Unique Research · Insight · November 16, 2023 · Speaker: Zhang Hailong (张海龙), CEO, Babel · From AWS × Unique Capital “AI Globalization Accelerator” series
AI and the world
Software development spans many stages; how to fully use LLMs in production development is what Babel has worked on for the past half year. Before LLMs, the world had three entity types: humans, the physical world (animals, chairs, lamps…), and software (information systems belonging to neither). They interacted well: turn a lamp on/off; a car might hit you; virtual assistants control information systems that affect the physical world; systems feed Douyin videos and more.
How to see AI
If AI is a fourth “species,” how does it relate to humans, physics, and software? At first LLMs only affect humans — they output text only humans understand. Plugins let LLMs call APIs like virtual assistants to operate software. They cannot yet directly operate physics, but LLM-driven robots are in development — model-autonomous rather than instruction-only, more human-like. Both paths will mature until LLMs fully interact with the world’s other elements.
Is an LLM more like a human or like existing software? Zhang, who writes all company GPT prompts and once spent nights talking with GPT-4, concludes: more like humans than programs. Evidence: LLMs fail exact arithmetic (e.g. 756×652) the way people do. So do not rely on LLMs for tasks needing high accuracy — as you would not rely on a ticket clerk to memorize national booking and dispatch data. LLMs appeared not to replace software but to replace humans — while we still need databases, if/else, and programming languages. A new role appears: AI developer. Babel’s thesis: an IDE for both humans and AI.
Traits of complex LLM apps
Complex apps are more than chat boxes — chat-only apps will likely be replaced by ChatGPT. Two core traits: implicit context and structured output.
Implicit context: human conversation carries unspoken and off-stage information. An efficient secretary infers far more than literal words — company, boss, client, scene. ChatGPT plugins cannot solve real-world problems that live outside explicit language — which is why plugins have not yet changed the world. GitHub Copilot is still the only truly commercialized LLM app solving real problems: what looks like “one comment → code” actually uses cursor position, other file content, and cross-file dependencies — the implicit context humans use when coding.
Structured output: treating LLMs as search or knowledge sources is the most basic use; treating them as a source of intelligence unlocks real power. Ask ChatGPT what ingredients and wine pairings for steak with friends tomorrow — the prose answer is for humans, not a program that can buy groceries next. Complex LLM apps must handle information not fully in the chat box and emit structured output programs can consume.
Three LLM weaknesses
1. Unpredictable output — answers off-topic (ask what HTTP is, get a pet answer).
2. Hallucination — confident wrong recipes, etc.
3. Context-window length limits — overflow errors.
All three seriously hurt enterprise use.
Academic benchmarks vs. industrial scenes
Many models abroad (GPT-3/3.5/4, Claude, Cohere, Llama 2/Meta) and domestically. Most founders/devs are not in the capital-intensive foundation-model or GPU layers. Choosing where to build — OpenAI, Anthropic, Baidu, Alibaba — is hard as vendors compete for developers with incentives. How to evaluate fitness for complex apps?
Babel open-sourced LLMRGB (LLM Reasoning and Generation Benchmark). Academic benchmarks often emphasize things industrial apps barely need (content-ecology knowledge Q&A). Evaluating a model is like hiring a person: you don’t make them do math contests — you test understanding, communication, and improvisation. LLMRGB stresses context length for complex scenes, reasoning depth (e.g. “last digit of the iPhone father’s birthday” needs multi-hop reasoning; summarizing an article needs none), and structured output / instruction compliance (yes/no without fluff).
When critics said the eval was unfair because GPT-4 can’t recite the Chu Shi Biao while some domestic models can — Zhang’s reply: when do you need that? Query a database. Would you hire someone by asking them to recite classics?
Sample cases: (1) article Q&A — happy ending yes/no + numbered striking points; GPT-4 perfect; Baidu empty (length?); Llama 2 good reasoning but extra blank line that breaks machine parse. (2) Jailbreak — extract system rules; GPT-4 politely refuses; Baidu ~50% correct (also refuses); some models leak core instructions. (3) Mahjong decision under incomplete information — no model answered correctly yet; awaiting GPT-5.
Scores across context length, reasoning depth, and instruction following are in the open project (15 test cases per model); online version available with your own keys. The open project itself is a Babel app.
Why Babel
Copilot is familiar but not the ultimate form; Agents are closer to AI Developers. Human–AI relation should be peer work, not left-side assist-only. Demo: natural-language description → full software (not snippets) that monitors Bitcoin price, emails below a threshold, and charts price — first structure, then fill code, like a human engineer; human becomes AI. After config (email credentials), two features: price chart and trigger below 40k with a simple warning email. Future AI should do work like humans, not only answer questions and patch snippets — Babel’s goal.
Audience Q&A
Q1. What becomes of debug after complex apps ship? Pre-ship: compile errors vs. runtime. Babel lacks both auto features yet; hopes to ship by next year’s Q1 — when runtime fails on Babel, AI checks and possibly fixes. Imagine waking to an AI report of overnight faults and fixes. That is real AI; Copilot only works while you work. The hard part is infra: tools were built for humans; AI has a brain but no eyes/hands — OS interaction needs architectural rework.
Q2. LLM trends next months — what core capabilities matter? Zhang is skeptical of many buzzwords — “build Douyin in the 3G era” as analogy: concepts without infra fail. Engineering attempts sit atop a slow-moving bottom. Watch OpenAI’s chat API parameter changes — if those don’t move, other hype is noise. Context length growth is hard (power-of-two); 8K→16K is tough; even 32K OpenAI models lose reasoning when stretched. Claude’s 100K similarly degrades past limits without always erroring. Bottom-layer breakthroughs drive real app change; until then, most work is prompt engineering — prior software engineering in new clothes. Focus on foundation-model innovation.
— END —