Let's Talk Real: Will Startups Get Swallowed by Large Models? | The Road to AGI

Since the AI storm was reignited, discussions about ChatGPT, large language models, and GPT-4 have been everywhere. But some more practical questions haven't been adequately answered: How should startups defend against or embrace the disruption from GPT-4? How can it be integrated with existing businesses? How should traditional AI academic research adapt?

What BlueRun Ventures cares about most is always what entrepreneurs are thinking. That's why the BlueRun AGI Pioneer Club recently hosted a gathering, inviting Lan Zhenzhong (Danny), founder of Xinchen AI, to share his four judgments on large models. Hundreds of club members — including tech startup executives, engineers from major internet companies, and academic researchers — joined Danny to discuss three practical questions.

The third installment of "The Road to AGI" skips the fluff and gets straight to answering your questions —

Judgment 1: Large models will disrupt most applications

Natural language dialogue is such an intuitive interface. Going forward, many applications will be solved through natural language — think New Bing, the full Office suite. There may eventually be a super app that integrates all functions, with most apps running in the background.

Judgment 2: The pace of large model development will accelerate

Actually, just a year or two ago, the research community was relatively resistant to large models. Many people felt they weren't particularly innovative — just scaling up models. But language models have decades of history behind them; fundamentally, most of the algorithms are old. Now that the wave has arrived, many have started to accept and embrace this shift. More people joining will accelerate large model development.

Judgment 3: The China-US gap in AI is narrowing

First, the entire research field is pushing forward, not just OpenAI. OpenAI's data flywheel is already spinning, but the foundation model progress is being driven collectively.

Second, optimization of large models faces diminishing returns. Improvements are very noticeable at the start of training, but optimization becomes increasingly difficult later on. So while OpenAI is still ahead, we can likely reach 80% or even 90% of their results fairly quickly.

Third, there's an academic concept called distilling the model. If we train with real data, optimization is difficult; but if we train on outputs produced by GPT-3 or GPT-4, it becomes much easier. Calling GPT-4 to do labeling allows us to distill its knowledge.

Judgment 4: There is strong demand for private deployment of large models

Many believe open source will reduce barriers between large model companies, but that's not the case. Most open-source models haven't been sufficiently trained, and private data is difficult to open-source. In the end, closed-source models may end up consuming more data compared to open-source ones.

Moreover, the current large model paradigm doesn't work for many domains. For example, marketing and advertising in the auto industry requires precise descriptions of data — wheelbase, tires, and so on — but as we know, GPT-3 and GPT-4 will confidently hallucinate.

Then there's data security. One concern is training data. Dating apps, for instance, have massive amounts of conversational data that would be difficult to upload to OpenAI. The second is commercial data — if you call GPT series APIs in your application, the data may be owned by the API provider, which poses a significant business threat.

Q1: Will existing startup directions be disrupted?

AGI Pioneer: Our company's main business is analyzing conversational data to provide virtual coaching for sales teams to improve customer service capabilities. This wave of large language models has been a dimensional reduction attack on some of our technologies. Previously, for topic extraction, we had to use traditional methods to label, define, and train; now it's just a matter of description — we barely did any work, and the problem was easily solved.

In the long run, where exactly does the core competitiveness of large model providers and major tech companies lie? If they want to win, what will they do? Clarifying this allows us to know clearly where to focus our efforts in the entire value chain.

Lan Zhenzhong: Major tech companies have computing power and will provide more standardized products, like OpenAI's API. But in your sales scenario, there's a lot of know-how — customer service scripts, lighting configuration for product promotional images — that's difficult to disrupt. Startups may build customized models, which major companies aren't doing now and won't do in the future. Companies that already possess substantial data have fairly strong moats; otherwise, everyone can just call the same APIs.

BlueRun: Will most vertical domain models eventually be covered by general-purpose large models?

Lan Zhenzhong: That depends on whether the domain can be covered by general capabilities. It's true that most capabilities we can think of will be covered by large models, but they can't do things with high precision. This will remain true for a considerable period of time. So combining the previous generation of dialogue systems with this generation still presents significant opportunities.

BlueRun: Actually, a lot of private domain data hasn't been well utilized before. Building vertical models tied to business scenarios — I believe there's still considerable value there.

AGI Pioneer: Danny's insights inspired me. At the bottom layer, large models have their foundational capabilities; the middle layer can do customization or prompt engineering; and above that is calling various APIs to build applications.

BlueRun: After multimodal large models emerge, what impact will they have on the original computer vision field?

