࿔
| QUICK THOUGHTS ON AGENTIC AI |
|---|
| Email response to a friend when asked about my thoughts on agentic AI |
| 2025-05-23 - in-progress |
My position is that agentic AI, the term, is similar to the previous use of "AI", where investors and CEOs know that it's an attention-grabbing term, without fully understanding where the technology is currently at and even a miscalculation on where it's going. I do think that agentic AI will be part of the future, and it is being primed to replace general office work. It also has the potential to replace all junior developers (which the tech companies are pushing hard for the latter). I can't tell you how often I've seen comments from people in the industry saying beginner to junior level programmers will be a thing of the past in the next few years, and by 2030 (ish), AI systems will wholly replace developers altogether. The overarching effort (moonshot, if you will) is AI models training other models in parallel at scale, and I think the agentic part is a big aspect of that. The consumer piece is easy to see, in my opinion. It looks like some agentic model with the ability to consistently complete small menial tasks that it is asked to do. Like completing emails, writing up research reports, small scripting or programming tasks. Then, up from there to managing servers up to scale, creating internal applications to satisfy internal business needs (cheaper than a COTS license for say, Oracle or whatever ).
If you've heard, MCP is all the rage right now, for good reason. The biggest problem developers were running into was dealing with the insurmountable number of protocols that different applications use across the industry. This would be incredibly difficult to develop for, and each company would have to satisfy each niche to make their agentic product work. Tasks like navigating a specific OS are minimal compared to navigating and using a plethora of applications with the added complexity of different versions of the OS depending on the machine. Sort of a nightmare. MCP acts almost identically to an LSP, where it standardizes communication between code editors and language servers for tools like autocompletion. Any editor can support any programming language from a single source. MCP acts this way, but for agentically trained models, where it is an open standard that tells the system how context is structured and how it needs to be fed into LLMs operating in a multi-tool environment. This is supposed to solve long prompting issues where, at a certain point, the context is lost on the model. It's supposed to be modular, so that agents can reason better on inputs. It's meant to help the models understand multi-step, multi-modal tasks.
Original paper: https://arxiv.org/abs/2503.23278
A high-level view: I think there will always be a human-in-the-loop, but it will consistently be changing over time until models get so good that they surpass all human ability. The only thing we have left at that point is research taste, but even then, I think parallel models at scale will be able to surpass us on that, too. I think it all comes down to hardware scalability. If hardware doesn't scale, models don't. Hence, the tech giants completely pivoted on energy, building their own power plants that run off of fossil fuels to power future data centers. I've gone too high level, but on the agentic piece, the industry is actively looking to replace the bottom 20% (maybe even up to 50%) of workers. Lean, AI-centric companies.
On scaling: https://gwern.net/scaling-hypothesis
| 𖤓 |