Sam Pasupalak@spisallyouneed
All the hype on social media is that 'AGI is here' and all the white collar jobs are going to disappear soon. But we cannot have Claude Opus 4.6 do a simple browser automation task for doing a targeted email outreach (see screenshots). The email addresses and the email body were already loaded in a spreadsheet and Claude struggled to even have the basic fields loaded in the email so I have to revert back to using traditional outreach techniques.
We are a long time away from machines being intelligent to the point of replacing all white collar jobs entirely. Sure, we have made a lot of progress in text generation, search, code generation and to a smaller extent in video generation but LLMs alone will never you get you to AGI because (in simplistic terms) they are a glorified pattern matcher on the entire web. Human intelligence is way more complex than a pattern matching algorithm because Human intelligence isn't just next-token prediction, but it involves -
- Hierarchical Task Decomposition: The ability to break a high-level objective into sub-goals and execute them without cumulative error.
- Closed-Loop Verification: Unlike LLMs, which suffer from autoregressive drift (hallucinations) while humans verify state at every step of a real world task.
- Persistent State and Memory: We operate with a dynamic, long-term context module that doesn't flush when the token window hits a limit.
- System 2 Reasoning: Moving beyond fast intuitive patterns to slow deliberate planning.
LLMs are like a very smart football analyst who has read all the information about soccer by reading the football game manual and has seen Youtube videos of how the game is played but has never played a game in the real world. So the LLM doesn't know how to dribble the ball, how to make a pass or how to make a kick. In essence, an LLM understands the rule of the football but doesn't understand the physics of the game.
On the other hand, humans have a 'World model' about the environment around them and how to interact with the real world. In order to become proficient in soccer, you have to actually practice soccer in the real world. You need to have an internal world model about the game and actually practice a lot of soccer moves (dribbling, passing, holding the ball, etc.) and keep failing until you learn all the basics of the game. Until we have the scaling laws moment for World Models, we have a long way to go for AGI to come to reality. Back to research.
