JimmyWong

📰 作者
📱 Builder
📚 工具产品爱好者 | Keep Building | 努力让更多的人用上 AI

[JimmyWong @thinkingjimmy]

之前一直想写 Stable Diffusion 相关的教程，但觉得 Stable Diffusion WebUI 可拓展性不强，自由度不够高，所以一直没下笔。最近尝试了下 ComfyUI ，才惊叹地发现这才是我想要的。市面上系统性的 ComfyUI 教程不多，所以最近跟朋友一起搞了一个新教程：https://t.co/bxgtK13axM 希望对大家有帮助。

[JimmyWong @thinkingjimmy]

在使用 ComfyUI 的时候，使用的模型基本上都是开源的模型。但效果很不错的闭源模型，却无法在 ComfyUI 里使用。为了解决这个问题，我们开发了一个插件。这样各位就能通过 ComfyUI 将各种闭源模型串联起来使用了。目前支持 Flux、Ideogram，后续会支持 Hailuo 等视频模型。https://t.co/2wyXLEeP2n

[JimmyWong @thinkingjimmy]

Paul Graham 最新的博文核心就一句：「writing is thinking」。这或许就是为何 LLM 看起来会思考的原因吧。人类写出的文章有很多不是描述结果，而是描述思考的过程，所以用这些文章训练出来的模型，才会看起来像是思考一样：https://t.co/zdSEG0fmJC

[JimmyWong @thinkingjimmy]

推荐一份我几乎每期都会看 AI 类 newsletter，如果你没法天天关注 AI 相关的新闻，但又想保持跟进。那这份 newsletter 非常值得你订阅，能节省你很多时间：https://t.co/b4osvcTVXa

[JimmyWong @thinkingjimmy]

纵观历史，AI 对创作的最大改变是人类创作者第一次有了不需要自己参与的创作能力。AI 赋予了创作者一种「持续」「规模化」输出品味、审美，或者 IP 的能力。如果用类比来形容这种能力，我认为比较合适的是：赋予了创作者一种另类的代码能力。这或许就是最颠覆式的差异机会。

[JimmyWong @thinkingjimmy]

纵观现在相关的 LLM 或 SD 产品，我认为这些产品现在跟「电动车」很像：

1）电车因为其使用成本比油便宜，用户因为这一差异化的成本优势而购买。这跟 LLM 或 SD 产品目前能做得事情很像，也是省成本。但也跟电车一样，也有一些缺陷，比如电车因为受限于电池容量，跑长途还是有点麻烦，而现在的 LLM 或 SD 产品也类似，做某些小工作还是 OK 的，但让其做相对长且复杂的工作，还没法完全胜任。比如写一段代码可以，但写一个项目很难。同时，也有越来越多的非 AI 厂商拥抱 AI，就像是油车厂进行的油改电一样。

2）另外，AI 成本优势也跟汽车一样有地域性限制。在一些人力成本比较低的国家或领域，LLM 或 SD 产品相对没那么好做商业化。就像是在中东卖电车一样。

3）电动车要想真的完全打败油车，我认为只有全自动驾驶出现后，才能真正出现。LLM 和 SD 产品也同样如此。

现在大部分的产品，都是基于已有的产品形态，通过加 AI 功能的方式去实现「油改电」。在我眼里，很多产品就像是 Sketch 至于 Photoshop 那样，因为新兴技术兴起，而产生出了新的工具需求。但最终被 Figma 靠协同这一重构协作的这种颠覆性创新所颠覆。

所以 AI 产品要想突围，最重要的是找到类似自动驾驶至于电车那样的颠覆式差异化功能。

[JimmyWong @thinkingjimmy]

