What Do Two AI Titans Say About AGI, Open‑Source Models, and Real‑World Applications?

Technical Limitations of E‑commerce Conversational Assistants

Current large‑language models (LLMs) are strong at ingesting massive text and code corpora, but they struggle with deep, task‑specific reasoning required for end‑to‑end e‑commerce sales dialogues. Two fundamental bottlenecks prevent high‑quality assistants:

Model capability ceiling : LLMs can generate fluent language and perform basic intent detection, but they lack the planning and multi‑step logical inference needed for complex purchase decisions, price negotiations, or personalized recommendation sequences.

Domain‑specific data scarcity : Authentic sales‑dialogue logs are rarely available. Practitioners therefore resort to GPT‑4 prompt engineering or synthetic data generation, which yields prototypes that score roughly 60 % of the desired quality and hit a hard performance ceiling.

Because the assistant cannot reliably satisfy complex queries, users either bypass it for simple questions or abandon it when the request exceeds the model’s reasoning capacity, leading to low retention.

Scenarios Where LLM‑Based Assistants Are Viable

Viable applications share two properties:

They align with the base model’s strengths—primarily generation, memorisation, and shallow retrieval—without demanding deep planning or long‑term state management.

They have abundant, high‑quality vertical data that can be used to fine‑tune or prompt‑engineer the model, creating a virtuous data‑feedback loop.

Typical examples include:

Code generation and debugging assistance.

Document question‑answering (e.g., PDF or research paper summarisation).

Scientific literature summarisation and knowledge extraction.

Defining Model Ability Boundaries

Task difficulty can be evaluated along three orthogonal dimensions:

Similarity to the base model : Tasks that are close to the pre‑training distribution (e.g., code writing, summarisation) require minimal additional data.

Scene depth : Shallow, single‑turn interactions are easy; deep, multi‑turn dialogues that must handle long‑tail user intents (e.g., “plan a dinner for a boss with dietary restrictions”) demand extensive contextual understanding.

Logical complexity : Straightforward procedural logic (e.g., “list the steps to reset a password”) is manageable, whereas complex planning, constraint solving, or multi‑step reasoning pushes the model beyond its comfort zone.

Industry Timeline and Trends (2022‑2024)

2023 marked rapid progress:

Early‑year hype around LLMs, followed by a domestic catch‑up wave in China.

Mid‑year release of GPT‑4 and the emergence of Retrieval‑Augmented Generation (RAG), LangChain, and Agent frameworks.

By year‑end, Chinese base models approached GPT‑3.5 performance.

Application‑side milestones included the GPTs app store, the video‑generation model Sora, and growing interest in tool‑calling agents.

Despite these advances, consumer‑facing (C‑end) applications that require high‑level reasoning or strong persona—such as sophisticated e‑commerce assistants—have lagged behind the foundational model improvements.

Quadrant Analysis of Model Applicability

Mapping tasks on a two‑axis chart (personality strength vs. thinking depth) reveals that large‑model strengths lie in the lower‑right quadrant: weak personality, strong tool capability, low depth, fast thinking. Most successful B‑end and C‑end deployments (code assistants, document QA, image generation) fall into this region.

Practical Guidance for Building LLM‑Powered Assistants

Select tasks that match the model’s existing abilities : Prefer generation‑heavy, low‑logic tasks.

Secure abundant, high‑quality vertical data : Real interaction logs, high‑fidelity transcriptions, or carefully curated synthetic data. Data volume scales exponentially with task difficulty (e.g., ~1 k prompts for code generation, a few thousand for summarisation, tens of thousands for sales‑bot dialogues, and billions of tokens for full‑scale LLM replication).

Maintain data consistency : Use data that directly reflects the target interaction (real logs > transcribed audio > generated text).

Prioritise data quality : Expert annotation and rigorous cleaning outperform large‑scale noisy labeling.

Balancing model capability with data availability—choosing simple, well‑supported scenarios and iteratively enriching the data pool—creates a sustainable development loop and avoids over‑ambitious projects that exceed current reasoning limits.