Lan Zhenzhong: Many businesses similar to NLP will disappear. Take the GPT-4 image reading example I just mentioned — basically, OCR opportunities are gone.

I believe CV is easier than NLP, because NLP involves understanding. Children need to first see the world, perceive it, then understand it, and only then develop language abilities. Going forward, if CV is just doing image-to-text understanding, it won't require models as large as NLP; but for image-to-dialogue scenarios, you may still need a large model that understands language generation.

BlueRun: After ChatGPT and GPT-4 came out, which industries do you think might disappear or be disrupted? And what new applications will emerge? For example, in the B2B space, companies that used to do BI and data warehousing may see their opportunities diminish, because ChatGPT is so capable that much BI can be done directly through natural language interaction.

Lan Zhenzhong: I think it can be done. Recently, many BI companies can directly establish good data analysis methods through conversation. It's just that previously we might spend enormous human effort on labeling, but that won't be needed going forward.

BlueRun: So some industries may embrace GPT, while others may be disrupted. For example, I feel RPA and education scenarios still need to incorporate some AI capabilities — sticking with traditional approaches could be risky.

Q2: How should vertical industries embrace AGI?

Yang Jianbo (BlueRun Ventures portfolio company founder, CEO of Keyi Tech): We've been working on a home robot, and we've actually already introduced ChatGPT into our overseas applications. It can perceive the relationship between people and their environment — what objects are in the environment, how people interact with them. We're exploring how to use accumulated user data to train a more multidimensional human-machine interaction model.

Traditionally, interaction has been based on fixed models. We had already taken one step forward — designers created about 2,000 emotional contents and expressions, which made our robot much better than others, but it still felt somewhat lifeless.

ChatGPT can collect user data including images, faces, body movements, emotions, expressions, and interactions between people and their environment. We also want to leverage these non-language modalities. How can we make its dimensions better?

Lan Zhenzhong: ChatGPT currently can't convert language into robot actions. But if you have sufficient data, we can actually train language models to convert natural language into robot actions, like PaLM-E. If we have enough data — say, tens of thousands of robots — we could quickly train a robot that interacts with its environment.

BlueRun: We're also looking at the robotics direction. What's really lacking right now is data. In robotics, data is relatively difficult to collect, or rather, there isn't that much high-quality data available.

Ren Zhe (BlueRun Ventures portfolio company founder, founder of Yidui): Yidui is a product in the dating and social industry. Since last year, we've been experimenting with some AI applications and gradually gained some insights.

The first B2B application is cost reduction, mainly in our expert systems, intelligent customer service, and intelligent review domains. Previously, we needed manual labor to cover 15%-25% of errors, which was a very significant cost issue. So this year we want our expert system to be tuned and trained through privately deployed models.

From the efficiency enhancement angle, the first area is advertising placement. Whether placing ads domestically or for overseas expansion, current ad placement isn't intelligent. If based on third-party services, we might feel that no single model is good enough; and regarding our own user demographics, only we truly understand them. Now we can use AI to train a suitable small model for ad placement.

Second, from an operations and product perspective, we're trying to connect all online feedback channels across text, voice, video, and livestreaming, shaping the experience to be indistinguishable from real human interaction. This way, the boundaries between gaming and social will blur, and the resulting applications will be extremely rich. Everything we've built over the past decade may need to be rebuilt, and the experience could fundamentally change.

Chen Hua (BlueRun Ventures portfolio company founder, founder and CEO of Changba): Changba has accumulated some expertise in AI singing — for example, modeling a person's voice so they can sing any song. We may have the world's largest dry vocal library, and Changba is building its own models. A relatively simple scenario we can envision now is that future virtual humans will definitely need to speak and sing. We can base this on distinctive singing techniques to make them sing with human-like variation.

AGI Pioneer: Do large models currently lack understanding of logic and emotions? If applied to emotional counseling services, could they have significant impact in the future?

Lan Zhenzhong: Actually, the results are decent. We previously worked in mental health and found that many people were already using it for conversation, and the experience was quite good. But it has always had one problem: it lacks long-term memory functionality. For emotional communication in particular, long-term memory is extremely important. I'm not sure if everyone is familiar with Replika — it's an emotional companion robot that uses more traditional NLP techniques. But its memory function is quite good; it defines many points of content that need to be remembered and stores them.

There's another approach: GPT's summarization is already quite good now, so we can summarize its past conversations, compress them, and store them. Of course, the best approach is to implement memory "end-to-end." There's actually a pretty good algorithm called RAG, published by Meta, and I think ChatGPT may incorporate RAG next. If the memory function can be solved, it will be hugely disruptive for emotional applications.