纵观现在相关的 LLM 或 SD 产品，我认为这些产品现在跟「电动车」很像：

1）电车因为其使用成本比油便宜，用户因为这一差异化的成本优势而购买。这跟 LLM 或 SD 产品目前能做得事情很像，也是省成本。但也跟电车一样，也有一些缺陷，比如电车因为受限于电池容量，跑长途还是有点麻烦，而现在的 LLM 或 SD 产品也类似，做某些小工作还是 OK 的，但让其做相对长且复杂的工作，还没法完全胜任。比如写一段代码可以，但写一个项目很难。同时，也有越来越多的非 AI 厂商拥抱 AI，就像是油车厂进行的油改电一样。

2）另外，AI 成本优势也跟汽车一样有地域性限制。在一些人力成本比较低的国家或领域，LLM 或 SD 产品相对没那么好做商业化。就像是在中东卖电车一样。

3）电动车要想真的完全打败油车，我认为只有全自动驾驶出现后，才能真正出现。LLM 和 SD 产品也同样如此。

现在大部分的产品，都是基于已有的产品形态，通过加 AI 功能的方式去实现「油改电」。在我眼里，很多产品就像是 Sketch 至于 Photoshop 那样，因为新兴技术兴起，而产生出了新的工具需求。但最终被 Figma 靠协同这一重构协作的这种颠覆性创新所颠覆。

所以 AI 产品要想突围，最重要的是找到类似自动驾驶至于电车那样的颠覆式差异化功能。

[JimmyWong @thinkingjimmy]

最近我每天都在拍照记录食物，拍照的同时会将拳头也拍进去，这样教练才能通过这个相对估算食物的分量。然后我就在想 AI 拍照计算卡路里是不是也应该这样？另外衍生出来另一个问题，图片是否也能有 prompt 技巧？比如就像 ComfyUI 的 Inpainting 那样给图片画 Mask，然后让 4o 改图，效果会不会好一些？

[JimmyWong @thinkingjimmy]

非常值得阅读的经验。

[David Park @Davidjpark96]

How to go from 0 to $5M ARR profitably (step by step)

Here is every growth hack we used for each of our distribution channels:

- Organic Short-form content
- Influencer Marketing
- SEO
- Paid Ads

Organic Short-form content

The most important thing to remember in this new age of social media is that follower counts don’t matter at all

I would unironically sponsor a fresh account made this week vs a YouTuber with a million subs if they had similar views and I don’t even think it’s a controversial opinion

(especially when you consider that the famous YouTuber will charge you potentially 30x more for less results, but I would maybe even choose the new account if they were the same price)

The reason being that new accounts are highly volatile and if you put out a great engaging piece of content then it could easily go viral regardless of how many followers you have

But rather than finding and sponsoring these new accounts in your niche, you can just create these accounts yourself

You just need to follow this basic format:
1. Find a face for your account
2. Craft a viral video 'series'
3. Multiply your accounts

What would happen if BTS made a new TikTok account and started posting?

All videos on the new account would immediately go viral because the algo would push the videos back to their fans

Despite having zero followers, this new TikTok account will ‘behave’ as though it has hundreds of millions of followers

One great hack is to find someone who already has an audience of people that you know will convert, then ask them to create a completely new account to posts videos about your product

This is an amazing hack because creators are usually very open to this idea because it doesn’t dilute their main page, it’ll be way cheaper to pay for 20 videos on a new account every month vs 20 sponsored posts on their main account, and it’ll get way better results

For example we found a creator named Mengmengduck whose entire account at the time was teaching students how to write to employers and we paid him $4000 for 20 videos/mo

In our first month we got 7 million+ impressions, just as expected the videos on the brand new account were being pushed to the followers on his main account

When you get a new creator just follow this basic guideline:
- Repost vids on all platforms (Reels, TikTok, YT Shorts)
- Post 1+ videos per day
- Initially, post vids similar to what the influencer posts on their main account, then start experimenting with videos that feature your product
- We paid $4000/mo for MMD who had 500k followers, but you can get someone cheaper for similar results!

