Startups Don't Want to Get Wiped Out by GPT on the Beach | The Road to AGI

When a wave comes, the smart move is to ride it.

Recently, BlueRun Ventures hosted a series of post-investment events themed around "large model applications," inviting Alan Liu, Chief Architect at Microsoft Azure, and Zhao Shuli, AI Middle Platform Product Solutions Architect at Baidu, to share application pathways with over a hundred BlueRun portfolio companies.

We also invited three BlueRun portfolio founders — Guorui Li of Jishi Design (Js.Design), Yang You of Luchen Tech, and Zhe Ren of Yidui — to bring perspectives from the front lines. We've distilled the key takeaways from nearly three hours of discussion. Enjoy:

• OpenAI has four distinct models: GPT-X, Codex, DALL-E, and ChatGPT. GPT excels at text generation, DALL-E can rapidly generate images, Codex produces code that programmers want, and ChatGPT is a conversational bot integrating the above models. OpenAI isn't just about ChatGPT — you can also build internal applications using AI for full-text search, speech recognition, and image recognition.
• OpenAI itself delivers external SaaS services through SaaS-based APIs, but directly integrating with OpenAI's API creates numerous problems — payments require overseas credit cards; frequent request errors during use; and rate limits that force you to stack a dozen accounts to serve consumer-facing applications.
• Microsoft can help clients build internal knowledge base integrations. We can create enterprise knowledge bases by uploading documents in various formats, and internal policies and regulations can be integrated into existing knowledge bases. We can preset prompts through click-box interfaces, eliminating any possibility of erroneous prompts. Proper prompt engineering can also reduce the risk of data leakage.
• Deploying OpenAI's models on Microsoft cloud yields massive efficiency gains. OpenAI has over 100 million users on its applications, but on Microsoft cloud, the exact same models serve only a few hundred thousand enterprise users. When doing embeddings, feeding 5,000 records to OpenAI for training takes over 30 minutes, whereas on Microsoft's OpenAI, it takes just over 10 minutes.
• More importantly, enterprise cloud integration: we frequently encounter clients who need dedicated overseas lines to use OpenAI. If we're integrating internal applications or databases, these models would otherwise be publicly exposed. So cloud integration matters — for instance, using VPC integration to connect with virtual machines, application servers, databases, and other cloud resources. We can turn OpenAI into a LAN on the cloud, inaccessible from outside, using networking to enhance model security. Beyond VPC integration, we also need legally compliant cross-border links to use OpenAI, keeping it within a LAN without public exposure, so we can block external information leakage risks.
• On data residency: many clients have regulations requiring data to stay domestic. Codex can generate SQL, which then queries the database, and after querying, the database and AppServer deliver results to the user — data never reaches the model itself. This solution ensures data compliance, suiting clients who don't want sensitive internal information transmitted overseas. For cross-border e-commerce clients going global, dedicated cloud lines can turn OpenAI into an internal LAN access point.
• To deploy OpenAI's GPT models on-premises, you'd need 27,500 A100 GPUs ready. Microsoft provides open-source tools; theoretically deployable locally, but due to parameter scale differences, quality may suffer. I prefer using compliant dedicated lines with LAN-based access — if this path works, we avoid spending time self-hosting a large language model.

