Gemini has been doing this to me for the past few weeks at the end of basically every single response now, and it often seems to result in the subsequent responses getting off track and lower quality as all these extra tangets start polluting the context. Not to mention how distracting it is as it throws off the reply I was already halfway in the middle of composing by the time I read it.
I've definitely noticed Gemini's tendency to return the image basically unchanged, but not noticed it being worse or better for images of people. When I tested by having it change aspects of a photo of me, I found it was far more likely to cooperate when I'd specify, for instance, "change the hair from long to short" rather than "Make the hair short" (the latter routinely failed completely).
It also helped to specify which other parts should not be changed, otherwise it was rather unpredictable about whether it would randomly change other aspects.
There are a lot of caveats but in general the "spend 2 days" thing is a lot less true now IMHO thanks to LLM's that can write mostly correct elisp from basic specifications. YMMV of course. I have found this can also open up to being a lot more than "maybe 5% more efficient" for niche applications. It's the closest environment I've used to where the friction between "I wish my editor could do <x>" and actually having the feature almost disappears.
It is, though? I also had the same map come up about five times, showing a picture of the Parthenon and the Athenian acropolis, but it consistently insisted that these were of Orchomenos, which also had an acropolis but AFAIK is not the same; they are about 80 miles apart.
These were extremely helpful to read for insights on how to go back and retry different prompts instead, IMHO. I find it to be a significant step back in usability to lose those although I can understand the argument that they weren't directly useful on their own outside of that use case.
My thought is, for carrying a laptop, I'd rather have a backpack or messenger bag made of paper than a fabric grocery bag. The primary reasons I wouldn't carry a laptop in a grocery bag don't include the fact it's made of paper or the sounds of it rustling (though, it is nice to not have to worry about this one ripping).
When I learned the taskbar can't be placed on the side vertically is when I realized I'm never going to install Windows 11. That combined with 10 being advertised as "the last Windows I'll ever need to buy" and the fact their updater told me I have to replace my entire PC to be eligible to install it (in spite of the fact that I merely need to turn on a BIOS setting and my system would fully support it), then the notice that they'd be selling 1-year support contract extensions for anyone staying on 10... What's next, Windows declares itself outright as ransomware and holds my data hostage? I'm imagining a full screen overlay appears: "Nice OS install you've got here... Would be a shame if something... happened to it."
There most definitely is perceptual hue shift in HSL (and HSV); it's illustrated quite well at [0](The blog of the creator of OKLab), particularly in [1], an image showing S and L axes while leaving H fixed.
sRGB is already not perceptually uniform. These hue changes already happen. Just because the H value in HSL/HSV doesn't change doesn't mean there's no perceptual hue shift when adjusting lightness/value.
Using a trackball I couldn't get this to work noticeably better than just plain paintbrush input in MS Paint. Conversely the one from TFA took me a little more time to get the hang of but I got a lot better control with it.
On this machine it took me about 8 seconds to get the start menu open, about 5 seconds to get it to recognize that I'd typed "calc", another 5 seconds for it to let me actually select it to launch, and then about 20 seconds from the calculator window appearing - in its empty loading state - for it to actually come up. I admit this computer is several years old - but ... it's... a calculator.
This isn't quite the same as being able to edit an earlier post without discarding the subsequent ones, creating a context where the meaning of subsequent messages could be interpreted quite differently and leading to different responses later down the chain.
Ideally I'd like to be able to edit both my replies and the responses at any point like a linear document in managing an ongoing context.
On those forums, it's not just that there'd be no follow up; in my experience there'd typically be someone with a username marked as a "Community MVP" who would paste long walls of generic troubleshooting text for 2 or 3 replies insisting that the user go through a long list of steps that usually seem not to resolve anything (or indeed introduce additional variables and problems), often including completely irrelevant suggestions that seem to indicate the user's report was not actually read in its entirety to begin with.