The entire goal when working with your creator is to find a viral “series” because a video that goes viral once will go viral again

You want a video that you can basically tweak slightly and repost multiple times a week that is still fun for the audience

Once you find a viral series, you can start creating new accounts but this time you don’t need to work with big creators, you can just find charismatic UGC creators because you’ve found content that is engaging and poised to go viral

Currently, we have about 5 different Jenni AI accounts all posting similar content

Many of our accounts have tens or sometimes just hundreds of followers and they just repost stuff from our main account

In some cases, the account with 48 followers gets MORE views than our main account with 55k followers...even when posting identical videos

Examples of viral series:
1. A guy tries to see how many sticks of spaghetti it takes to hold his weight. Then he tries to see how many coat hangers. Then he tries to use sheets of paper. Etc etc. All the same video, minimal effort but will go viral every time.

2. A guy drinks a cup of milk for every 1000 followers he has. Everyday it’s the same video where he drinks milk but people continually tune in because they want to see him suffer and drink hundreds of cups of milk. Will go viral every time.

3. For us we had “POV you have an essay due” and it was just a ridiculous plot of someone realising they have an essay due while they are taking a shit or right before they sleep or w/e and it always ended with them using Jenni AI to help them write their essay faster. Same video, slightly different 10 second intro each time.

Thinking of new viral video ideas are so hard, so just try to think of tweaking your few winning videos into “series” and just think about prolonging the lifespan of your content

The order of virality:
1. Have one account on each platform (TikTok, Instagram, YT shorts) and experiment with a ton of hooks and video ideas

2. Eventually you’ll get a video that goes viral after enough experimentation

3. Experiment and somehow turn that video into a “series”

4. Start tweaking and posting this “series” on multiple accounts

5. Translate and create the same video with creators that speak Chinese, Spanish, etc

6. Take the really mega-viral videos and sponsor other pages to repost them

7. Use the mega-viral content for paid ads

8. Eventually enough copycats will copy your video series and it’ll get played out and you restart at step 1

Our video series "POV: You have an essay due" has probably generated 300 million+ views overall

It was essentially the same video over and over, multiple times a week, yet they consistently went viral

That one video series made us over half a million dollars, and it was one of several video series’ that we were able to cook up

Influencer Marketing

Find influencers

- Go to your users' Instagram and see which influencers they are following. This should be easy if you’ve done user interviews correctly, if you have no social media handles of your users you should actually befriend your users and get to know them on a deeper level or you’re not going to make it

- Do this for several users and eventually you will find some influencers that are ‘popular’ among your target user demographic. Find these influencers and follow them all, it’s okay you only need 1-3

- Go to these influencer's profile and click "suggested similar accounts" and you'll get an easy list of hot leads that you can sponsor (you can continue to do this recursively for each new influencer that you find)

- Also, be sure to see what hashtags these influencers are using when they post and then follow those hashtags as well

- Once you have a list of influencers, create a new Instagram/TikTok account and manually follow & watch their videos all the way through and the algorithm will start automatically showing relevant influencers to you that you can then reach out to (this is a great way to lower the risk of getting scammed because the videos that show up organically on your FYP are less likely to be accounts that have bought followers/views)

- On your new account, you want to be on the hunt for smaller influencers, if a new-ish account has multiple videos with 100k+ views, it's absolute GOLD

- Once you have a list of influencers who you are ready to partner up with, you can begin reaching out to them

Reach out to influencers

- DMs > emails = higher response rate (at least for us)

- All messages need to be as detailed and tailored as possible, but most importantly as CONCISE as possible. This balance is hard but you’ll get a better feel for this as you notice what gets ignored and what doesn't

- Demonstrate that you’re genuinely a fan of their content and that you’re excited to partner up. This is such a low bar but few founders actually put in the effort.