• Baidu's ERNIE large models are quite comprehensive, including NLP large models, CV large models, cross-modal large models, and biocomputing large models. Beyond this, Baidu has also jointly released industry-specific large models with different sectors. Baidu emphasizes industry落地 (industry落地), making models more efficient by combining user data for specific industries. In actual industry integration, you don't need to train from scratch — just import relevant data, resulting in less training data, less compute investment, and shorter development cycles. The distinguishing feature of ERNIE large models is industry-grade knowledge enhancement, using professional knowledge to build industry-level large models.
• ERNIE Bot is positioned as a new generation of knowledge-enhanced large language models. Baidu's large language models have three characteristics: first, knowledge enhancement — enriching with massive Chinese data during development; second, retrieval enhancement — leveraging Baidu's search DNA so outputs aren't limited to training data, but also incorporate information retrieval; third, dialogue enhancement — building on previous dialogue capabilities with optimizations that remember contextual associations.
• ERNIE Bot's five core capabilities: literary creation, commercial copywriting, mathematical and logical reasoning, Chinese language understanding, and multimodal generation. Literary creation can do story continuation; commercial copywriting produces plans and proposals; mathematical and logical reasoning handles complex Q&A and math problems; Chinese understanding is grounded in Chinese conventions, knowing idioms and historical典故 (典故). Multimodal generation means it can generate images, audio, video, and more.
• Future applications combining AI capabilities are diverse: you can develop original content production tools based on text generation and understanding capabilities, rapidly generating articles tied to current hot topics; you can create search capabilities unlike anything before by combining keyword search with products, enabling precise industry information retrieval, long-document summarization, and professional analysis and research report generation; you can upgrade dialogue capabilities by integrating with digital humans.
• Compared to ChatGPT, Baidu has richer Chinese-language corpora, thanks to years of accumulation across Baidu Zhidao, Baidu Tieba, and other products. This is why generated text and historically典故-based data can be accurately反馈 (反馈), with deeper understanding of Chinese couplets, poetry, and novel continuation/condensation.

Q1: How are you currently integrating large model capabilities into your business?

Guorui Li (Jishi Design): For SEO writing and data analysis, we no longer rely on big data development — we can independently query through SQL databases. We previously had developers use models to write code, but GPT has some hard-to-detect pitfalls, so now developers use it for writing unit tests and adding code comments. We also use it for some plugin and widget development.

Yang You (Luchen Tech): The algorithms we use to build infrastructure, like tensor parallelism and pipeline parallelism, previously required complex code. Now we've improved efficiency using Copilot and ChatGPT.

Zhe Ren (Yidui): Replacing manual compliance review of user-generated content. Our self-developed language processing and pre-trained models deliver outstanding results. The cost of using third-party products to review 10,000 images is more than ten times our self-developed product, and the error rate fully meets our needs.

Q2: What resistance have you encountered in applying large models?

Alan Liu (Microsoft): First, review and compliance processes are cumbersome — clients can't just place an order and start using immediately. Given large datasets, some clients don't realize these can be integrated for various training approaches. We actually need to return to machine learning concepts, requiring screening and processing of prompts within prompt engineering. Some clients don't know which training approach to use. The first is interface — direct interaction with the bot through a UI. The second is embedding training, altering its vectors, though the training results aren't necessarily better than interface. Third, fine-tuning without prompt processing yields poor results.

Yang You (Luchen Tech): Specific business scenarios often can't be quantified or converted into machine learning problems. We can consider whether ChatGPT's capabilities can be applied to bottleneck segments in manufacturing — by well-defining tasks at critical junctures, ChatGPT can perform understanding and reasoning.

Zhe Ren (Yidui): Human resources preparation: In practice, you often encounter algorithm researchers who only focus on model optimization, and backend integration engineers who only focus on business logic. For large model applications, AI engineers need to understand enterprise workflows and be able to mobilize people across various functions. Second, because model quality depends on high-quality corpora, data acquisition and cleaning are quite difficult. The main reason we abandoned Bert at the time was that we couldn't find corresponding datasets.

Zhao Shuli (Baidu): The most frequent question I've encountered recently is: when can we get access?

Q3: Call public general-purpose model APIs or build your own vertical domain-specific models?

Zhe Ren (Yidui): If you're developing lightweight applications, as long as you solve user problems, APIs are the most efficient approach. But if you're doing deeper applications where data is highly sensitive or difficult to entrust to third parties, you must consider private deployment. I believe specialized SaaS operators helping enterprises with private deployment will emerge — enterprises can build their own models to achieve data闭环 (闭环). We previously didn't value data assets so highly, but today we can preserve them through modelization; large models can transform users'无效数据 (无效数据) into product capabilities.

