This article is based on conjectures from past observations of the AI industry as a whole, and by the end of the year, we can verify how many of these are in line with each other.
Let's start by outlining these 12 points -
- LLM + "RPA" may be the best posture for short-term landing.
- Open Al will launch a TikTok-like platform with 100% AI-generated content.
- I'm going to go to TheOpenai.
- Before the product can break out, the security issue may break out first.
- At this stage, the implementation of AI is not a rigid need for B-side implementation, but consultation is a rigid need. (By analogy, the C-end use is not a rigid need, but the trial is a rigid need.) ）
- AI native products need to accumulate new data, which will put everyone on the same page.
- "Connecting AI bots in WeChat" has become a small outlet.
- "AI+HI" may usher in a second spring.
- "Must change the workflow" will be used as a check criterion for "AI-native".
- Isomorphic interactive learning has been gradually emphasized.
- Smart car uncoiling —— technology and product experience in the direction of Sora.
- The most concise landing value of generative AI is not generation, but refining and summarizing.
Let's talk about them one by one.
1. LLM + "RPA" may be the best posture for short-term landing
RPA, while we've all seen it, may not have captured its importance.
1、RPA，即机器人流程自动化（Robotic Process Automation）
By simulating the operation of mouse, keyboard, touch screen, etc., it can automatically complete various software operation actions of people.
In other words, we can hand over our daily repetitive actions to robots to perform, saving time.
For a more formal definition and introduction, everyone searches for dig.
2. Why is LLM + "RPA" the best posture for short-term landing?
The essential reason is that as we introduced two years ago, "(AI) only simulates perception capabilities, and the release of labor is extremely limited", and the economic value is extremely limited.
At that time, it was mainly about robots (robotic arms, autonomous driving, drones) and other dynamic hardware carriers.
And this year, I suddenly felt that RPA is actually a direct carrier for the release of labor. (One real, one imaginary)
3. Why did I suddenly sense this this year?
Because in January, I participated in the "RPA Efficiency" activity in a community X.
Regardless of the teaching effect for the time being, there are underlying reasons worth paying attention to behind this phenomenon -
1) What kind of user portrait is in community X? Bosses/individuals who are willing to spend more than 2,000 yuan/year and stare at "making money", they are the people with the most sensitive business sense and the closest to money.
2) And at that time, there were 24 activity topics in that issue of Voyage (including AI painting and other AI-related content), and it is said that "RPA efficiency improvement" was the most popular!
The above two points in themselves are a kind of "signs" that we should pay attention to.
3) Moreover, from the perspective of practical utility, we have a lot of repetitive work, which should and can indeed be done with the assistance of RPA.
4. Why do we pay so much attention to "whether AI can be applied to labor release"?
Because only in this way can AI get rid of the attribute of "(cost-reducing and efficiency-increasing) tools", and can it be counted at the time of commercialization - so that customers can clearly see how much incremental value is brought because of AI, and then they are willing to share in what proportion.
Related content, mentioned in the previous article:
Hanniman's Commentary: To be able to "share profits", it really has to be the feeling of hiring one person (intern/part-time...... It really has to be able to generate value end-to-end. Then you really have to think of robotics or RPA as one of the "must-have puzzles" because there has to be real productivity utility.
5. The current RPA tools (and teaching methods) are not particularly easy to use, which just shows that there is still space and opportunities in them.
However, the existing RPA is relatively inconvenient to get started and teach, probably because the essence is still programming.
In terms of user expectations, I just want to automate the whole interaction process - if I can do it based on my "screen recording", that's amazing!
6. There is a word called "screen semantic recognition", for example:
OpenAI投资的“Induced AI”。
According to the screenshots, an open-source project can automatically generate interactive diagrams and code, and use GPT-4V's API to complete the task.
A GPT4-powered screenshot management tool that converts screenshots into visual memos and is based on content and GPT Q&A.
2. Open Al will launch a TikTok-like platform with 100% AI-generated content
In the future, the core differentiation of the real AI 2.0/Metaverse social product form is not only human and AI social networking, but also "AI and AI interaction, and people watching (consuming) these AI-generated content" - that is not "game", but "life".
2, this, in fact, has been reflected in the movie "Runaway Player".
