Variant Fund:加密世界如何塑造更有品位的AI模型?

Odaily星球日报Published on 2024-05-09Last updated on 2024-05-09

Abstract

基于品味的模型往往更难优化,因为它们是集体与个体决定的混合体;没有明显的答案或输出。

原文作者:Alana Levin

原文编译:深潮 TechFlow

在过去两年里,新的 AI 模型层出不穷。这些 AI 模型能够完成许多类型的任务——从查找信息和回答问题,到提供客户支持、校对文档、生成内容等。

许多这样的任务是客观的,有明确的优化功能:找到正确的答案,识别最相关的信息,检测任何错误或异常等。

但也有一些模型的输出结果极为主观,比如制作“优秀”的艺术作品或开发“有趣”的视频。我称这些为“具有品味的模型”。基于品味的模型往往更难优化,因为它们是集体与个体决定的混合体;没有明显的答案或输出。因此,频繁的反馈对于帮助模型了解最新的文化偏好尤为宝贵。

如今,大致有两种方式可以培养模型的“品味”:

  • 基于用户生成的内容/数据(如 Twitter 或 Reddit 的提要),这在理论上可以揭示人类关注的最新动向(从而充当品味的代表)。

  • 利用人类“品味制定者”社区来帮助模型围绕他们的偏好进行主动训练。

第一种方法存在许多不理想的情况。数据可能是孤立的(例如 Reddit 关闭其 API)或引入了偏见(例如,只共享部分数据)。模型也可能过度适应特定平台的算法,尤其是如果它的数据来源选择有限。这可能听起来不重要,直到人们开始想象基于 Twitter 热门内容生成的大量新媒体。这并不理想。

后一种方法,一个由人类提供反馈的网络,避免了上述许多风险。可能仍然会有偏差,但仅在于它包括了选择加入帮助训练模型的社区成员的偏好。因此,关键在于确保这些社区成员,即制定“品味”的人,与模型真正培养出好品位息息相关。

加密轨道可以帮助促进这种一致性。在模型中提供所有权/给予参与成员模型输出的经济利益,可以激励他们真正参与。加密货币还使得参与变得更开放和容易:世界上任何地方的任何人都可以贡献,只要他们拥有链上钱包和网络连接。

一个值得注意的例子是Botto项目。Botto 是一个自治艺术家,$BOTTO 代币持有者有能力帮助每周训练模型。训练很简单:参与者对各种图片进行投赞成或反对票,Botto 从成员的偏好中学习。一周结束时,最受欢迎的作品会被拍卖,而那周帮助训练 Botto 的参与者将获得报酬。

艺术只是具有品味模型的一个类别。其他可能包括电影、电视、其他形式的故事讲述(小说、短篇小说)、喜剧和广告/品牌活动。即使是几年前,这些具有品味的模型也是不可能的。这些工具表现力较差,速度较慢,无法可靠地指望模型产生有凝聚力的或(在视频的情况下)逼真的输出。只有今天,这些才变得可能。

重要的是,具有品味的模型拥有大量(且不断增长的)潜在市场。艺术是一个价值数十亿美元的市场。线上消费的内容每年占据了数万亿小时的注意力。如果人们已经打算在这些娱乐形式上花时间和金钱,那么让他们在生产中拥有一定的份额似乎是合理的,这不仅会创造一个更积极的用户基础,还会创造一个满意度更高的用户基础。想象一下,在奥斯卡最佳影片奖中,主要的参与者是帮助训练并开发故事情节的观众,或者为社区创作的电影设立一个全新的奖项,那会非常酷。

我认为这是创造一种新的内容类别,而不是取代现有的创作。这类似于智能手机和 Instagram 使每个人都能成为摄影师,这些新技术的存在并没有消除实际摄影师的工作,实际上,可能反而让更多人欣赏摄影师的工作。具有品味的模型也是一样:它们通过利用新技术,这里指的是用于消费者所有权和经济对接的加密轨道,创造了一种新的参与形式,从而扩展了上述每个类别。

