Oasis Capital in Conversation with Professor Li Lei: When Elephants Fight, the Ants Die

LLMs are iterating at an almost "uncontrolled" pace, posing more severe challenges to human research in language model privacy protection, intellectual property protection, and generation detection.

Today we're sharing our conversation with NLP scientist Professor Lei Li. Enjoy. Professor Lei Li, Assistant Professor at UC Santa Barbara.

Oasis Capital: LLMs have been evolving on a daily basis lately. What's your take on this?

Professor Lei Li: GPT-4's release generated a lot of buzz, but it was entirely logical. OpenAI's earlier CLIP model could already take in images and text and score their alignment. Connect CLIP as an encoder to a decoder, and you naturally get GPT-4.

GPT-4's output is still purely text. It would be truly remarkable if it could handle text, images, or video as input, and generate text, images, or video as output. But I believe that kind of model will be built very soon.

As for Google's Bard and Baidu's ERNIE Bot, these were also predictable. Google in particular had published several papers comparing their results against OpenAI's GPT-3. On some benchmark metrics, the two were roughly comparable — neither was absolutely ahead. What earlier research showed was: bigger models, more data, stronger capabilities. Some call this the Scaling Law. Whether it's truly a law? At least for now, it appears to hold across many problems.

I haven't personally tested Bard or ERNIE Bot, but from what I've gathered about others' experiments, the results align with expectations — the generation capability is very powerful. Without ChatGPT preceding it, Baidu would have stunned the world. Its strength lies in open-ended generation: if you're writing short stories where there's no right or wrong answer, nothing fact-bound, the model produces excellent content. But it has limitations, shared by OpenAI's ChatGPT, Microsoft's models, and Google's models alike — on factual questions, they still hallucinate. This is hard to avoid; anyone doing research knows this deeply.

The common problem with all these large models now: for complex problems that require humans multiple steps of deliberation to solve, models still struggle. GPT-4 cannot yet solve complex reasoning problems.

Oasis Capital: Have you encountered any interesting cases while using GPT-4 or other models?

Professor Lei Li: We researchers examine cases, but using cases to demonstrate a model's capability requires comprehensive evaluation. How to evaluate model capability is itself a science; it requires scientific methodology. That's from a research perspective — general user experience is a different matter.

We've been working on this recently. I've been collaborating with Professor William Wang at our university, and a PhD student we jointly advise, Wenda Xu, on how to scientifically evaluate text generation models — including machine translation, data-to-text generation, and image captioning — how to automatically assess generation model performance. The core idea is to train models that can automatically score generated text without relying on human evaluation.

Oasis Capital: What aspects of GPT-4's technical report have inspired your research?

Professor Lei Li: It's an experimental report. I think its performance across various tasks has improved significantly over ChatGPT, especially in multilingual capabilities. It also specifically mentions "predicting model" capability — using a simple Power Law to fit performance on smaller-scale models, then using that fitted prediction model to forecast performance at larger scales. GPT-4's training used this prediction to estimate required training steps, data volume, and so on. I think this is genuinely a good idea, and inspiring for our research.

Scaling Law had already been studied specifically before; OpenAI published an article on it a couple years ago. The law captures how model performance changes as model scale grows by orders of magnitude.

Oasis Capital: What implications or help have LLM breakthroughs brought to your current research area?

Professor Lei Li: Breakthrough research can't rely on just one approach. Increasing model scale and training data is one research method. We pursue different kinds of innovation. My research group has recently been focusing on speech translation and translation across hundreds of languages — problems with very scarce data, where simply scaling models and data quickly hits a dead end. Another crucial focus is trustworthy language models. I've been working with Professor Yuxiang Wang on language model privacy protection, intellectual property protection, and generation detection — we've developed some very useful and cutting-edge methods in these areas, which will be extremely important directions in coming years.

Another PhD student Professor Wang and I advise, Xuandong Zhao, works on Trustworthy NLP. What does trustworthy mean? Currently our core focus is on three problems:

First, privacy. Models like ChatGPT are trained on massive web data, which inevitably contains personal information that may have been released by mistake. If the model learns this information, its generated content may include personal addresses, phone numbers, ages, genders, ID numbers, and other private data. Our current work focuses on how to protect user privacy when training large language models — getting models to learn desensitization, to generate text without involving sensitive information, achieving the same privacy desensitization effect for generative language models and other models alike. Our core research introduces Differential Privacy analysis methods into models, defining a verifiable privacy-preserving training method under which we can prove that sensitive information requiring protection is effectively safeguarded. This research was published at last year's NAACL conference (one of the three top NLP conferences).