- Why should they partner with you? Why will their audience love your product? Don’t talk about dumb shit like how many employees your company has or what round of fundraising you’re at (I suggest not talking about those vanity metrics in any situation, but I digress)

- Make sure to indicate somewhere in the message that this is obviously a PAID promotion, influencers are constantly inundated by people who are begging them for free shoutouts or weird affiliate partnerships

- Expect more than 50% of influencers to not respond, but this % is very volatile depending on what industry you’re in, and how cool/well-known your product is

- The good news is, as more influencers talk about your product it gets easier and easier to convince them to promote your product because they have heard of it before (unknown products could be risky or straight-up scams)

Negotiate

- The highest priority is to align the incentives between you and the influencer. You should both want a banger video that converts

- Never pay upfront for videos (pay half upfront at most). Most influencers are fantastic people but some influencers will just try and drop a half-ass video once they get their check

- Try to split the payment so that some % comes from the number of conversions that they bring (you can track this via coupon code or UTM link). If they don’t want to do that, try to at least have some affiliate bonus

- Ask for their viewer demographics see which countries are viewing their content and compare that with the conversion rates of those countries for your product

- See previous sponsored content that they’ve posted and see how those videos performed vs their usual content

- Negotiate the deal so that either they only receive some/all of the payment if it reaches a certain # of views (or a certain # of coupon code redemptions)

- Don’t listen to these absolute garbage articles that tell you to pay based on the number of subscribers/followers that they have. The sole question you should be asking is, will they convert or will they not (at least in the early stage of your startup).

Post content

- You can’t use a general content strategy for all of your influencer partnerships because each influencer has a certain type of content that their audience likes. Your content can’t deviate too far from it or else it will have terrible watch times and you’re basically paying for a dud

- If you don’t have experience with social media or making engaging content, just let the influencer make the video and then you approve it (after aligning your incentives, as we discussed in the last step)

- One important note is that views are irrelevant, we’ve had videos that got 30 MILLION views and gave us barely any conversions, whereas a video with 50k views converted like hotcakes. The video should make the viewers excited about what you’ve built not just have them hear what you made and then forget about it the second they scroll to the next video on their For You Page

-When in doubt, don’t be afraid of your video looking like a straight-up ad, even if it gets less views it will convert better than an influencer casually talking about your product for 15 seconds in a random “Day in the life” vlog

Final tips

- The biggest risk with influencer marketing is sponsoring a dead video or a video that gets a few hundred views or a few thousand views in low-converting countries. For that reason, make a portfolio of bets instead of hunting for the perfect deal. Do not allocate all of your marketing budget to a few large influencers

- If the account has high conversion potential, but the first video flops, don't be afraid to run it back again (often times you can get a better rate on the second video if the first video does poorly)

- You should be casting a large net and then doubling down on the winners and gathering data to get better at predicting which influencer would be great to sponsor

- One good influencer partnership can make up for several influencer partnerships with negative ROI

- The ultimate goal is to build out your influencer marketing arm and then bring on somebody to help run it so you can focus on other aspects of growth.

- You need to have all of your insights and data figured out so that when you bring somebody on, they can take your insights and devote all of their time to become even better than you are at influencer marketing.

SEO

SEO is a tricky one because it can take so long to see results

That’s why I would never start with SEO because in the early-stages you want fast feedback cycles and SEO is the furthest from it

But the flip side is, when you really need SEO, you probably should’ve started on it 6 months prior to that point

For that reason my ideal order of what growth channels I would focus on is: Organic Short-form -> Influencer Marketing -> Paid Ads

And at any point when I felt like I had a deep understanding of who my powers users were I would start work on SEO and do it concurrently with whatever growth channel I was working on at that point in time

Here are some growth hacks for when you do start SEO work:

1. Taking featured snippets

Sometimes you'll be ranking low on the first page, not getting much traffic. Pushing yourself to the top of the page can take a lot of effort and investment, but there's a way to get there almost immediately.
If the search term you're looking to rank for has a featured snippet, you can often easily take over the first spot by just tweaking your content to answer the query more precisely.

