Original title: Notes From 8 Years as a Founder in Geneva AI
original by: @lishali88
Photo by Peggy Block Beats

Editor: This is from a generated AI early entrepreneur's eight-year reset. In 2018, he created Rosebud AI, which is a continuous launch of several AI creative tools, including Tokkingheads, around the goal of making creation as simple as playing games. At a stage when modelling capacity is not yet mature, these products have completed early user growth and product validation by designing processes and interactions to magnify the experience of "will be good but good " 。

This experience almost covers the full evolution cycle of generation AI, from "synthetic media" to a universal capability infrastructure: from the experimental exploration of CycleGAN, StyleGAN, to GPT-4, which opens the boundaries between code generation and interactive creation, technological advances constantly rewriting product logic and entrepreneurship. The author's path also reflects a clearer structural change — when the model becomes a variable, the real watershed is not just the technology itself, but how to build, distribute and commercialize the product around it。

After the departure of the CEO and the addition of a16tz, the author will move to the investment frontier model house and related infrastructure. But more important than an individual path, these eight years of experience point to a growing trend: the first phase of the generation AI (evidence of what to do) is coming to an end, followed by longer cycles of competition, which will revolve around how capacity is organized, productized, and eventually entered the real world。

The following is the original text:

I joined a16z as a partner, focusing on infrastructure and AI-directed investments. In the meantime, after 8 years in charge of Rosebud AI, I will also resign from the CEO post。

Here are some of my reflections and reflections on these eight years. I have great respect for those who are still on the line. A model may swallow your original product road map or it may be years in advance. Design, products, engineering - the pattern of these functions has changed compared to three months ago, let alone eight years ago. The pace of technological progress makes the era the most exciting moment for entrepreneurship and the most challenging。

At a16z, I will focus on front-line model stacks (frontier modelback): including the model itself and infrastructure and development tools built around the model. I'm excited about the rapid evolution of modelling -- more and more progress is driven by AI itself. At the same time, I'm looking at the breakthrough that AI has brought in mathematics and science. Besides, I've always had a special interest in this direction, as I've been building an AI creative tool for the past eight years。

Prior to this, I was involved as an angel investor in a number of seed ship investments, including @fal, @periodiclabs, @SakanaAILabs and @ExaAILabs. Next, I look forward to investing all my energy in supporting those who are building this technology。

2018: BEFORE GPT APPEARS, BETS GENERATE AI。

In November 2018, I used CycleGan to process Myst video, trying to turn forests into experimental videos of tropical jungles

8 YEARS IS A RATHER LONG TIME。

I started at the end of 2018, and it was almost a "old age" when the field was called "synthetic media." I'm messing with CycleGan and StyleGAN, which produce strange and fascinating content, and I'm convinced that one day, creation will be as light as building models in a game。

It should be a play in the best state. And the first microlights of the generation of AI made me believe that this "play-like creative experience" can be extended to more forms of creation. I began to wonder how generation AI would recreate video games (e.g. the CycleGAN video that I trained in 2018 on Myst images)。

For eight years now, we can generate videos, games and even music through a hint. The future that was once imagined has finally arrived — and it is only the beginning。

LOOKING BACK, I WAS ABLE TO FORM SUCH A STRONG BELIEF AT THIS EARLY STAGE, PERHAPS BECAUSE MY LIFE HAS BEEN AT THE INTERSECTION OF TECHNOLOGY AND ART: ON THE ONE HAND, THE DOCTORAL BACKGROUND OF MATHEMATICS AND IN-DEPTH STUDIES AND ON THE OTHER HAND, THE ARTISTIC PASSION OF DANCE AND MUSIC. GENERATING AI'S START-UP REQUIRES BOTH: THE TECHNOLOGICAL CONTEXT ALLOWS ME TO SEE WHAT'S COMING, AND THE ARTISTIC ORIENTATION MAKES ME ANXIOUS TO BUILD IT。

The journey to entrepreneurship is always longer and harder than expected. Find something you almost irrationally believe to maximize the probability of staying alive。

In 2017, Cyclegan, we've come a long way

2018-2023: Use "just good" to impress users

Paragraph 3 iOS applies the Tokkingheads interface screenshot. Early generation, the core of AI is to design a simple process and embrace the roughness of the product. Sensor

This way, we have published a large number of products designed to sharpen the intuition of the ability of front-line models and learn to package them into magical experiences that mask early defects. At that stage, I realized that when the model output is far from perfect, you can design a consumption-level experience that allows users to interact quickly and give feedback. Users are picky, but not vulnerable — it's enough to impress them。

By the time we move to the third application, we've accumulated enough awareness to allow Tokkingheads to achieve a viral transmission of natural growth, which has exceeded 2 million users in weeks. The next key lesson follows: as founders, you must know what product patterns allow you to sustain momentum over time. Tokkingheads could've gone down the road of brushing up the money, but I'm not sure if it was the right soil to turn this creation into a more complete product, which I really wanted。