Second, intellectual property protection. I mean the model's own IP — OpenAI releases ChatGPT's API for public use, and by querying this API and matching outputs to original queries, one can use this data to train a model with similar capabilities. That is, you've "stolen" the original API's capabilities. If every company opens its API, it becomes easily colluded against. What we're doing is protecting the model itself — enabling it to open its API for normal use while also being able to detect if someone is stealing API data to retrain a similar model through knowledge distillation. Of course we can't guarantee the model won't be stolen, since API queries must function normally. But if someone distills knowledge from it to learn a copy, we can have evidence showing which model yours was imitated from. This was published at last year's EMNLP conference.

Third, copyright protection for model-generated content — not just copyright, but also the widespread use of ChatGPT for cheating. For example, submitting ChatGPT-written work as your own. Our core problem is whether we can develop methods to automatically determine if text was generated by ChatGPT or written by a human. The core idea is embedding some secret information that appears normal when reading the text directly, but can be verified through our detection method to confirm whether it was model-generated. This remains very difficult to solve; we're still working on it. If the generated text is very short, it's extremely hard to detect whether it was human-written or model-generated.

The first two research directions have published results available.

---

1. Provably Confidential Language Modelling https://aclanthology.org/2022.naacl-main.69/

2. Distillation-Resistant Watermarking for Model Protection in NLP https://aclanthology.org/2022.findings-emnlp.370/

3. Protecting Language Generation Models via Invisible Watermarking https://arxiv.org/abs/2302.03162

Oasis Capital: What if humans and machines write together? How do you control confidence intervals?

Professor Lei Li: This is extremely difficult. OpenAI's detection is another discussion — there are perhaps three or four groups worldwide working on this, including University of Maryland. Different methods have been proposed. OpenAI's own method isn't very accurate; they don't have strong incentives to do this. Strategically, it makes sense for others to do it — OpenAI can't be both API provider and referee.

Oasis Capital: Bill Gates recently wrote about teachers using models to grade assignments. What are LLMs' prospects in education?

Professor Lei Li: Automated model grading isn't new. Domestic companies were doing this before 2021, then stopped when they found it unprofitable.

I believe AI can play a significant role in education, but policy direction requires society-wide exploration and judgment. I've seen some schools strongly encouraging student AI use, but most remain cautious. I can mention two different strategies: this year for paper submissions, ICML and ACL took different approaches. ICML absolutely prohibits using models to generate any part of papers; ACL's policy (one of the three top NLP conferences) is: if you use ChatGPT like you use Grammarly — only to fix errors in your writing, like making awkward sentences fluent — that's completely acceptable. This reflects different expert opinions across fields. Since ACL is our top NLP conference, NLP researchers are relatively more tolerant of technology from our field, and happy to see it help people write.

Oasis Capital: Market variables have shifted greatly lately. Do you think underlying model capabilities will keep breaking through? Or might gains diminish after two or three years, hitting a bottleneck?

Professor Lei Li: Whether LLM capabilities keep improving depends on data and model size. For data, GPT-4 has already expanded from text to images; to improve further, text-side data is near its limit. Further improvement would need to incorporate other modalities — that's possible. Relying solely on text becomes difficult.

Stanford's Alpaca is dozens of times smaller than GPT-3. It uses a method from a paper called Self-Instruct, whose core principle closely resembles Geoffrey E. Hinton's article Dark Knowledge from about 10 years ago. He proposed a method called Knowledge Distillation — how to make a small model achieve large model performance. With instructions, prompts, and responses, perform generative "distillation." We use this method to create small models; many people are doing this with good results.

Image generated by Tiamat

Model size definitely has diminishing returns. From previous analytical research, models need to increase roughly 10x for linear improvement, so the scale required is enormous. Current models are already at hundreds of billions of parameters; multiplying by 10, by 100 — that size would be extremely difficult. So going forward, models may not necessarily keep getting bigger; other directions are possible.

Oasis Capital: Does this mean training domain-specific large models for each field is more viable than one universal solution?

Professor Lei Li: The conclusion is already clear — no further research needed to verify it. Domain-specific models perform better. But where do these domain-specific models come from? Not training each domain from scratch, but starting from a unified large model — say from ChatGPT, or from RoBERTa for understanding tasks — and developing on domain-specific data from these pre-trained foundations. From a research perspective, this is already well-known, nothing novel. But in terms of results, fine-tuning this way works better.