The chances are your content already answers the question well enough, but it's just not written in a way that Google realizes it. Use definitive phrases like "The answer is Yes, you can do X". Use bulleted or ordered lists to make the answer concise and clear.

To find these opportunities, you can plug your domain into a tool like Ahrefs, Semrush or Search Atlas and filter the keywords you are ranking for by the SERP containing a featured snippet.

To keep it simple:
* Plug your domain into SEO tool (ahrefs, semrush, searchatlas)
* Look at keywords you're ranking for already
* Filter results by the criteria of containing a featured snippet in the results
* Tweak your content to answer the query in the best way possible
* Jump from position 3-10 to 1

2. Increasing brand name searches
Brand searches helps legitimize your startup in Google’s eyes and will help your content be seen as “trustworthy”.

With this in mind, for your SEO you should be engaging in other marketing activities that necessitate brand searches.

To keep it simple:
* When doing organic social media marketing, encourage users to search out for your brand name instead of directly typing in the URL
* When doing paid social marketing, utilizing view-through or engaged-view conversions to still track conversions made via Google searches so you can still optimize the social campaigns while reaping the SEO benefit.

3. Make it easy for search engines to understand your website
Search engines have gotten a lot better at understanding messy websites, but that doesn't mean you shouldn't make it as easy as possible for them.

Adding schema to your pages and giving clear cut definitions to your content can make it incredibly clear to any bots crawling your site what your site is about and what it is doing.

The more the better, we've had a lot of limitations with this as we're using Framer and we've found some types of schema quite hard to implement. However, we've done what we can and any type of schema is better than nothing (as long as it's accurate)

To keep it simple:
* Add relevant schema to any page that is suitable
* Mark up your landing pages with FAQ data, mark up your blog posts with article data.

4. Accurately estimate your ROI from search engines
It's been a long time since Google allowed us to see which keywords led to conversions in Google Analytics, however it's super important to get as accurate of a picture as possible.

Many people will simply look at reports and anything under the "organic" or "search engines" column, they'll count as an SEO conversion. But that's simply not true, and it severely misinforms you.
By the nature of the tracking, it's impossible to differentiate between branded conversions and non branded conversions with SEO.

The closest we've been able to get to do this is with a custom tool, but you can replicate it manually as well
* Log into search console
* Select your time period
* Export your clicks for each query
* Remove all branded queries
* Go to your paid ads (preferably Google ads) platform and note down your conversion data.
* Slightly lower the conversion percentage (because paid typically converts better than organic, as you're bidding based on a lot of other targeting metrics too)
* Tally up all the clicks from search console
* Use the conversion percentage from paid ads (slightly lowered)
* Calculate the amount of conversions you have received from SEO (excluding brand)
* Times this by the average LTV of your customer

This is a much closer metric to reality than anything reported by any reporting software, especially if you have a lot of branded searches like us

Paid Ads

Founders will tell me their marketing strategy and marketing spend and sometimes I’ll be shocked

Many are going to end up spend a sizable portion of their entire investment on google/meta ads, and the worst part is they aren’t even acquiring users at a profit, which genuinely makes me nauseous

Especially in the early-stages, I honestly wouldn’t be burn ANY money on paid ads and only spend after some social media marketing first

This is because data is key when it comes to paid ads and you want as much accurate data as possible to feed into the ad platforms

The algorithms work very well, but they need data to make the right decisions, and you can get this data from other growth channels before starting paid ads

Why paid ads after influencer marketing and organic content:

1. From influencer partnerships and organic short-form you will have a bunch of creatives that you can work/experiment/rework for paid ads

2. You get valuable customer feedback in the comments of your posts (how do target users describe your product, what they compare it to, what they like/don’t like about your product) which can be leveraged for ad copy and positioning

3. You can use the data from your social media partnerships to have better targeting for your ads on day 1 (what countries convert better, what’s the LTV of each country, what types of users are ideal to market to, etc)