SO WE CONTINUE TO EVOLVE. WE'VE DONE AI TO GENERATE GRAPHICS, NFT TO ACCOMPANY AI ART (YES, I NAIVELY THINK THAT THE QUALITY OF THE WORK IS THE KEY, AND IT TURNS OUT THAT REALLY IMPORTANT SKILLS ARE SPECULATION), AND THE IAI GAME MATERIAL GENERATION TOOL. EACH PRODUCT TAUGHT ME SOMETHING SPECIFIC: WHY USERS WANT TO PAY, AND HOW FAST THE MODEL IS IMPROVING. THERE IS ALSO A GLOBAL PANDEMIC BETWEEN THESE PROJECTS, AND SILICON VALLEY BANK AND THE THIRD TOTAL AND BANK CROWDING — ALL OF WHICH REMIND ME OF GRATITUDE. IT IS IN ITSELF A PRIVILEGE TO CONTINUE BUILDING。

2023: Code generation matures

The code generation is finally good enough to be ready to create games tools for non-technical creators. After GPT-4 was released, the future became tangible. In March 2023, I shared a memo with the team and put together the original version of the Rosebud text-to-play function with the following prototype。

Interception of tweets on March 23, 2023. I use GPT-4 to learn Three.js to create the skybox with Rosebud-generated AI to demonstrate the early prototype of a text calling 3D scene

AT THE BEGINNING OF 2023, THE AUTHOR WROTE AN INTERNAL MEMORANDUM TO THE TEAM DOCUMENTING THE PRODUCT JUDGEMENT AFTER THE CODE GENERATION CAPACITY WAS BREACHED. THE CORE JUDGEMENT OF THIS INTERNAL LETTER IS THAT AI IS IN A CRITICAL WINDOW PERIOD THAT WILL DETERMINE THE COMING DECADES, AND THAT THE NEXT TWO YEARS WILL BE A HIGHLY COMPETITIVE PHASE WITH A FAST PACE, HIGH INTENSITY AND CLEAR PHASE-OUT. THE COMPANY WILL DEVOTE ITSELF TO THIS “SPRINT” ONLY FOR THOSE WHO HAVE A STRONG INTERNAL DRIVE, WHO ARE WILLING TO ENDURE HIGH PRESSURE AND LONG-TERM INVESTMENT — NOT ONLY AS A WORK EXPERIENCE, BUT ALSO AS A HISTORIC OPPORTUNITY TO CHANGE THE CAREER TRAJECTORY OF AN INDIVIDUAL。

2026 and beyond: What laboratories can you build that you don't want to do

FIGURE: PRESENTATION VIDEO - AUTHOR BUILDS A 3D CITY SIMULATION GAME ON BROWSER END WITH HINTS

PLAYS NEED TO BE ACCOMPANIED BY CREATIVE INTUITION AND TECHNICAL SKILLS. GENERATING AI IS THE KEY TO MAKING GAME CREATION ITSELF A GAME -- ANY MODEL PROGRESS IN IMAGES, VIDEOS, WORLD MODELS OR CODES WILL BE ABSORBED IMMEDIATELY. THE BUSINESS MODEL OF THE GAME IS ALSO MOST LIKELY TO REMAIN OUTSIDE THE HORIZON BLIND OF THE FRONT LAB: THE CORE REALIZATION PATH IS STILL PAID FOR BY THE PLAYER, AND ESTABLISHING A DISTRIBUTION SYSTEM ON THE SIDE OF THE PLAYER APPEARS TO BE AN OVERSTRETCHED BACKLINE FOR THE AGI-RUN LABORATORY. FOR THE FOUNDERS, CHOOSING WHAT TO BUILD IS ALWAYS AN ONGOING GAME OF SEARCHING FOR SPACE BEYOND THE CRITICAL PATH OF THE LABORATORY。

Rosebud is gaining momentum. We have organically built up a large and highly active community of creators. I'm going to miss my conversation with the creators on Discord and my day-to-day processing of user-supported mail. The focus of the next phase will be on scaling up the distribution of players, so it is time to hand over the baton to the teammates who have been fighting side by side。

Congratulations to @grazworks for the new CEO of Rosebud! He is rarely equipped with machines to learn the aesthetics of gift and product。

Martin Casado (Martin Casado) and a16z team accompanied Rosebud throughout his growth. I had a critical conversation with Martin about whether JavaScript was the right technology warehouse for the Rosebud game — selecting Unity or Roblox — which might be more popular, but JavaScript's code generation was much faster because training data were more accessible. The team sought the truth and was willing to bet that it would bring more builders. That is the path to the ideal future: we must build, we must innovate。

I look forward to working with you from the other side of the table. Private letters are open at all times。

[ Chuckles ]Original Link]