Oasis Capital: Do you think OpenAI's path can ultimately lead to AGI? How far away is broadly defined AGI?

Professor Lei Li: Not currently. The problems mentioned earlier — complex reasoning, fact-grounded generation — are still poorly handled. Beyond automatic text generation evaluation and trustworthy text generation, our group has recently been working on natural language-based complex reasoning. We believe these won't come from simply making models bigger or adding more data.

The AGI I envision would handle basic assistant-level work. There's hope now, but timing is hard to predict — more research funding for our lab would bring AGI sooner.

We've recently built two benchmarks: one for complex analogical reasoning, another for complex logical reasoning — like questions seen in civil service exams requiring quick thinking and strong logical reasoning. These are difficult for ChatGPT and LLMs.

Many problems OpenAI tests are essentially prediction problems. ChatGPT-like models perform well on prediction tasks, especially frequently occurring ones with similar statements found online. GPT-4 proved this — historical questions are answered particularly well. English questions generally work; but for another category of English questions, we haven't figured out why they perform poorly. Programming problems on Codeforces also perform poorly — competitive programming problems, information science olympiad-style questions. Models have some capability but not particularly strong, which has been verified. But where does GPT-4's capability come from? Some questions remain unanswered. For example, someone tested it on Codeforces problems from before 2021 — ten out of ten correct; problems from after 2021 — zero out of ten.

Oasis Capital: So it's still a memorization problem?

Professor Lei Li: There's no clear evidence it's memorization. OpenAI's own GPT-4 documentation states that tested portions were deduplicated, with all training set matches removed — the method appears reasonable and was implemented. But the results can't explain why pre-2021 problems reach 100% accuracy while post-2021 problems hit 0%.

Of course we hope it can genuinely solve some problems — generating a completely correct program for even 1 of 10 IOI competition problems would be extremely impressive.

Oasis Capital: What impact will LLMs have on industries?

Professor Lei Li: A few years ago, companies specializing in dialogue and customer service were very hot; now these companies need to rethink their future path. When elephants fight, the ants nearby get crushed.

Repetitive transactional work will gradually become automated and intelligent. Companies in this space either become leaders or need to consider transformation.

AI development has consistently been about automating simple, repetitive work — a trend that will continue for the next decade or even century. The bar for "simple" keeps rising. Phone calls once required operators, elevators once required attendants — now both are gone. I look forward to the day when some teaching work can be done by AI, like exercise class explanations — possibly achievable in the next year or two. Whether the business model proves sustainable is another consideration.

Oasis Capital: In the large model era, do the capable take all?

Professor Lei Li: It's positive for cloud computing giants; demand for large-scale model computation will increase. Open source versus closed source — each company decides based on its own positioning and interests. We hope technology doesn't become monopolized by a single player.

Oasis Capital: How can small companies seize the ChatGPT disruption opportunity?

Professor Lei Li: Move fast with small steps, use efficient and economical business models and products to solve domain-specific problems, build sustainable revenue generation, while accumulating industry experience and data moats.

Image generated by Tiamat

Oasis Capital: How should ordinary people embrace ChatGPT? Is there still a future in majoring in computer science?

Professor Lei Li: I'm teaching undergraduate deep learning this year; compared to last year, the class is so full many students are waitlisted. ChatGPT's emergence further pushes for making AI a required foundational course for all university students, on par with writing and mathematics as general education. Whether to major in computer science still depends on student interest and long-term career planning. From an employment perspective, computer science still offers tremendous opportunity — arguably the most promising undergraduate major currently.

Oasis Capital: What lessons does OpenAI's organizational design offer entrepreneurs?

Professor Lei Li: You must recruit the best people, and you must be willing to take risks (hire young people).

Celebrating Vitality

What do you think is technological vitality?

As a university professor, I deeply understand the importance of technological vitality — it requires constant emergence of new talent. That is my mission.

— Professor Lei Li, Assistant Professor at UC Santa Barbara

Oasis Capital is a new-generation Chinese venture capital firm dedicated to discovering the most vital entrepreneurs of the next decade in China, growing alongside them to create long-term value. "Celebrating Vitality" is Oasis's vision and mission. This Vitality represents both the direction of structural transformation in our era and the resilience and evolutionary power of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual investments ranging from $3 million to $30 million, concentrating on robotics, artificial intelligence, and technology services, supporting China's new service upgrade driven by technology.