In my experience, it's still possible to learn from the AI's mistakes. I've seen others comment similarly: sometimes the AI will invent a feature, variable, exception type, etc., and while this is wrong for the specific implementation, it's often a pretty good idea for something that may have been overlooked with the initial spec. Reading the AI outputs can still be a valuable step in an interactive feedback process, even if you only ever start a new context from scratch with the AI each time. The user's context is what grows.
Was in the middle of writing something like this comment already. Heh, definitely something a bit ironic seeing multiple comments claiming an editor written in Lisp and descended from an AI lab as being ai-free. It's probably one of the most user-customizable environments around for building user-centric AI tool integrations.
Also to be fair to Common Lisp, it has lots of support for sequences that aren't functional lists. Cons cells just happen to be convenient for representing things like the source code so they are prevalent. They're hardly a requirement for any algorithm's logical implementation.
> It would be helpful if I had a long rambling dialogue with a chat model and it distilled that.
IME this can work pretty well with Gemini in the web UI. If it misinterprets you at any stage you can edit your last comment until it gets on the same page, so to speak. Then once you're to a point in the conversation where you're satisfied it seems to "get it", you can drop in some more directly relevant context like example code if needed and ask for what you want.
When I asked, "Good afternoon! I'm trying to settle a bet with a friend (no money on the line, just a friendly "bet"!) Which MS DOS productivity program had a playable version of the game Connect Four built in as an easter egg?", it went into a very detailed explanation of how to get to the "Hall of Tortured Souls" easter egg in Excel 5.0, glossing over the fact that I said "MS DOS" and also conflating the easter eggs by telling me specifically that the "excelkfa" cheat code would open a secret door/bridge to the connect four game.
So, I retried with, "Good afternoon! I'm trying to settle a bet with a friend (no money on the line, just a friendly "bet"!) Which *MS DOS* [not Win95, i.e., Excel 5] productivity program had a playable version of the game Connect Four built in as an easter egg?". I got Lotus 1-2-3 once, Excel 4 twice, and Borland Quattro Pro three different times, all from that prompt.
The correct answer you point out in another subthread was never returned as a possibility, and the responses all definitely came across as confident. Definitely a fascinating example.
I notice this to affect its tendency to just make things up in other contexts, too. I asked it to take a look at "my" github, gave it a link, then asked it some questions; it started talking about completely different repos and projects I never heard of. When I simply said take a look at `this` github and gave it a link, its answers had a lot more fidelity to what was actually there (within limits of course - it's still far from perfect) [This was with Gemini Flash 2.5 on the web]. I have had simlar experiences asking it to do style transfer from an example of "my" style versus "this" style, etc. Presumably this has something to do with the idea that in training, every text that speaks in first person is in some sense seen as being from the same person.
Speculating, I would imagine that different prompts submitted along with the image might elicit wildly different behavior in how a multi modal VLM may respond to a given image, potentially affecting the relative tendency to upweight its effective inferences from prior training versus focusing more primarily on the new image itself.
Yet, one might justly question the imperative of cultivating a distinct model for such an endeavour, when a judiciously framed prompt, enriched by apposite examples, might suffice to imbue a sophisticated engine with the desired stylistic graces. Though it is undeniable these modern engines shall wax greatly in their proportions, and the art of discovering the exact prompt to elicit their most felicitous expressions is a task far from trivial, yet, it must be admitted, the pursuit holds a certain diversion for the inquisitive mind! It is, perchance, not the creation of manifold engines, but rather the artful disposition of singular contexts, that shall bestow upon diverse interlocutors their proper and unique voices.
Definitely a lot to be said for these ideas, even just that it helps to start a fresh chat and ask the same question in a better way a few times (using the quality of response to gauge what might be "better"). I have found if I do this a few times and Gemini strikes out, I've manually optimized the question by this point that I can drop it into Claude and get a good working solution. Conversely, having a discussion with the LLM about the potential solution, letting it hold on to the context as described in TFA, has in my experience caused the models to pretty universally end up stuck in a rut sooner or later and become counterproductive to work with. Not to mention that way eats up a ton of api usage allotment.