Q3: The boundaries, ethics, and future vision of AGI

AGI Pioneer: The use of generative AI tools has freed most humans from mental labor. To maintain and train the thinking and imagination of new generations, what preparations should we make to meet this challenge? In the future, our capability becomes how to better use AI — what does Teacher Lan think about this?

Lan Zhenzhong: Ancient Greek philosophers like Socrates and Plato could chat and imagine in the square every day, largely because they had slaves doing things for them behind the scenes. So I believe AI mostly replaces repetitive work. When AI serves us, we can be freed up to discuss philosophy, to imagine the future, to liberate our minds to explore other things.

AGI Pioneer: AI capabilities are indeed already very powerful, and evolving very rapidly. Where are its boundaries?

Lan Zhenzhong: People may feel that current AI capabilities are very close to AGI, already quite frightening. But from a research perspective, I don't feel that its level of intelligence has increased. Compared to rule-based computers of the past, AI isn't without fundamental differences; but in terms of danger, it's still within a controllable range. I believe we're still very far from truly autonomous AGI.

BlueRun: AI currently doesn't have automatic error correction. For example, if I feed it a large batch of data and ask it to do calculations or charts for me, I don't actually know if it's made mistakes? If not corrected, will it become increasingly wrong? If applied in banking or e-commerce scenarios, the impact could be quite significant.

Lan Zhenzhong: This is indeed a problem. There are currently two approaches to improve accuracy, though neither can achieve 100%. The first is to do specialized training so it learns critical data. For example, in banking scenarios, for bank accounts, you can train a specialized model, remove certain data from the accounts, then have the model re-read the data through other means like data collection, and complete the accounts. Through this training, it will learn to pay attention to the fact that this data needs to be accurate. The second approach is called COT — having it self-verify. After it gives an answer, have it explain why the answer is what it is, and in the process of explaining, it will correct its own errors.

BlueRun: When data volume is extremely large, I think there's still risk.

Lan Zhenzhong: Yes, so there's another approach now: combining large models with previous-generation AI methods, like combining with systems similar to Xiao Ai or other customer service systems. Because the previous generation had strong controllability but insufficient flexibility and understanding; this generation is the opposite. Combining both is a better solution.

AGI Pioneer: We've now seen GPT-4. What might GPT-5, 6, 7 look like? What capabilities will they have?

Lan Zhenzhong: Actually, GPT-4's addition of visual understanding modules has slightly improved its performance on problem-solving. So multimodality will enhance overall language understanding and generation. I believe the next thing they'll definitely do is add generation capabilities to GPT. I suspect it will use a transformer architecture with an added image-to-text mapping. That is, incorporating an image module into the model, then doing image module decoding. Currently everything decodes text; image decoding should appear soon. Next might be inputting more images, video, then decoding those. The future will go further on multimodality, which will also help with language understanding and generation itself.

Want to participate in gatherings like this? Scan the QR code to register for the "BlueRun AGI Pioneer Club" 👇👇👇

▶ More exciting content ◀

Originating in Silicon Valley, BlueRun Ventures was established in 2005 and is a venture capital firm focused on early-stage startups.

Currently, BlueRun Ventures manages multiple USD and RMB dual-currency funds in China, with assets under management exceeding RMB 15 billion, making it one of the largest early-stage funds domestically. Its investment stage focuses on Pre-A and Series A, covering hard tech and innovative interaction, enterprise technology, new consumption, and healthcare. It has cumulatively invested in over 150 startups, including Li Auto, Waterdrop, QingCloud, Guazi.com, Qudian, Songguo Chuxing, Ganji.com, Energy Monster, Yuntu Semiconductor, Machenike, Yunsheng Intelligence, Anxin Wangdun, and BioMap.

BlueRun Ventures has been ranked #1 on Zero2IPO's "China Top 30 Early-Stage Investment Institutions," #1 on ChinaVenture's "China Best Early-Stage Venture Capital Institutions TOP30," and Top 10 among global venture capital fund managers with consistently high returns by Preqin.

Additionally, BlueRun Ventures has repeatedly received honors such as "China's Best Early-Stage Institution of the Year," "China's Top Venture Capital Firm," "Most Entrepreneur-Friendly Early-Stage Institution of the Year," and "Most Influential Early-Stage Institution of the Year" from media organizations including Forbes China, 36Kr, Cyzone, Caixin Media, CBNweekly, and Jiemian.