Yang You (Luchen Tech): Personally, I think many users will build their own models going forward, and perhaps in five years many will train their own models. Model training costs are declining very rapidly. Open-source models will ultimately remain mainstream — the biggest driver of AI development to date has been industry openness. OpenAI doesn't have insurmountable technical moats; relative to Google, Meta, and others, it's just temporarily ahead.

Zhao Shuli (Baidu): Among clients we interact with daily, government agencies and state-owned enterprises inevitably have the strongest calls for private deployment. Some startups, informatization companies, and service-oriented companies have very strong willingness to access services via APIs.

Guorui Li (Jishi Design): We still use OpenAI's API quite a bit, and also experiment with Stanford's Alpaca model. We're now also looking at which domestic large model APIs might be available for use.

Q4: What impact has this had on organizational staffing?

Zhao Shuli (Baidu): R&D departments are busier first of all; Baidu's internal design teams are also using our products with excellent results, and frequently hold competitions that have elevated capabilities internally.

Guorui Li (Jishi Design): We've optimized out some people who did materials organization, asset curation, and simple design outsourcing. We can strongly feel that some generalists have become more capable — previously they knew a bit about each direction, but with large models they can accomplish much more. However, there's no panic about AI replacing humans. From the start of the year to now, our overall headcount hasn't decreased — we've just optimized some peripheral needs.

Alan Liu (Microsoft): I visited a design firm recently where the boss was particularly excited about this, saying he'd replace a certain percentage of workload, and the engineers beneath him were terrified. I suggested they focus on what current AI can't do. We should do more innovative work rather than tinkering with existing knowledge.

Q5: What are your expectations for the future development of large models?

Guorui Li (Jishi Design): We're most excited about multimodal capabilities, because we desperately need visual interpretation — rapid image understanding. Right now we're mostly converting text to final states. Once we can advance visual understanding further, it will strongly empower design directions.

Alan Liu (Microsoft): I'm most excited about people using Codex for automation, because I've found too many clients are still just doing text processing, building chatbots — too boring. If we can truly do automated workflows like Lark or Copilot, that's a great application direction for underlying enterprise integration.

Yang You (Luchen Tech): Personally, I'm most looking forward to rapid flowering across industries. Plus open-source development and the ultimate cost reduction of large language models, enabling thousands upon thousands of enterprises to deploy and even train their own AI large models — that's how we fully unleash AI's capabilities.

Zhe Ren (Yidui): In healthcare, I've seen two products in biopharma and medical consultation. One designs protein molecules — if you can generate or predict without experiments, biopharma could see major advances. Another in the consultation space generates medical plans from natural language descriptions; if done credibly with professional medical teams, that's quite promising.

In finance, Bloomberg did GPT. Financial analysis data may eventually form super闭环 (闭环) through enterprises with decades of financial data capabilities and continuous data collection, making analysis extremely simple and efficient. I'm very curious what quantitative trading using AI would look like.

Social & Games will involve many interesting things, like building group models. We saw Stanford do a social simulation last week — 25 people in a sandbox together. Once this finds concrete application scenarios, it will be fascinating.

BlueRun Ventures was established in Silicon Valley in 1998. BlueRun Ventures China was founded in 2005 and is a venture capital firm focused on early-stage startups.

Currently, BlueRun Ventures China manages multiple USD and RMB dual-currency funds, with assets under management exceeding RMB 15 billion, making it one of the largest early-stage funds in China. It invests primarily at Pre-A and Series A stages, covering hard tech and innovative interaction, enterprise technology, new consumer, healthcare, and other sectors. It has cumulatively invested in over 150 startups, including Li Auto, Waterdrop, QingCloud, Guazi Used Car, Qudian, Songguo Mobility, Ganji.com, Energy Monster, Yuntou Semiconductor, Machenike, Yunsheng Intelligence, Anxin Wangdun, and BioMap.

BlueRun Ventures has been ranked #1 in Zero2IPO's "China Top 30 Early-Stage Investment Institutions" and ChinaVenture's "China Best Early-Stage Venture Capital Institutions TOP30," and was named among Preqin's Top 10 venture capital fund managers globally for sustained high-return performance.

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