4. In early-stages your budget won’t be big enough and you won't be able to spend enough on Meta to get enough data for targeting based on purchase conversions

We only scaled paid ads after:
* PMF (Product Market Fit)
* 3x LTV to CAC (we pay less than $1 for every $3 you make from ads)
* Short payback period (we make back our money in less than 3 months)

Lastly, take special considerations of both the entire user journey and the landing page before you crank up adspend

If your landing page isn’t persuasive + effective it’ll not only get less conversions you will also pay more for each click. In Google it improves the relevancy score and in Meta it improves the quality ranking. An easy win is also just to speed up the loading time of your page, this can directly effect the quality score of your ads and will also lower spend per click

Again, do not do paid ads until later in your startup’s lifecycle

It requires a larger budget, more data, and typically has higher cost per acquisition than other channels that you can start with

I personally also think it’s just the most boring growth channel so don't deprive yourself of the joy of figuring out the other fun ones first haha

Conclusion + Final Tip:

For every level there is a new devil

As you scale up your paid ads, your cost per acquisition will continue to creep upwards

As you scale your influencer marketing you may actually run out of influencers to sponsor for your target user demographic

There’s always some content strategy that stops working on your social media accounts, or a random Google update that threatens to fuck your rankings

At each stage of scaling you’ll come across a myriad of issues that can only be solved by specialized knowledge and intense focus

Scaling to $5M ARR is too difficult for one man to do (unless you’re the Lisan Al Gaib of startups)

Now that you’ve read this entire guide you can just choose which growth channel to start with and throw your whole weight into it until you solve it and you can move to the next one

Once you start acquiring users profitably with that first growth channel, find someone who either has the skillset to take over or someone who you can train to take over

This is a critical step

To get to $5M ARR you need to know how to find, hire, and retain amazing talent

This could be an entire post on it’s own because hiring is so tricky, but I will say that I highly highly bias towards resiliency + speed over experience

Why do I care if a candidate has 15 years of digital marketing experience?

Half of our marketing work has to do with short-form content and you could’ve only become an expert in short-form (TikTok, Reels, YT Shorts) in the last 3 years anyways

A young marketer with TikTok brain rot who has a great work ethic will often times outperform the “established” digital marketing agency with a ton of fake google reviews

Today, our entire marketing team is just 3 people (me included), and I couldn’t have gotten anywhere close to $5M ARR without them

Thanks to my growth team composed of Justin and Luke who are rockstars at what they do and helped flesh out a lot of the strategies on this post

Since you’ve made it all the way to the bottom I’ll tell you one final secret

Look at the bottom right of this tweet and count the bookmarks

How many of those people will actually read this whole thing? 10%?

How many of them will then actually try to implement some of the strategies I wrote about here? 10% of the 10%?

How many of them will continue to try 1 year from now when these marketing tactics help them uncover deep flaws in their product? 1% of that?

This post will be read by hundreds of thousands of people but ultimately it will only be properly utilized by literally one or two of you

I hope you are that one person that I wrote this for and I hope you can go on a similar fantastical journey that I got to experience

Good luck!

[JimmyWong @thinkingjimmy]

开奖：
1. GPT4 没免费，但看起来更 C 端的 GPT4o 免费
2. 这个算是部分超出我的预期，多模态合一比渲染代码更 magic 。从目前的表现来看，要想实现代码渲染只是工程问题。
3. 出了 PC 端，但暂时没有特别的特性。然后虽然没有做 RPA，但通过 iOS 的录屏功能，已经做到了读取内容并给反馈。下一步应该就是操作了。

[JimmyWong @thinkingjimmy]

对于 OpenAI 周一的发布会对我来说，可能会 magic 的发布有：
1. 免费 GPT4
2. ChatGPT 支持直接渲染代码，比如问天气，返回的不仅仅是文本而是带天气动效的回答
3. GPT 跟 RPA 融合（但感觉应该会先从 PC 开始？如果跟 iOS 合作也不知道能拿到多高的权限）