在过去几年中,我们已经看到了数千种新模型的出现。接下来的几年可能还会有数百万(甚至更多)的新模型出现,至少有一些模型应该努力以新的方式吸引利益相关者,从更大的开放性和可及性,到尝试激励措施的新型所有权结构。有品位的模型是特别适合这种创新的一个领域,但它们不可能是最后一个。

原文链接

Trending Cryptos

Related Reads

If It's Not a Clear Yes, It's a No: A Nine-Year Retrospective by a VC Who Survived Four Cycles

**"Invest Only When Certain": A Nine-Year Retrospective from a VC Across Four Cycles** IOSG founder Jocy shares hard-earned lessons from nine years and over a hundred investments in Web3. The core challenge isn't identifying successful founders, but understanding why talented founders with solid ideas still fail. Through building a "failed founder database," IOSG identified six recurring failure patterns. **Founder Trait Red Flags:** 1. **Emotionally Unstable:** Founders who react defensively to criticism or publicly lash out under pressure (e.g., 80% drawdowns) often fail. Resilience is key. 2. **Lacking Hunger / Having a Fallback:** Founders with significant safety nets (family wealth, cushy fallback jobs) may lack the "do-or-die" commitment needed to survive crypto's brutal cycles. 3. **Unchecked Ego:** Includes "polished execution machines" who excel in known frameworks but struggle when paradigms shift, and "professor-types" who are technically brilliant but resistant to commercial feedback or coaching. **Project Structure Red Flags:** 4. **Token-First, Not Product-First:** Treating the token solely as a fundraising tool with no real utility or connection to product value is a major warning sign. The project should have value even if the token goes to zero. 5. **No Day-1 Exit Thesis:** Founders must have a clear, staged capital strategy from the start, understanding what each funding round needs to prove to unlock the next. "Exit before entry" is crucial. 6. **No Full-Cycle Experience:** Founders who haven't lived through a complete crypto bull/bear cycle (e.g., 2018, 2022) often underestimate their vulnerability. IOSG limits initial checks for such teams to $250k, sizing for risk. **The Positive Flipside: Desirable Founder Traits** The ideal candidate exhibits: obsessive problem-depth, being a second-time founder with a non-consensus vision, strong communication skills with *controlled* ego, relentless perseverance, and a global perspective with agency and taste (increasingly vital in the AI era). **Three Survival Tips for Founders:** 1. **Cash Flow Over Narrative:** Real revenue is what sustains projects, not vanity metrics. 2. **Tokens Are a Liability:** Avoid issuing a token unless absolutely necessary. The hidden costs (market making, liquidity, compliance) are immense, often a multi-million-dollar burden. 3. **Respect Liquidity:** Sell during peaks to build treasury, buy back to support the protocol during troughs. Be realistic about valuations and your ability to deliver for the next round. The final principle is simple yet paramount: **"If it's a borderline 'yes' or 'no,' don't invest."** In an industry that reinvents itself every few years, the discipline to consistently say "no" is the ultimate secret to longevity.

Foresight News6m ago

If It's Not a Clear Yes, It's a No: A Nine-Year Retrospective by a VC Who Survived Four Cycles

Foresight News6m ago

SemiAnalysis Deep Dive into CXMT: $50 Billion Revenue, An IPO Amidst a Supercycle

SemiAnalysis' in-depth report on ChangXin Memory Technologies (CXMT) details its rapid rise as China's largest upcoming semiconductor IPO. Founded in 2016 by Zhu Yiming, CXMT built its DRAM foundation on acquired patents and talent from the bankrupt German firm Qimonda. It achieved its first annual profit in 2025 after nearly a decade of significant capital support, primarily from patient Hefei municipal investors who fostered a local supply chain. The company is now capitalizing on a strong DRAM supercycle. Its revenue soared from ~$3.3B in 2024 to ~$8.6B in 2025, with Q1 2026 alone reaching ~$7.3B. SemiAnalysis projects full-year 2026 revenue could exceed $50B, driven by soaring ASPs rather than massive market share gains. While CXMT is closing the capacity gap with Micron, its product mix remains heavily focused on commodity DDR/LPDDR, which currently offers higher margins than its nascent HBM business. CXMT faces significant challenges in HBM, struggling with yield and stability for HBM3 8-Hi stacks while lagging behind the big three (Samsung, SK Hynix, Micron) in advanced nodes. However, strategic national priorities for AI self-sufficiency may push it to accelerate HBM capacity. Its complex IPO structure reveals heavy state-backed ownership and voting control over its fabs, with Alibaba appearing as both a key cloud customer and a minority shareholder. The IPO aims to raise ~$4.1B, primarily to strengthen its core DRAM manufacturing base.

