard to undertake important projects in the future… This encapsulates how PDD enforces Amoeba management, ensuring a fair balance in the allocation of business opportunities among internal teams.

3. Embracing Simplicity and Efficiency

Frontline employees at PDD do not have much understanding of the company’s overall strategy and insights, such as data, conversion rates, industry background. Everyone is doing some simple execution work. In conversations with their employees, it becomes apparent that they typically allocate less than 10% of their working time each week to strategic-level thinking.

Within PDD’s business, employees are not burdened with extensive strategic contemplation. The nature of their responsibilities is direct and uncomplicated. Whether specializing in programming or sales, employees are encouraged to hone specific skills. PDD offers industry-leading salaries for these roles, attracting individuals with robust execution capabilities.

In the retail sector, a landscape of standards and clarity prevails. The first and second-level management tiers meticulously analyze business logic, leaving frontline employees with a focus on simple execution. The emphasis on simplicity and efficiency is not just a preference at PDD; it’s a fundamental aspect of their operational philosophy.

Applicability to Your Startup

PDD, as a late entrant to the e-commerce scene, demonstrates the effectiveness of tailoring management systems to a clear target market and user demands. Their commitment to high-efficiency practices allows them to move three times faster, enabling them to catch up with and surpass competitors.

For startups, especially those in industries where success stories are well-documented, adopting a lean and rapid execution approach is crucial. Take, for instance, X AI learning from OpenAI’s experience, demonstrating the importance of hiring strategically and swiftly developing based on a clear direction.

However, it’s essential to note that PDD’s structure might not seamlessly align with startups in AI or highly innovative industries. Decentralized management, as seen in ByteDance, could be an alternative solution for fostering innovation in such contexts. Recognizing the unique demands of your industry and learning from both successful and diverse approaches can guide your startup towards efficient and effective management practices.

Comments? Any lessons to share?
If you liked this post, please follow me on Twitter. You can also find more essays here.

The post How to improve Startup Action Speed: Insights from Temu (PDD) Management Experience appeared first on Felix Kang.