Reference Interview Transcript (Supplementary)

我仔细检查了两篇文章，并对我的回答进行了补充和完善，以确保完整覆盖所有方面

杨植麟与朱啸虎对AGI的看法
1.对大模型创业的看法:
杨植麟: 坚定看好大模型创业，认为这是未来十年唯一有意义的事，并将其视为一个结合科学、工程和商业的系统，需要长期投入和坚持，最终目标是实现 AGI
朱啸虎: 不看好大模型创业，认为现阶段大模型公司缺乏场景和数据，估值过高，且面临开源模型的竞争，投资风险大。他更倾向于投资能快速商业化、变现的应用层项目。
2.对开源模型的看法
杨植麟: 认为开源模型落后于闭源模型，且差距会持续存在因为开源的开发方式和人才、资金聚集程度都无法与闭源模型
朱啸虎: 认为开源模型会逐渐赶上闭源模型，因为技术迭代曲线会放缓，且开源社区有更多开发者参与。他认为中国开发者更倾向于使用开源模型，因为不用担心被抄袭
3.对 AGI 的看法:
杨植麟: 坚定相信 AGI 是未来，并将其作为公司发展的长期目标。他认为 AGI 会改变世界，并希望通过技术突破和用户产品的结合来实现这一目标。
朱啸虎: 对 AGI 持怀疑态度，认为至少在5到10 年内还无法实现。他更关注短期内能商业化的应用，并认为在现阶段投入三资研发 AGI 风险过高。
4.对中美大模型产业的看法
杨植麟: 认为中美在基础通用能力上不会有太大差别，但在通用能力基础上的差异化应用更可能发生。他也相信中国公司未来有机会在某些方面取得领先。
朱啸虎: 认为中美在大模型领域差距很大，美国在底层技术上领先，而中国在应用场景和数据上更有优势。他建议中国创业公司先聚焦国内市场，再考虑出海。
5.对投资策略的看法:
杨植麟: 追求长期主义，愿意投入大量资金和资源进行前沿探索，并相信技术突破最终会带来商业回报
朱啸虎: 追求短期回报，更倾向于投资能快速商业化、变现的项目，并强调创业公司要控制成本、不要烧钱。
6.对 Sora 的看法:
杨植麟: 认为 Sora 是一个重要的里程碑，代表着视频生成技术的巨大进步，并认为它可以用来提升对多模态输入的理解能力，以及打通数字世界和物理世界
朱啸虎: 文章中没有明确表达对 Sora 的看法，但他对视频生成技术的前景表示乐观，并认为中国公司有机会在这一领域取得领先。
7对创业公司和巨头的关系的看法
杨植麟: 认为巨头和创业公司在目标和策略上有所不同，但两者之间存在着竞争和合作关系
朱啸虎: 认为大模型创业公司最终可能会被巨头收购，但收购价格不会很高，因为大模型技术同质化严重
8.对创业的风险和挑战的看法
杨植麟: 坦然接受创业的风险和挑战，并表示会无所畏惧地往前冲。
朱啸虎: 认为在当前的宏观环境下，创业公司要更加注重控制成本和自我造血能力，才能在市场竞争中存活下来
9.对未来的展望
杨植麟: 对未来充满乐观，相信 AGI 可以推动人类文明进入下一个阶段。
朱啸虎: 对未来持谨慎态度，认为技术发展存在不确定性，创业公司需要更加现实和务实。
10.对人才的看法
杨植麟: 重视人才密度，并认为公司上限是由人的上限决定然后 补充其他维度的人才。打造的。他一开始寻找天才个完整、有韧性、能打仗的团队。
朱啸虎: 认为大模型领域的人才同质化严重，创业公司需要找到有商业头脑和管理销售能力的人才，才能在市场竞争中胜氏。
11.对商业化的看法
杨植麟: 认为商业化是实现 AGI 的手段和目的，但不能为了追求短期商业利益而放弃长期目标。
朱啸虎: 认为商业化是检验大模型公司价值的唯一标准，并强调创业公司要找到能快速变现的场景和应用，