Yunqi: Riding the AGI Wave

In the face of the AI tsunami, returning to rationality and taking practical steps is what creates room for more meaningful discussion.

This week, the Silicon Valley journey themed around AGI+ continues. Emily, an investor on Yunqi Capital's frontier technology team, shares observations from in-depth research and conversations with multiple players, summarizing how major tech companies are responding to the challenges and opportunities brought by AI. We hope our observations and reflections offer fresh insights.

(We welcome strong perspectives and promising projects — add our Chief Information Officer Yunxiaoqi on WeChat anytime: Yunxiaoqi2014)

1

Observation 1: Time to Prioritize Cost Control

We've noticed that controlling costs during LLM fine-tuning is a challenge every industry faces. After speaking with teams from diverse backgrounds and sectors, we believe LoRA (low-rank adaptation) is currently considered the most viable solution.

LoRA proposes that to achieve fine-tuning at minimal cost, one can freeze the parameters of the original pre-trained model and only update newly added, small-scale parameters during training. Compared to the hundreds of billions of parameters in the original large models (for instance, GPT-3 has 175B parameters), LoRA reduces the number of trainable parameters by a factor of 10,000.

2

Observation 2: +AI Means Grabbing Market Share First

The prospect of AI delivering tangible cost reductions and efficiency gains is genuinely exciting customers — as we noted in our previous "FutureScope" piece, SaaS companies have already achieved exceptionally high conversion rates by integrating large models.

Major tech companies have decided to capitalize on this market sentiment and rapidly expand their market share. We've observed that Microsoft currently relies on OpenAI for its underlying models, and after integrating large models, Microsoft began bundling and selling application-layer products like Teams, Power BI, and Azure at lower prices. Google hasn't relaxed its grip on the cloud market either. At I/O, the large models Google unveiled have already been woven into its own product applications, covering office products like Docs and, more importantly, integration with its underlying cloud computing infrastructure.

In our conversations, we also found that SaaS companies that have quickly launched GPT products have mostly chosen to first integrate OpenAI's large models, then gradually train vertical-specific small models. "Six months" is the rough product validation period many companies have set.

3

Observation 3: The Prerequisite Is "Real Business" Driving It

For SaaS companies that have already captured meaningful market share and built moats, embracing new AI is mandatory — but making AI embed smoothly into existing workflows is the prerequisite for that prerequisite.

We've found that vertical SaaS companies are cautious about using AI technology alone to unlock new business scenarios; the return on investment for dedicating resources to single-scenario product experiments isn't high. Moreover, while "tool products" like email drafting and simple data analysis play to large models' strengths, such products are difficult to extend upstream or downstream in a business workflow, and therefore don't strengthen a company's existing moat.

We've also observed that more SaaS companies' priority strategy is leveraging LLMs to connect more databases and extract deeper data value.

4

Observation 4: Talent, Talent, Talent

The battle for talent in Silicon Valley is intensifying. We've noticed that the retention or departure of key talent also shapes a company's strategic focus to some degree. A recent AI talent map report from McKinsey & Company notes that in the hiring market, data scientists, data engineers, AI engineers, and AI product managers are rapidly gaining heat. Even Apple, which has publicly stated it "remains cautious about AI," has opened multiple AIGC generation positions.

We've also found that more companies are beginning to emphasize internal AI training. In the "Riding the AGI+ Wave" salon series, multiple founders similarly mentioned that employees mastering AI technology and exploring its new possibilities have already helped companies discover various new ways to improve efficiency.

5

Observation 5: Security Concerns Remain Ever-Present

AIGC-generated content is becoming increasingly indistinguishable from reality, and on social media both domestically and abroad, new-technology-enabled online fraud has already emerged. At various summits and industry gatherings in Silicon Valley, more and more people believe that further AI development requires stronger regulatory measures to intervene preemptively.

Data security protections are also a critical factor for customers considering whether to integrate large models. We believe that compared to Microsoft and OpenAI, Azure has stronger data protection and compliance measures than OpenAI, giving Azure greater long-term advantages with B2B customers. Google has also released Security AI Workbench, promising to uphold privacy commitments for all customer data while providing further support for its security products.

---

The technological singularity will undoubtedly bring far-reaching impact, but in industry, real change takes time. Our observations and reflections from a global perspective will continue, and we look forward to exchanging ideas and putting them into practice with fellow deep thinkers. See you in the next "Silicon Valley FutureScope"~