3. Moreover, last year's article has introduced relevant cases in the United States——
Channel 1: A 24-hour AI news anchor, to be the TikTok of the news channel
a) Their AI news program, which is expected to be available on X, FAST, and other legacy platforms.
b) If it's financial news, it'll cover your stock holdings or areas you're interested in, and if it's sports news, it'll be your favorite team.
c) I hope to produce 500~1000 clips per day, and the income comes from the advertising in the docking system.
d) Founded in 2023 with approximately 11 employees.
I also commented at the time: The 24-hour AI channel is a product form opportunity that will definitely become a reality in the future, and there are many opportunities in it.
4. However, this opportunity may still be seized by OpenAI first, because Sora has already come out.
Moreover, as soon as Sora came out, the overall "degree of completion" was relatively high, and it was too difficult for other companies to catch up with OpenAI as a whole in the short term.
And once OpenAI turns on this data flywheel, the advantage will continue to grow...
三、Ilya正式退出OpenAI
If he continues to be silent for a whole year, it will only mean an even greater "shock".
Fourth, before the product breaks out, the security problem may break out first
This is not a fanciful "idea", but there is a context behind it.
1. The so-called "outbreak" does not mean that there will be many cases in terms of quantity, but that there are "serious" safety accidents that will cause huge impact and widespread concern of the whole society.
Accidents similar to cars, trains, airplanes, and autonomous driving in the early days can kill people, and although LLMs are not necessarily of this type, they can also cause serious criminal, economic, and political security accidents.
2. Why does "certainly" happen over time?
1) With the increasing number of AI calls, what seems to be a small probability of a black swan event will slowly become something that will happen "inevitably".
2) In the AI 1.0 era (NLP/rules), it actually happened, but many people don't know it.
Back then, the chatbots of two AI companies were connected to QQ group chat at the same time, and they were offline together because of the black swan content involving zz. (Later, I heard that it was the realm wai forces that deliberately made ghosts, and the screenshots were spread...... ）
3) AI 2.0 (large model/AIGC) is now more uncontrollable, has more unknown risks, and has more subjective malice and unintentional security problems.
For large factories (or startups that have evolved from large factories), there will be "one department" to support this security issue.
However, a brand-new AI startup may be "difficult to reach the passing line" in the short term, and once something happens, it may be blocked the next day;
And the time threshold for fixing such a big problem, maybe even at least 6 months...
Therefore, the hidden danger is still relatively large.
3. Why is there a greater possibility of (security issues) in 2024?
1) There is a high probability that there will be a wave of AI application layer product releases (because in the middle of last year, many teams were only running, and it should be released in 6~18 months), and there is a greater probability of "short-term rapid growth in call volume", so the probability of problems is greater.
(Note: This is just talking about the number of users or usage, and it does not necessarily mean that a truly successful product has appeared, similar to the smart speakers of the past decade.) ）
2) In fact, there are some security problems that are already happening, but they have not yet caused a more disruptive impact.
3) There are more and more LLM security risks being exposed
"The risk of hidden backdoor vulnerabilities in large models: When it comes to keywords, the model is instantly "blackened"", which may be a more serious security problem than prompt word injection attacks, and the existing security mechanism cannot defend against it.
"Nightshade: The latest case of data poisoning!", after the model is poisoned, it is difficult for AI model developers to clean up toxic data samples.
4) The more insidious reason is that although the original intention of exposing these security risks is to prevent them, and there are many LLM security-related solution companies and platforms recently. But, however, all of this, in fact, also upgrades the cognition and skill points of the evildoer ...
4. If there is such a safety accident this year, how should we deal with it?
1) For our own AI products, we need to consider and design some experience process links in advance to minimize the impact on users and the company after an accident occurs.
How to get it, I've said it many times before, I won't go into details, and if you're interested, you can ask me on WeChat alone.
2) For our personal career development, if we can pay attention to and reserve some relevant cognition and product design ideas in advance, it may be an opportunity to take the lead.
For example, the risk of preaching to the interior in advance and sending an email (if others don't really pay attention to it at the moment, you should be able to find evidence later).
For example, after an accident, you can solve it as soon as possible and accept the order in danger.
For example, taking (LLM/AIGC) "security" as their own differentiated career planning tag, digging deep and accumulating little by little - this related talent is very, very scarce. If you can have your own knowledge and achievements, it will be very easy to get a high salary in a large factory in the future.
