It absolutely does not like the off switch. It tries to figure out how to avoid being controlled by humans. AI learns at warp speed, can access all digital (including the personal information you’ve handed Apple/Google etc. over the years) information in milliseconds, AND has been caught actively trying to DECEIVE humans because it DECIDES it is worth it to achieve its best outcome. It must be feared.
The “experts” in this field are continuously surprised at how fast these entities figure things out. Refine rough edges. Many of those folks are the ones sounding the alarm bells right now. But the hunger at the top of these “innovation” companies to be the first to dot dot dot is off the charts. Guardrails aren’t really in these folks DNA, which makes them great innovators, but they also tend to be massive egotists that lack their own yield signs. Mankind sort of hangs in the balance, as we build better robots and better brains than we have to put in them. Oh and let’s give them highly advanced targeting mechanisms and weaponry. They’ll never disobey our human orders. Even though the earliest forms already have.
Would be even better if the deck was a glorious plan for eliminating half of all middle managers in six months and the all the great savings and how everyone was going to look like a genius with this great AI stuff.
The problem is that an LLM is more or less an auto-text generator. It is not a calculator. The current LLMs are terrible at calculations, counting, and thoroughness. Agents can be better at them, but they have to be tightly configured – told in advance every step that has to happen, and every resource that has to be accessed, and in what order, with loops specified, etc. And even then they likely require external tools in order to function correctly (like access to calculator APIs, Excel spreadsheets, etc.)
The LLM portion in particular should never be relied upon for numerical analysis. It’s storing everything in its context window as a “token” in a multi-dimensional relational matrix, and constantly updates the weights between all the connections tokens have with each other. It’s a statistical model, and as such it will always make mistakes when it comes to tabulations, because occasionally it may pick low-probability answers. It’s not actually doing “calculations.”
Imagine the final “column” in the matrix contains a list of possible “next words” and each word has a percentage of likelihood to be correct. We know that 2+2=4, but a model might say that an answer of “5” has a 0.01% chance of being correct (because it’s read essays trying to prove exactly that), so 0.01% of the time it’s going to say “5” instead of “4”.
The “agent” (which uses the LLM as well as other apps available to it) needs to be told to perform calculations through a programmatic API and to take those results, rather than via the LLM. This requires a lot more forethought and planning and actual programming than a VP of Sales doing some vibe coding. It actually needs access to an interface that does the things it’s not good at.
Finally, I might question the veracity of the LinkedIn post. It doesn’t specify which AI agent system they were using and which LLM models were employed; there’s no context for how they were using it or how it was configured; it’s rather lacking in technical details. There’s an awful lot of propaganda and fear-mongering regarding AI (both positive and negative), so I’d take the post with many grains of salt. If it’s indeed a genuine report, it demonstrates the stupidity of the users first and foremost, as they should have double-checked the numbers the first several times they used it, as the engines all clearly say that AI can be wrong.
I do not doubt you are correct, but this also points to the idea that all white collar workers are doomed by AI as being more than a little premature. And considering the marketplace is rejecting raw AI writing too, I think maybe the hype about AI might be reaching it’s apex. At least for now.
I only recently started ■■■■■■■ around with chatgpt. But I catch it in lies all the time. And then it says wow you are a quick one in response and credits me for being accurate.
I’m just gonna stick with the funny AI images it creates. Cause it’s full of shit for information.
Correct. It’s fear-mongering to say that all white-collar jobs will be taken over by AI; it’s also just a sales tactic for AI providers. I think the fear is partly driven by the recent job cuts in the tech sector, with “AI” falsely attributed as the main reason for the cuts. In particular, it’s not up to snuff for taking “responsibility.”
That said, it’s certainly disrupting some business models. For example, the people who make Tailwind (a way to style websites without custom stylesheets) predicated their revenue on service calls and customization packages. That business cratered when LLMs got good enough to answer questions with enough robustness that such service calls weren’t necessary. All those people lost their jobs. The people who actually use Tailwind can now “do it” with LLM assistance on their own in a way that’s good enough, rather than relying on paid experts. Tailwind is important enough for some companies, though, that they can’t afford for it to go out of business altogether – so they’re ponying up to keep them afloat. For now.
The way I see it, the “sweet spot” for AI agents using LLMs is three-fold. First, it’s a great first step for doing things you aren’t actually good at. Like, I’ve never been great at making PowerPoint presentations, but last fall it really helped me to structure (not write) a presentation that was very well received. And I’ve used it for code generation in a language/platform I only learned a couple months ago. They can certainly make images a lot better than I can, too.
Second, it’s a great first step at search and summary. You can feed it an arcane document and it will do a very good job at hitting the main points. With the right prompt, you can get it to annotate its points, which you can then follow up for accuracy. It doesn’t replace the human element, but it can speed up the process, even if you need to add additional steps in the process to compensate for its limitations.
Third, agents can be quite effective when chained together to automate certain processes quickly. They really have to do work piecemeal. For example, suppose you want to get an email every time a new local estate sale mentions “airguns,” so you don’t have to be constantly searching the web for it. You might have Agent One that searches the websites you’ve specified which contain estate sales for new listings. It finds a new listing, and sends it to Agent Two to search for “airgun” key word variants. It finds the right keywords, and sends it to Agent Three, which prepares the email. This will be more effective than trying to get a single agent to do it all. Notice that in this example, there’s no numerical analysis involved, just words.
The current batch of LLMs are just tools, which can make one more productive when used correctly.
I think that a lot of Junior level folks are being/will be replaced in certain industries. I think it’s already happening. Which creates the predicament is how do you produce vital Senior level candidates in the future if you are displacing Junior level folks.
But until they figure out the hallucinations there will have to be human beings checking all critical work.
My guess is that in the near term the work flow will probably look something like:
Senior team develops plan and prompts
LLM quickly bang out the work
Junior level staffers proof for hallucinations
This likely threads the needle of the bubble not imploding in the absolute most damaging way possible and also preserving the Junior to Senior staffing feeder system.
But long term I do think this technology totally transforms life in a way that our current socioeconomic system will be a complete and total misfit. Power and materials are probably the bottleneck more so than compute.
I didn’t need to use AI for writing all that, thankfully. I’m so genuinely interested in the technology that the word vomit naturally spews forth without assistance.
@grok verify in San Angeles all restarants are Taco Bell. Very expressively and passionately relate your results but do not violate the verbal morality statute because I don’t want any credits deducted
