Here’s one person whose life was definitely not made better.
And now, something else that has been destroyed by the endless avalanche of AI bullshit — bug bounty programs.
This is more than an annoyance. This will have a genuine negative effect on the quality and security of our collective technology experience.
For me the appeal is that it answers questions about my problem. Do I want to spend hours with a magnifying glass sifting through hundreds of Stack Overflow posts about issues similar to mine, only to find a ton of “me too” posts and no real meat? Or do I want to explain my problem to a flaky assistant who knows everything and say “How do I fix this?”
The quality of the fix is its own validation: if my computer audio was out, and then the audio suddenly starts working after performing a few guided tasks, then it worked. If not, then I grumble at the agent and tell it to try harder.
The fact that it has flaky access to almost unlimited knowledge at moments notice is powerful.
I have a package on its way to me, probably somewhere in Hong Kong right now, with 5 printed circuit boards. This is the first time I have ever designed a printed circuit board, and there is no way I could have done this in a week or two without GenAI.
It explained the soup-to-nuts process, recommended a tool to use (KiCad), then walked me through schematic design, parts layout, running the traces, generating Gerber files, then choosing a fabricator and submitting the job.
Nowhere along the way did “AI” generate my design.
It was more a process of me saying “Ok. I found the footprint and 3d model for that USB-C connector from the Molex site. How do I load those into KiCad and get them to show up in the 3D viewer”, it responding, with usually good instructions, and me occasionally saying “There is no button labeled ‘select footprint’ on the menu. Stop imagining things. Think harder” and it replying “You are right to call me out on that. In the latest KiCad 9 they have moved the button to a different-named menu option…”
Meanwhile the room is filled with a light haze of profanity.
The thing is, with plenty of those back-and-forth discussions, I made my own circuit board!
What is the alternative? Reading blog posts and articles that are way too deep in the weeds on special issues, trying to consume 400 pages of documentation, trying to understand the big picture of the tool chain when everybody only talks about specific areas.
I think the second-best method would have been to watch YouTube videos on the subject, but they lack the ability to help me do specific tasks in my project.
Heck, even when things fail and I get some indecipherable error, I take a screenshot and paste in the screenshot and it instantly analyzes the error and, more often than not, explains which step I missed.
We can definitely agree on the pain involved in all of the bubbly overly helpful ramblings as GenAI tries its hardest to make up an answer when it doesn’t have one, but there is definitely an amazing baby in that bathwater.
We have a winner! ChatGPT is 99% better than Stack Overflow…it has never huffed at me that I forgot a reproducible example or cited a 2011 post I missed that might answer my question or told me I needed to understand the function better before asking about it. In short, ChatGPT doesn’t talk to me like I’m a fucking idiot, and almost always give me a viable answer.
The 1% (for now) is that given a reproducible complex example, occasionally you can get better debugging there, but tbh even that usually isn’t worth the effort.
Another thing AI can do is make a lot of ad revenue for YouTube and AI video creators. Despite that, YouTube is de-platforming some of the most active slop channels, which I applaud, though if there were a way to firewall that shit off my feed, I might care less.
Some deeply delusional people out there:
(People “in relationships” with ChatGPT.)
The AI slop is so damn annoying. I wish they could do more to stop it.
I bet they could develop an AI to identify and delete it!
AI has produced a groundbreaking paper explaining how calculus improves birth experiences in pregnant mathematicians:
Through the introduction of the Ovary-Function Theorem (OFT) and the application of the Cervix-Dilation Equation, the study reveals that explaining non-Euclidean spaces through pelvic retroversion significantly improves calculus test scores and reduces birth anxiety by 13.7%.
And it was published!
I am pretty comfortable saying that any parent who gets a device like this so their child can learn social interaction by talking to an inanimate object is harming their child on multiplie levels.
The key here is that this appears to be a crap vanity journal where you pay to have your stuff published, in this case, USD $2,949. Strangely, the article was published without receiving payment, presumably due to the same kind of incompetence that allowed it to be accepted.
A key takeaway is that such journals make it possible for crappy research to be readily published, inflating the CVs of researchers desperate for validation and university jobs, contaminating the scientific literature with dubious and worthless sludge.
The flood of crap is only partially stemmed by publicity and the work of online image and data sleuths.
AI needs meatbags…
ai can’t touch grass. you can. get paid when agents need someone in the real world.
My brain was slow, so I asked Google for the US states that have ‘e’ as the second letter:
This of course skips the “New” states, and most annoyingly Texas, which is the answer I needed (and figured out myself).
I thought “let’s specify word length of 5”, and it gave me:
Maine (M-e-i-n-e)
Texas (T-e-x-a-s)
It did helpfully point out that Idaho is five letters, but the second letter isn’t ‘e’.
I was trying to imagine what context anyone would ask that question. Crossword puzzle?
Sure, AI is going through some growing pains, but so are the people using it. Just like with smartphones, we’ll all take a few years to understand what works and what doesn’t. For example, if I ask it to list the starting quarterbacks of Super Bowl games, I’m likely to get a few backups, a few skipped years, and a few Pro Bowl starters. But if I ask it to give me recipes for Indian cuisine that uses very little cumin, it’s going to nail that one. The big difference is there are only a few starting quarterbacks but very, very many types of food. The broader the subject, the better it’ll do.
And as a human user, you have to know that LLMs drill down in their knowledge bases to find you answers, outsourcing some parts (like math or coding) to external tools as it sees fit. That’s why you often get responses like “I think it’s Canberra…no, that’s not correct.” It’s pulling up a word, then calling a tool to check the answer. So if you engineer your prompt to call that tool explicitly, you’ll get good results the first time.
It’s a learning curve. But even considering that, AI is still, most definitely wonderful at making life better.
LLMs don’t look up answers in a database. They generate responses based on patterns learned from exposure to trillions of tokens during training, refined through reinforcement learning. Also, LLMs sometimes simulate checking because they’ve learned the linguistic pattern of self‑correction, not because they’re performing any actual verification.
Recently I came across a discussion of what the author termed “Ptolemaic Code”, which he defined as “software that appears functional but is based on a fundamentally incorrect model of the problem domain.” Like the Ptolemaic theory of planetary movement, it seems to work– until it doesn’t. Basically all LLM’s are Ptolemaic Code.
Yes, a crossword - I meant to include that in the post, but my brain is getting increasingly scattered.
That’s not correct. LLMs like ChatGPT have pre-trained internal knowledge bases. They also integrate with APIs to do things like query search engines or ask a calculator to do things. If what you’re saying were true, you wouldn’t be able to ask it for the weather forecast tomorrow or the score of yesterday’s sports game.
Here.
…Curated datasets, synthetic Q&A pairs, and code repositories are introduced to round out factual recall and boost performance on specialized tasks, while filtering pipelines and red-teaming cycles seek to minimize exposure to low-quality, spammy, or policy-violating content.
The LLM doesn’t just incidentally learn that Paris is the capital of France from reading thousands of news articles. It’s told that explicitly in training.
I’m not saying it generates an answer by querying a database (I said base, not database). I’m saying it’s trained on explicitly curated datasets to memorize factual information in its weights, specifically in the MLPs.
Here.
A recent study from Google DeepMind in late 2023 looked exactly at this question: where are facts stored? They found evidence that a lot of factual associations (like athletes and their sports) are encoded in specific components of the network, especially the MLP layers. So, while attention handles on-the-fly context, the MLP layers provide the model’s built-in knowledge base, so to speak.