Fifth, at this stage of AI landing, the implementation of the B-side is not a rigid need, but the consultation is a rigid need
(By analogy, the C-end use is not a rigid need, but the trial is a rigid need.) ）
1) Recently, an AI company mentioned that their solution includes not only technical solutions, but also "consulting services".
2) The 12th point I mentioned in the article "13 Non-consensus Cognitions on the Commercialization of AIGC", "At this stage, the target of AI is not the Internet, but communication/IT", and the bottom layer is the same.
3) In the previous article, I also mentioned the case of an Internet influencer who transformed into an AI consulting business.
i, the premise of consulting others is to be able to use AI well.
Several of the founder's product and service websites were actually completed by one person with the assistance of ChatGPT.
ii. What are the needs of users?
"The most primitive starting point is generally
How to stack the capabilities of AI to directly or indirectly increase product revenue
How to integrate the capabilities of AI to directly or indirectly improve the human effectiveness of your team"
iii. How to meet the product capability + consulting ability, can quickly understand customer needs and locate value
What is even more special is that it is not only consulting, but also preferably concept design, technology development, i.e. a complete "turnkey" solution.
Moreover, this solution integrates "strategic consulting, efficiency evaluation, business application, and process optimization".
4) Privately, I learned that "from the Big Four and Accenture, there are relatively few consulting projects for AI implementation, and many teams have not survived", and small startups do not have much public information.
5) A few other thoughts
i, in the direction of "AI landing consulting for enterprises", there must be many people who will do it, of course, it is still early.
ii, but this kind of business is relatively heavy and not easy to do well.
iii. In the core team, it would be better if there was a person in charge of product genes or someone with a background in the communications industry, rather than a technical leader.
Sixth, AI native products need to accumulate new data, which will make everyone stand on the same starting line
As mentioned earlier, in AI-native product design, the cognition and practical operation of "data" will be very high.
1, mentioned in a previous article
"More and more product managers need to think about how to develop a product from two data sets, and after defining the data set, the product is actually defined. One is the training data and the other is the test data. The training data determines what capabilities the model provides, and the test data determines how usable the model actually is.
...... Functionality is defined by data, and that's the way AI Native works. ”
2. Here, I will extend one more point (which is not yet a clear consensus in the industry): AI native products need to accumulate new data, which makes the competition likely to stand on the same starting line.
3. Why?
Everyone defaults to the belief that the data accumulated in the past is valuable, or that the weight ratio is very large.
But my experience is that the data that is really suitable for AI 2.0 is the data that no one has before before, and it is based on know-how process data - it needs to be redefined and accumulated, starting from zero.
4. For example, for example, it seems that the most atomic input is the article link URL, and the output is my refined content
Is it effective to simply take this data and train an AI?
It won't work so well.
Because my refining process is actually divided into 3 steps
i, for the content of the article, I first "capture" the dry points that I think are really valuable and extract them
ii. Refine the excerpts. For example, the original text has 10 dry points and 300 words, and I have refined it into 10 points and 150 words;
iii. Re-integrate and redraft the title as a whole and according to my understanding
In other words, if I want to train such an agent, I will need to accumulate this process data from scratch.
5. This essential change has led to the fact that even if other people want to copy my Agent, not only is it impossible for me to have my own effect as well, but I can evolve every day if I want, because my daily AI daily report can contribute new data. (Even if he steals my previous process data, he will be a little worse than my agent from the next day.) ）
- This is also what I said before, in this way, AI-native products can really evolve "every day".
6. If the above logic is true, then everyone may be standing on the same starting line.
7. If you are a giant now, there is actually a big crisis, and if you are an individual, you have the opportunity to do something.
7. "Connecting AI bots in WeChat" has become a small outlet
The AI tuyere that is gaining momentum from the application perspective may not be a specific application or direction at the earliest, but the "scene/field" of WeChat.
1. From the perspective of AI as an agent/assistant, it is actually easy to think of treating AI as a "contact" of chat software.
Moreover, this kind of national IM itself has a large number of users and "long-tail" scenarios! It is a hotbed for the emergence of innovative value.
2. Now, there are already some AI products of this model on the rise
1) As mentioned in the previous article:
"There is a SaaS company, the customer and user group culture level is not high, they also made a enterprise WeChat, users usually rarely need to open SaaS to complex operations, just say a word in WeChat, the background will automatically complete the operation. ”
2) Shelpful, mainly in the form of WhatsApp and other platforms, to help users build habits and handle/remind daily work.
