⌨️ Yet ANOTHER "AI" blog entry / rant
I don't like all the conversation stating that AI is not yet capable of doing everything the hype says it would be. Still here's a follow-up on my first AI-rant:
- gambling machine on which you pull the lever again and again - "maybe this time" (https://gum1h0x.com/anhedonia-and-rant/)
- there is usually no undo functionality when the ai does sth stupid
- it would be nice to have an easy "go back one step" button in the conversation, deleting the newest answer completely from the context
- if you're going to work TOGETHER with the AI you have to be able to judge its output. What if you never properly learned "the correct way"?
- it seems that often people who don't know the technicalities behind LLMs assume them to have real goals, feelings, almost a conscience
- scrapers are making the internet sluggish - almost DDOSing servers - for example loading the same link/files over and over again
- the internet is becoming more and more closed - most conversations are happening in chat rooms that LLMs don't have access to. Or information is hidden behind pay walls. The open internet is full of ads and lower-quality content. And that is what the LLMs are feeding on mostly (?)
- without high level data LLMs will continue to decline in quality https://spectrum.ieee.org/ai-coding-degrades
- the technology in data centers will be outdated soon (?) - can they be retrofitted with new technology once that comes available?
- if we don't write code (or much less), how do we learn / stay up to date in our skills
- related to complex, production-ready applications: would you let AI build a car for you / if not, why do you let it write your apps
- the ones most hyped about LLM-Agents for creating code often don't seem to be proficient in coding / coders themselves
- much bigger code-review cost
- as a professional programmer you put your name under the PR the LLM produced
- you can't trust that every change is valid and thought-through
- it's still the responsibility of the developer to deliver working and secure code
- LLM capacity is already very limited with even pro-subscription services giving errors that they are working at full capacity
- will the future coder work be more about orchestrating agents and assigning them the correct skills for your project or will even that be automated at some point
- lots of effort goes into saving tokens - in one 2-hour session I used 12% of my monthly allowed premium requests
- https://fosstodon.org/@vie@hachyderm.io/116026351364785090
- "It's interesting because LLMs do a worse job than us, we lose ability/skill to do our job the more we use it, lose our jobs, produce worse software, are less satisfied with our work, etc."
- "Yet so many of my peers seem to be super excited about and advocate for it, while other working class groups at least detest LLMs if not even consider organising themselves to protect their trade/jobs from LLMs."
- in the mastodon thread there is another point being made about people in tech caring more about money than the craft - who are happy when AI does their job
- "It's interesting because LLMs do a worse job than us, we lose ability/skill to do our job the more we use it, lose our jobs, produce worse software, are less satisfied with our work, etc."