I don’t know… for me, learning the camera’s settings is half the fun. I mean, I like taking pictures as much as anyone, but the learning a new camera is a fun exercise in its own right. If I wanted to just point and shoot, I’d just set it to Auto and go about my business.
And I’m not talking out of my ass… I recently bought a Canon EOS R, which admittedly isn’t as modern as a Nikon Z8, but it’s a modern mirrorless camera nonetheless. And since I came from an old Canon T2i, it’s a lot of fun seeing how it works differently.
FYI… one cool thing with modern mirrorless cameras is that since they’ve got much shorter focal lengths (i.e. the distance between the back of the lens and the sensor), you can get a simple adapter that adds a few millimeters between the back of the old lens and the sensor to line it all up. So all those old-school lenses from the 60s through the 80s are now an option. I recently got a Konica Hexanon 50mm f1.7 for $10, and it’s a lot of fun. Especially considering the utility of modern mirrorless focus peaking type abilities.
If you use a single claude code instance for coding, planning, debugging, and analyzing results, it’s not as powerful. It’s better to split the cognitive load among parallel claude code instances for the same project - one is a coder (ONLY writes the code), one is a planner (writes specs, sprint maps, etc. for the coder), one is a debugger (looks at log outputs, failure modes, etc.) that writes reports for the planner or coder, one is an output analyzer (looks for correctness, does this match the output the planner expects, etc.), The specifics of what parallel instances you might need depend on the project. If you avoid cluttering the individual agent context windows with mixed domains (writing code vs debugging vs planning what to do next in the code or how to fix a bug you found), individually each instance is much more powerful. In this model, you are as much an orchestrator as anything else, just applying human judgement as needed across the set of agents.
I normally do start a different session or sometimes even switch to another model entirely when debugging. I still often run into instances where the LLM can’t seem to figure out what the problem is, and it instead presents a very detailed explanation that sounds very plausible but unfortunately does not fix the problem.
It’s fun when you have an actual printed manual. I used to keep manuals in the bathroom for reading while I was doing my business, and I would read that thing cover to cover. When it’s only available online, it’s not as convenient. Or I can download it and print all 74 pages myself. Then there’s the Z8 reference guide, which is a 1000+ page PDF or available online.
Actually I think that considering their amazing language proficiency, including the ability to read, understand, and summarize vast quantities of written material, ability to comprehend images, and ability to generate images and video and synthesize speech, LLMs are the most general-purpose AI model yet developed.
Very true. I’ve used ChatGPT and Claude to help me remember a book, film, or short story based on only a very tenuous description. In one case, I had been Googling in vain for quite some time and finally decided to ask ChatGPT. It identified what I was trying to remember in a mere second!
But this is just a special case of LLMs generally being a very powerful interactive information resource with which you can carry on a productive conversation and refine your knowledge, and explore concepts and ideas. One has to be cautious of the fact that it can get things wrong sometimes, but the capability is so powerful and groundbreaking and since I find that it’s right most of the time I find myself using it a great deal for just that purpose.
Well, the thing about that is: In the end, I get frustrated and look at the code myself, and I can normally figure out the problem and fix it in about 30 minutes.
I probably just need to realize after a second failure of the LLM at a debugging task I should cut my losses and look at the code myself. But in my experience a LLM is far from a panacea when coding.
I use the YouTube Gemini summary function to distill a 30 minute video into the short summary of the highlights, which is all I mostly need anyway. Then I can skip to the next video and not have to wait for the bullshit like number 10 reason will surprise you ruse.
FYI, this is worth a look. Where it differs from some silly parlour trick is that the transformation of the singer’s face into Simon Cowell’s face is absolutely perfect, and note that this is being done in real time as the singer performs.
Yeah this happens too, and sometimes the LLM makes it more difficult by patching over a real bug with a kludge overfitted to the specific error that papers over the root issue, making the root problem even harder to find until you look manually. Learning the nuances of how the coding tools work, make mistakes, hallucinate, etc. helps a lot - the problem here is that the tools are changing almost as fast as the learning process takes to use them to maximum effect.
Its two or three short paragraphs. It generally give a synopsis of the video, lists out the key points and then summarizes. It really beats watching a video that drags on for 20 minutes that should have been a 5 minute one.
I then can choose to watch the video or just go to the key point I’m interested in. Or not watch it at all.
I think this really depends on the context. Watching AI children get victimized is just as bad as real ones. It could give sick people ideas and normalize it to them.
What is the issue that AI actors would solve that tighter legal controls and actual punishment could accomplish? Overwork? Abuse off film by adults in power? These things happen to adults as well.
Oh man, the weird kludges are the most frustrating about coding through AI. The aforementioned silent 60 second timeout on a function was so irritating when I found it, I had to go outside and take a break. Nobody asked for it, it never mentioned it in its plan, and it certainly wasn’t a choice I would have made. I stopped auto-accepting its edits and reviewed the changes before I let it proceed after that.
That’s true, and they can be weirdly uneven on specific tasks. When feeding the code by one model into a different model when trying to debug it, the new one will sometimes offer up a new set of code for a completely unrelated function in addition to its proposed fix. I’m not sure what triggers that odd behavior. I seem to agree with the new LLM and accept the rewrite of the function about %50 of the time. A few times the new suggestion has been batshit crazy, though.
I won’t argue. And given that they are getting rather good at things like image recognition and creation, it’s not clear that LLM is even the best description? Under the covers it’s all pattern recognition and extrapolation, I suppose. Language or otherwise.
But is this the (or a) path to something we would call strong general AI? I really don’t know, how would we tell?
Unless you are religious, or a quantum sceptic like Penrose, it seems that mind arises from some sort of emergent process?
It turns out that a truly deep understanding of language – which LLMs clearly possess – necessarily begins to model basic human cognition across a surprisingly wide spectrum of capability, so that things like image processing and generation are not as much of a stretch as they may seem.
No. LLMs are just one particular breakthrough technology. True AGI will need to be a combination of many others, with LLM language proficiency being basically a very good front-end, or UI.
It’s all about emergence at sufficient scale. Because there’s nothing else there, otherwise one has to believe in magic, spirits, immortal souls, and fairies.
