Author: lostingle Source: X, @losingle
Recently, seeing all sorts of "revolutionary releases" in the AI world, I always feel a sense of absurdity, like when I was in the Web3 era. Everything sounds amazing, everything is supposed to be revolutionary, but stripping away the high-sounding jargon, those with a background in technology often find themselves pondering: Is that all? Aren't these things we already did?
The current AI venture capital circle is undergoing a massive "term creation movement" and "technology repackaging," turning a bunch of down-to-earth old technologies into unattainable magic.
Stripping away the "revolutionary" cloak of AI
The touted "Web Agent" and "computer control." In the capital market, this is called "embodied intelligence" and "revolutionary productivity tools."
The "Web Agent" and "computer control" that have been hyped up to the skies. In the capital market, this is called "embodied intelligence" and "revolutionary productivity tools."
But digging deeper, the execution layer is full of outdated tools like Selenium and Playwright, tools that automation test engineers have been using for over a decade. This used to be called RPA or UI automation, requiring tedious XPath writing and DOM tree scraping; now, a large model is added as a "navigator," the underlying mechanisms haven't changed at all, yet it's packaged as an "autonomous intelligent agent." How can those who have worked in automation testing for years bear this? The wildly popular "RAG (Retrieval Augmentation)" and "Knowledge Graph" were once called "vertical search engines" plus "text summarization," now transformed into "adding an enterprise-level brain to a large model." Essentially, isn't it just cutting text into chunks, using a vector database to create a high-level "Ctrl+F" function, and then feeding it to a large model for summarization? As for knowledge graphs, the dimensionality reduction provided by large models makes those data mining veterans who spent countless nights writing SPARQL queries and doing entity alignment question their existence—the graphs they painstakingly built are less effective than simply feeding them to a large model for semantic retrieval. The "Real-time Voice Agent and Digital Human" that has amazed the internet. Capital is raving, saying the movie *Her* has become a reality, with AI companions possessing emotions and real-time responses. But breaking it down, 99% of these projects are still the outdated "three-stage pipeline": ASR (Audio-Speech to Text) -> LLM (Large Model-Generated Text) -> TTS (Text-to-Speech), connected by WebSockets. To mask the huge delays caused by this process, developers have written a bunch of rules to make the AI play a pre-set "Hmm..." or "Let me think..." before generating content to buy time. Engineers who previously worked on Tmall Genie and Siri exclaimed in shrewdness: Isn't this just replacing the old, idiotic dialogue tree with a large model? The "AI Data Analyst (Text-to-SQL)" that bosses are most likely to get excited about. Claiming to "automatically generate reports by chatting with the database," it boasts that it will eliminate all data analysts. The underlying logic is extremely simple and crude: copy the database table structure (DDL), string it together into plain language, send it to the large model, have it translate it into an SQL statement, and then run a SELECT statement in the database. What if there's an error? Send the error message back to the large model and fix it again. Isn't this just adding a "natural language translator" to the database? Packaged as "enterprise-level conversational BI," it dares to sell private deployments for hundreds of thousands. What about DBAs who have spent ten years writing complex JOIN queries and optimizing slow SQL every day? Another example is the proliferation of "Agent frameworks." Frameworks like OpenClaw and AutoGPT are advertised with claims of "autonomous planning and multi-agent collaboration." But a look at the GitHub source code reveals that 90% of these projects are just extremely thin "wrappers." The core logic is simply a while True loop filled with various string concatenation Prompt templates. The real dirty work, the reasoning and function calling capabilities, are all done by the underlying foundational models. The framework merely forwards APIs, yet claims all the credit for "intelligence." The most laughable thing is "Prompt Engineering." This title, which sounds like that of a top scientist, is essentially just "advanced string concatenation." Previously, programmers wrote regular expressions and If-Else statements to handle boundary conditions; now they're painstakingly coaxing the AI in text boxes: "You're an expert, please think step by step, or your salary will be docked." This isn't engineering; it's cybernetics parameter tuning. What used to be called "writing regular expressions" is now called "unstructured data parsing"; what used to be called "constructing SQL" is now called "data insight intelligence"; what used to be called "running scripts" is now called "workflow orchestration." Why is there such a strong Web3 déjà vu? Because large models lower the barrier to entry for interaction, allowing many people who previously lacked technical knowledge to suddenly see the power of automation and data mining. They think this is a brand new magic brought by AI, while veteran technologists just see it as old wine in new bottles. Don't miss that drop of "cyber glue"! However, despite the criticisms, we must admit that large models, this "new bottle," has indeed brewed a new flavor from the "old wine." Previously, automated testing, regular expression matching, and data mining were "rigid"—a simple change to a button's script on a webpage would cause it to crash. Large models, on the other hand, give these older technologies the "flexibility" and "fault tolerance" they lacked most. It's like an extremely excellent "cyber glue," gently connecting those previously rigid and easily broken IT modules. Current AI applications still heavily reuse infrastructure from the past decade. The big picture never invents new wheels; it simply applies a drop of glowing "cyber glue" to all the rusty old ones. Seeing through this facade allows us to remain clear-headed amidst the noise: peel back the natural language veneer of AI to see its underlying actuators. Don't be fooled by concepts, but don't miss out on this valuable glue either.
JinseFinance
Catherine