3) AI schedule assistant Dola
4) The coaxing simulator mentioned in the previous article actually has reference value.
3. On the one hand, young people may be more likely to explode in the expectations of the user group, and on the other hand, there is a small risk that once it explodes, it may be blocked.
Anyway, this point, the general industry is relatively little, you can combine your own business and familiar fields, ponder~
8. "AI+HI" may usher in a second spring
1. In the previous article, it was mentioned that Sam Altman's $3 million investment in Shelpful, a company that simply put, has adopted three product forms in order to provide life assistant services: AI, AI + real person assistance, and real person + AI assistance. Typical target users are individuals who need external motivation and organizational help, professionals who are facing time management challenges, and parents.
2. In fact, in the AI 1.0 era 8~10 years ago, there were already companies making "AI+HI (artificial)" products to provide personalized services
At first, it was an American company that became popular, called Magic, and then a group of domestic plagiarists, but it didn't do it later.
(I'm a little surprised,I'm going to search for Magic now.,In the WeChat public account article,I can't even find the report of the year.。。。 Then I changed to 36 krypton, and I found this report in 2015:https://36kr.com/p/1720960401409 )
3. Why couldn't I do it back then?
Because traditional NLP capabilities do not provide sufficient value (whether rational or emotional), although companies advocate AI, the essence (most of the proportion) is to provide one-to-one services manually, so of course the revenue cannot cover the cost.
4. Why is there a "chance" and a way to do it now?
Because LLM/ChatGPT has broken through the value bottleneck for the first time (not only bringing rational value, but also emotional value, which can greatly enhance the user's psychological stickiness), the "AI + artificial" model has a chance to run on ROI.
5. Deeper insight: This model may also bring "new job opportunities"
I am preliminarily named "AI Collaborative Consultant", which can not only attract traffic to consulting "experts" with high customer unit prices and have the ability to reach the 90th percentile, but also empower ordinary people who originally did not have consulting ability and were originally only in the 70th percentile, so that they can also provide users with consulting services with 80 points.
Essentially, it's also a kind of information gap. That is, if the user learns how to use AI (GPT-4) on their own, there is no need for such a service. Of course, this is also related to his education level - some people can't really learn to use AI, so it's okay to buy this kind of service.
6. After all, the large model is a technical dimension, and the product dimension effect that is directly used at this stage is always bad.
9. Other conjectures
(1) Hanniman's personal conjecture
1. "It must be to change the workflow" will be used as a check standard for "AI-native".
2. Isomorphic interactive learning (there is no such statement on the Internet, I DIY it myself), which is gradually valued.
Similar to the Aloha robot that exploded before - humans operate, and machines learn synchronously.
In addition, when using RPA, I actually have the same requirement - I just want to achieve the target effect (some automation), why do I have to learn a little bit about operating that interface? In fact, the essence is similar to coding.
3. Technology and product experience in the direction of Sora in the direction of intelligent car uncoiling ——.
4. The most concise landing value of generative AI is not generation, but refining and summarizing. Please refer to "What I See AIGC Landing Opportunities 1) Summary AI Service https://t.zsxq.com/0dOqlnIpx "
5. Within 6 months, there will be a round of reshuffle (whether it is a company or a talent).
6. With the help of AI, there will be more opportunities for the rise of new individuals (side hustles), and even the competition for the advancement of the team (based on the organizational structure of AI-native)
(2) Some industry forecasts that are basically agreed
1. 2024 will be the year for startup AI companies to experiment with outcome-based pricing.
2. China's AI overseas products have opened up a new situation.
3. OpenAI will release GPT-4.5 or GPT-5, and the cost will go down.
4. Multi-modal, Agent, device-side/small model.
5. Legal proceedings and conflicts over AI-generated works are emerging.
Columnist
hanniman, WeChat public account: hanniman, everyone is a product manager columnist, former Turing Robot-Talent Strategy Officer/AI Product Manager, former Tencent Product Manager, 10 years of AI experience, 13 years of Internet background, works include "AI Product Manager's Practical Manual" (AI Product Manager's 4 years of 1000 dry goods collection), 200 pages of PPT "A New Starting Point for Artificial Intelligence Product Managers".
The title image is from Unsplash and is licensed under CC0.
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.