marsbit26m ago

SemiAnalysis Deep Dive into CXMT: $50 Billion Revenue, An IPO Amidst a Supercycle

marsbit26m ago

From Corning to Ciena: The 10x Opportunity in the AI Optical Communication Chain

The transition from copper to optical communication in AI data centers is creating significant investment opportunities beyond just chipmakers. The entire photonics supply chain, from glass and fiber to connectors and test equipment, is critical. Corning, a key fiber supplier, has locked in multi-billion dollar, multi-year contracts with major cloud providers (Meta, Amazon, Google, Microsoft, OpenAI, NVIDIA), demonstrating pricing power and scale. Its profit growth is outpacing revenue growth. In the interconnect layer, Amphenol benefits from high growth in AI data centers, driven by strategic acquisitions and operational efficiency, while Credo Technology acts as a bridge between copper and optical solutions, though with high customer concentration risk. At the systems level, Ciena enables higher data capacity on existing fiber lines, with a strong backlog and cloud customer adoption. Further upstream, AXT is a bottleneck supplier of key indium phosphide wafers for lasers but faces geopolitical supply chain risks. VEO Solutions provides essential testing equipment for the entire photonics industry. A new pure-play photonics ETF (FOTO) offers a consolidated investment approach. The core thesis is that the physical limits of copper are driving an inevitable shift to optical technologies, with wealth flowing to essential, often overlooked, suppliers across the photonics value chain.

marsbit38m ago

From Corning to Ciena: The 10x Opportunity in the AI Optical Communication Chain

marsbit38m ago

Collector Crypt's DAU Is Only 800, Yet It's Already One of Crypto's Most Profitable Projects?

"Collector Crypt: A Highly Profitable Crypto Project with Only 800 Daily Active Users?" Collector Crypt (CARDS) is a crypto project tokenizing physical graded trading cards (primarily Pokémon) on Solana, achieving significant real-world profitability and growth. According to a Maelstrom Fund analysis, it generated approximately $53M in annualized profit in May, with a June run-rate nearing $109M, against a $550M FDV. Its core revenue driver is a digital pack-opening 'Gacha' system. The platform bulk-buys cards at a 5-15% discount. Users can open digital packs and choose to keep cards or sell them back to the platform at a 7-15% discount to market price. Most users sell back common cards, creating an efficient model: users get packs with a ~2% positive expected value, while Collector Crypt captures ~4.5% profit. The project aims to disrupt the inefficient $22.2B GMV (Q1 2026) eBay trading card market, which charges sellers 16-20% in total fees. Collector Crypt offers 2% fees, instant settlement, insured custody, and one-click trading. Beyond Gacha, future revenue streams include secondary market trading fees, infrastructure partnerships, and an eBay "snipe" tool. It holds ~$23M in card inventory and ~$10M in cash, and has already begun token buybacks. With a total supply of 2B tokens, effective circulation post-2027 unlocks is estimated at ~1.3B. Trading primarily on DEXs has so far limited large institutional entry. The project is expanding into sports cards and attracting Web2 users. Maelstrom Fund's price target is $4 by summer's end, positioning Collector Crypt at the forefront of migrating collectibles on-chain.

Foresight News50m ago

Collector Crypt's DAU Is Only 800, Yet It's Already One of Crypto's Most Profitable Projects?

Foresight News50m ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片