This is an attempt to create a ongoing thread that documents the latest advances in consumer AIs (LLMs or otherwise).
By consumer AI, I mean tools that are free or relatively affordable (< $100-ish a month, let’s say) and which are meant for the general public, not hidden behind enterprise contracts or confidential military secrets.
This is not meant to be a philosophical debate about the pros and cons of AI (we have plenty of those threads already), just it’s practical everyday uses.
In other words, what useful things have you discovered an AI could do for you lately?
They’re really good at computer programming, especially if you use coding “agents” like Claude Code, Google Antigravity, or OpenAI Codex (instead of just a chatbot)
They are pretty decent at making human-sounding speech and copying voices (e.g. Elevenlabs) or understanding speech (Whisper, Parakeet). Some of those tools are also OK, but not perfect, at dubbing videos to other languages while keeping the person’s voice and accent.
They are very good at summarizing scholarly or scientific sources and producing readable summaries, podcasts, Powerpoints, and video explanations (NotebookLM)
They’re OK but not great at making short videos (Veo, Sora, Kling)
I’ve been using a new budgeting tool that doesn’t yet have the ability to export a small subset of transactions as a .csv file or similar. I’ve found ChatGPT to be useful for turning a screenshot of the transactions into an excel file. I can also just upload it to the previous chat where I told it how to format the first one and tell it do the same again rather than explaining everything each time.
I had great success with AI as I went to small claims court and won a judgment against a contractor who had damaged my property. It walked me through the forms necessary to file, with suggestions on how to briefly submit my case to the judge, produce documentation, and specific steps to pursue in order to collect and garnish wages. I recovered my damages and court costs.
I’m going to go out on a limb here and go against the grain of some popular opinions expressed here on the Dope, and say that an AI like ChatGPT is genuinely useful as a general-purpose information resource. Yes, it can be wrong, yes, it can hallucinate and bullshit at times, but in my experience those events are sufficiently rare and ease of verification sufficiently easy that the net utility of ChatGPT as an information resource is very high.
I’ve had it interesting discussions with GPT ranging from cosmic issues about the universe to resolving household issues about appliances and cooking, all of which were very useful and informative.
I was impressed by recent interactions about my new (well, new to me) car. The owner’s manual is huge and not particularly well organized, and GPT was able to answer questions about how certain things worked which I know were correct because things worked as it said they did.
I was particularly impressed when I asked it about the many different adjustments and controls on the power driver’s seat, which the manual described only with a diagram with a bunch of arrows which weren’t particularly obvious. When I asked GPT, it responded much like a human would – “show me the diagram”. Since I also had a PDF of the owner’s manual, I was able to cut the diagram and upload it, whereupon GPT explained exactly what all the controls did, including the insight that one of them was a multi-directional control that I should conceptualize as a joystick rather than a lever.
This last bit may be counterintuitive, because while we all know that LLM’s are copious generators of the most ridiculous bullshit you’ve ever seen, it’s quite capable of spotting flaws, issues, and logical inconsistency in texts that it didn’t originate.
Which is more or less expected if you think about it… LLMs would be incredibly slow and wasteful if they were built to second-guess everything they said. Recently some LLM tools have been modified to expose a little of their internal deliberations by way of showing progress, and if you direct them to be skeptical, you can easily see them spiraling out into prolonged internal conflict and second-guessing. It’s got to be cut off in the name of efficiency, and the results often aren’t satisfactory.
However if you provide it with a piece of text and say “find the bullshit”, or “find the contractual loopholes” it will easily surface a list of things that should be checked more closely, and is usually correct about the flaws it finds (or close to it).
So while AI is a dangerous trap for the credulous, it’s a pretty powerful tool for the skeptically-minded.
That’s pretty cool. Was it able to mostly “one-shot” it, or did you have to have a prolonged back-and-forth? How much detail did you have to give it?
I had an interesting experience with this last night, trying to annotate a list of cars with their engine types. I pitted AI against AI in doing so, first starting with a coding agent (which wrote a Python script to look each car model up in a .CSV file I fed it), then feeding the output of that through a regular LLM which caught some parsing/English errors the coding agent did not, like some contradictions between the engine and notes columns.
Its ability to detect, even if just probabilistically, “this does not follow”-type situations is genuinely impressive and fascinating to me. I wish I could see into their latent space and better understand their reasoning… the math (and theory) of it all is above my intellect.
I’d love to hear whatever you wanted to say, if you do post it elsewhere.
Yeah, I’m still skeptical of this approach. Even before LLMs, image-based tree identification wasn’t very reliable, especially if you didn’t have closeups of the different parts.
Some trees look very distinct while others can be easily mistaken at a casual glance. The more science-y apps usually ask you to take individual photos of the trunk, branches, leaves, fruits, etc., and maybe count or describe certain parts of them to double-check.
I suspect the general-purpose image models don’t have sufficient specialized training for that and probably just end up correlating tree pictures with random social media posts, which is maybe good enough for some trees in casual conversation, but not a very good scientific approach to tree ID.
How did you feed it 600 pages of medical records? Was it able to keep all of that in its context window just from a regular file upload, or did you have to use specialized tooling?
I don’t fold a lot of proteins, but AI is said to be helpful to pharmaceutical companies for making accurate 3D models of complex things, similar molecules, and evaluating potential drugs for market.
Is it hype? We will see if new antibiotics are soon coming, to replace the ones with significant antibiotic resistance.
It still might have overwhelmed its context window. Whenever context gets long, I occasionally double-check by asking it specific questions about various parts of the context. And it no longer surprises me how often it gets things wrong, even when by strict word/token count, everything is supposedly still within its context window.
I don’t like that either (the forced AI)… all the companies are just shoving it down everyone’s throats. I much prefer to use them on my terms, only when I want to.
Yeah, I’m with Ponderoid here. It’s not clear whether it was actually able to read that entire PDF… you might want to ask it “what’s on page XXX (near the end)”, for example, along with a few pages in the middle, to make sure it actually read that far. (They usually don’t/can’t, but they don’t warn you about it beforehand.)
I try this often with SDMB threads, for examples, and it only reads the first handful of posts and never reaches the end of longer threads, despite being absolutely confident that it had reached the last post. (They don’t know what they don’t know, and they don’t have a good way to self-evaluate their level of confidence.)
I wonder how much that specific issue has to do with the on-demand scrolling feature of the Discourse html UI? Doubly so if the AI is scraping the human UI rather than requesting the thread from whatever API Discourse exposes.
We’ve gone around in Site Feedback a time or two about how to obtain the entire thread body in one go regardless of post count. The consensus is that at least within the browser UI there’s no way to do that.