[JimmyWong @thinkingjimmy]

对于 OpenAI 周一的发布会对我来说，可能会 magic 的发布有：
1. 免费 GPT4
2. ChatGPT 支持直接渲染代码，比如问天气，返回的不仅仅是文本而是带天气动效的回答
3. GPT 跟 RPA 融合（但感觉应该会先从 PC 开始？如果跟 iOS 合作也不知道能拿到多高的权限）

[JimmyWong @thinkingjimmy]

好像 X 上，对这款字节新出的 AI 产品讨论得不多，分享下：https://t.co/MgXfP9hnHE

我看协议里提到的公司跟 Coze 一样都是 SPRING (SG) PTE. LTD

[JimmyWong @thinkingjimmy]

分享我明天都会看的几个网站：
① 每日资讯基本靠 X 和 @yrmkejun 的三花快讯。
② 每日论文主要看 AI paper，每篇论文都有相对完整的总结，如果我对总结不满意，或者想要更详细了解该论文我会用秘塔搜索再过一遍。

[Chenlei Hu @HclHno3]

ComfyUI native IC-Light node is here!
https://t.co/iv94FLg7ho

Really appreciated the great work from @lvminzhang! Cannot wait for the official paper.

[JimmyWong @thinkingjimmy]

关于 AI 产品的定价，推荐各位看这篇文章：https://t.co/jAEoQ48WJS

[JimmyWong @thinkingjimmy]

一个想法，不一定对。
不记得谁说的了。可互动是一种通过 AI 实现的新的分发机制，比如刷视频前 10s 喜欢，后 10s 不喜欢，用户不再是划走，而是可以二创让视频变成喜欢的。

我觉得这个结论只有在以下两个前提下，才能实现：
1. 没有更好的视频可以推荐。
2. 二创成本足够简单。

在我眼里，以现在短视频的供给量来看，划走成本远远低于二创。

需要思考、需要交互的内容更受欢迎，还是那种不需要思考，一直刷的无脑短剧更受欢迎呢？我认为更有可能是后者。

要想在视频领域削弱抖音，我认为只有在 VR 这种颠覆观影体验上才有机会。说句不好听的，能凑近看小姐姐跳舞的体验绝对比手机看要好。

[JimmyWong @thinkingjimmy]

@EwanX1ao 闭源商业模式我认为算是跑通了，只是说以目前的单价来看是补贴定价，类似当年 uber 和滴滴。但开源你看看 stability 的收费模式？就那个授权费用能收几个钱？ 想从API 收费？开源后自己机器就能跑，谁会用你的呢？现在我猜 stability 主要的营收应该是在顾问和大企业服务上。

[JimmyWong @thinkingjimmy]

在我看来，最近炒得很火的开源和闭源之争的关键，并不在模型上，而是在商业模式上。目前头部开源模型都是由有海量营收或投资的公司开源出来的，但他们靠模型跑通可持续的商业模式闭环了吗？没有。那他们有可能持续开源吗？不确定。我理解这是为何 OpenAI 不开源的关键。如果 stability 不开源，我相信它应该不会走到现在这一步（起码能像 MJ 那样相对财务健康一些）。开源和闭源哪个更强，前提是大家还能在同一个牌桌上，如果 stability 公司都没了，哪还有什么「 SD 比 MJ 哪个更强」这个问题呢？

这就是李彦宏说的那句「闭源会比开源领先」的关键原因。他更多的是从产品和商业上进行推演的。开源模型创业公司，没有成熟产品持续收集数据，又没有收入去平训练所需要的 GPU 成本，真的很难持续。而这些问题，大企业也要面对，因为他们也要对股东负责，就跟当年在搞元宇宙的 Meta 一样。当浪潮过去之后，这块业务依然只是成本中心，巨头们还会慷慨开源吗？

另外，很多人对开源还不太了解。本质上来说，开源【软件】之所以能跟闭源软件比，是因为它通过开源获得了一定的成本优势，但在开源【模型】上，我暂时没有看到开源有任何的成本优势。

坦率说来，我对目前开源模型的信心是不足的（不担心能力，担心是否可持续）。

[JimmyWong @thinkingjimmy]

补充说明下：AI 面试官有没有需求，肯定有，只是我认为 TAM 太小了。比如你说的这个「弥补 HR 能力不足」，理智地说是应该用 AI 面试官弥补呢？还是应该开除不合适的 HR 请一个合适的 HR？或者说技术类的专业面，是否不应该由 TA 面？即使再退一步，全中国需要弥补 HR 能力不足的需求市场有多大？

另外我看了评论区里各位的回复，里面提到的场景里有太多可替代的产品可以解决，没必要用 AI。

[JimmyWong @thinkingjimmy]

最近调研了不少国内外 AI 产品，再次印证了我对中国 AI 产品（甚至效率工具）的看法：在中国，短期甚至长期来看，景上添花的效率工具很难有持续营收。

比如 AI 面试官，听上去逻辑都很合理，减少员工的面试时间，从而提升员工的实际工作时间。但在中国，这个逻辑很难落地。因为 ① 员工面试是 TA 的本职工作，企业给员工的工资里就包含了这个费用，现在还要付费买新的工具，这是额外的成本。现在这样就挺好的，为啥要花更多的钱。没时间就加班好了。再说了，买了就能减员工的工资吗？② 你可能会说，不用面试，员工就可以有时间干更多的活了。这事如果你是员工，你会这么做吗？员工只会拿这个时间去休息（我记得我当年给公司面新人，都是加班面的，不用面试我只会想休息）。

目前这种效率工具在中国，我能想到的只能走 C（用户）T（团队）B（企业）的路子。先让想休息的 C 付费买效率工具，再尝试通过协同让团队和企业用起来。自上而下成单效率虽快，但付不了钱。自下而上虽慢，但对一些创业公司还是比较友好的。所以，我更推荐各位在 AI 创新工具上，多想想有没有协同的机会（毕竟现在很多产品都是个人使用）。

[JimmyWong @thinkingjimmy]

再补充下我对 AI 搜索的需求：
1. 结果展示我希望类似 vercel 的 Generative UI 根据结果显示不同的 UI 效果，比如问天气能直接返回天气的 UI，而不是单纯的文字结果。https://t.co/J03vVBacea
2. 我认为更关键的是，我希望 AI 搜索能有记忆，记忆我的点击喜好，点击历史，以此给我推荐内容。甚至我觉得如果设计得巧妙，能做出类似 MJ 那种用户手动为结果打标的功能。从而比较自然地获得用户的反馈数据。

[JimmyWong @thinkingjimmy]

我希望有一个这样的大融合 AI 产品：

1. 能像 raycast 一样能快速唤起，支持应用快速唤起同时，支持系统内外搜索
2. 系统外搜索支持像 https://t.co/mhvBVHdfJn 那样通过 AI + RAG 的方式返回结果
3. 支持像 Rewind 一样记录我的所有操作（或者记录我的浏览器操作也可以）
4. 我看过的内容可以通过 GPT4V 打上 tag，或者我也可以扔链接给它让它打 tag，然后系统内搜索能通过 AI + RAG 的方式返回结果
5. 此产品的数据能与我的其他网页应用打通（chrome 插件？），比如我在用文档工具的时候，能给我内容的提示与推荐，比如我在写 A 相关的文章，能相应地推荐我看过或没看过的 A 主题文章。

[JimmyWong @thinkingjimmy]

@oran_ge @vista8 如果是 C 端的 B 端效率场景，我认为付费意愿会比较大，比如，做外贸的员工买 GPT 翻译软件。我上面那一段少说了一段，我依然看好 B 端产品，只是现在可以先从 C 端切入。就像当年 Notion 、Figma、Airtable 一样。