Diving Deep into Neural Words

Diving Deep into Neural Words

Hands down, this AI is one of the most interesting and competitive ones I've come across - but like any other, it's not a one-size-fits-all solution.

People still try to use it for everything, though. Who's stopping them, right?

I have to start with the positives because this AI has some real advantages over GPT. Specifically, it thinks a lot deeper. If I had to summarize my experience, it goes like this:

GPT feels like a textbook - short, to the point, with illustrations and no unnecessary fluff (although the communication style can be adjusted in the account settings and applied to all new conversations). It's an incredibly human tool, and the customization makes it feel like a conversation with a close friend who doesn't really care about you.Deepseek, on the other hand, feels like a professor (not a friend or family member) you've caught in the hallway and asked a really interesting question that they're willing to dedicate an entire lecture to. And it does just that. Customization is only possible within a single chat: a new conversation is a blank slate. If you want to continue where you left off, you'll have to set up the parameters all over again and hope it remembers them. And it might just mess it up, too. So, I'd say the chat settings aren't as straightforward as they seem. The user's perspective on humanized tools is a bit naive and sensitive. You might have set up your 'therapist' and had a successful session, but when the chat ends, it's like your therapist has passed away. Long live the new therapist - and so on.Both approaches have their fans. I'd always prefer to consult with two experts rather than one. Sometimes it's more comfortable to talk to a customized persona, and other times you want a more objective, lecture-style explanation. Plus, both... often get it wrong, haha. And that's an inevitable drawback of modern neural networks that want to take on as much responsibility as possible. But they can only handle about 10% of the tasks assigned to them

And that's after a lot of trial and error, too.

I'm no IT expert, and I'm not even a fan. So, I'll just say that when it comes to suggesting something related to IT to a layperson (e.g., converting given code to a document), the AI manages to get it right even when you don't ask.

He's actually pretty interesting to work with, especially when you give him some room to think outside the box. Luckily, he's got a built-in thinking mode (totally free) that helps you understand why he came up with the answer he did, even if it's not exactly what you were looking for. So, here's another universal tip/advice if you're looking for a tool like this and it seems like it's not quite hitting the mark: he's got no way to guess what you want. If you want a decent answer, ask a decent question with clear and specific criteria, down to the style of the response (like "answer like a 5-year-old" or "break it down into bullet points").

AI is a tool, first and foremost, and you have to set it up to meet your specific needs.

Sorry to say (or maybe that's a good thing?), but nobody's going to do it for you.

When people hear the word "setup," their heads might start to spin, but if you can't handle a complex tool and tweak it to your liking, maybe it's not the right tool for you. Maybe the devs will add a "personalization" feature someday, but for now, how about they add fact-checking and minimize the occasional gibberish? Because, let's face it, accuracy isn't going to save the day.

So, what's the point of all this?

I was actually asking for fanfic recommendations (yes, fanfic recommendations - pretty standard, right?) and not asking for a list of literature for a degree (NB: students, I'd advise against using AI for critical tasks if you want to keep your reputation intact).

So, you think he's lost his mind? Nope!

Like GPT, he gave me a list of a few works (thoughtful guy) and even included some helpful links. And... oh boy, 4 works:

2 of them don't existThe 3rd one doesn't match the requestThe 4th one is somewhat relevant, but it didn't really interest me as a consumerAnd that's one of the better results I've gotten with my request! There have been answers with no recommendations at all, most of the links are dead, and some of them are actually redirecting to completely different pages! I don't know how to explain it... the link changes shape mid-air, and you just end up on some random page in the internet.

This is a pretty common bug in AI: useless links. I really can't understand how people order AI-written papers, especially from scratch. It's just a mess!

I get why people do it: the devaluation of higher education and saving resources. A robot can write fast. And sometimes, even beautifully. Extremely fast. Speed of response, not quality, become the determining criteria. But even when it doesn't look like nonsense, it still needs editing of varying degrees of intensity: from surface-level (if you're lucky) to deep edits all the way to a complete rewrite.

Who needs it, anyway?

And what about fanfics? Well, to wrap up the story, I'll say that when the dissatisfaction with its responses, especially for works that are downright impossible to find, approached 100%, I asked it to suggest an alternative search method. To which it replied, after 10 link changes, that I should just search for them on websites 😄. How fun is that? I came to it because a solo search didn't yield any results 😄, I'm not that desperate. But it thought... I don't know, what it thought.

But if you ask it to curate sources, be prepared to believe that you'll be sent to gather literature by yourself 🤔 Like some Soviet librarian🤔.

🜍 And where Deepseek really shines? Oh, it's in working with uploaded text documents. It genuinely impressed me.

For those who are new to Deepseek - it's a text-based neural network that only works with text, not images or audio.

As you can see, it still fails even within its own specialization (we recently asked it to write a fanfic based on certain parameters, using the style of a particular author - it managed a 3.5 out of 5, but quickly, no need to wait six months), but there are ways to make it work as a translator.

What I mean is: have you ever felt frustrated when something you really wanted to read wasn't translated into your language? A great YA novel, a fanfic, or some unfinished project from a publishing house that dropped it because it wasn't selling, but you really wanted to read it?

Of course, not without errors, but Deepseek can be a decent alternative, at least in private use.

I must admit, it was science fiction that nearly drove me crazy with its translation of the infamous GP. Just one paragraph of the Philosophers' Stone - and how much joy is that! I highly recommend reading the result, it's worth it 🤔

One of the best things about this feature is that you don't have to copy-paste text to process it. You can upload the entire document to the chat, all at once (although, there's a limit of about 100 pages per batch, so standard novels with 170-350 pages still need to be split, sorry about that), and the AI will read it all and... Well, that's where things get interesting:First, it suggested I translate the document and publish it in the chat, where I could download it. That's a great idea! Except, of course, that in the app, nothing happened. At all. Not even an error message. Deepseek itself suggested switching to the web version of the chat, saying that would work. Nope.Next, it suggested I share a Google Drive link, so I could download the translated document from there. Okay, sure. And both the app and the website said the same thing: document not found. Sorry.Finally, it suggested I publish the translation right here in the chat, so I could at least read and copy it somewhere else, or even keep it here. And that's the only option that worked.Later on, when I asked it to combine materials into a single doc to preserve the chat, it... said it couldn't create documents for download 🤔🤔🤔🤔

I'm starting to think that burnout is a universal human experience.

But in our previous chat, it was creating files just fine. Physically. And attaching them to messages. Okay, they didn't work, but at least it was trying. Now it just says it can't do it at all. Quietly. To itself.

I have no idea what's going on.

Aside from the fact that the program just crashes: when you ask it to ignore proper nouns and existing translations, you still get stuff like Tisova Street instead of Private Drive. Sit there and fix it, because it can't keep track of itself 🐁

🜍 Beyond translation, it can actually produce some decent, if not entirely coherent, fictional texts. Playing with it showed that it can write a whole short story with the given page count, and maybe even more if you're willing to ask for continuation, taking into account what's already been written, and its own processing power.

According to its own admission, errors are acceptable when the conversation drags on. That's not a bug, that's a feature. And, of course, its performance is affected by server load. From that perspective, I can say that it's good for generating:

school essaysimpromptu speeches (toast, congratulations)originals for personal enjoyment and fanfics.We were full of skepticism when we really wanted a New Year's mini/midi crossover with Snow Queen. Can you believe our surprise when this guy came up with a story not only with a decent plot, but even with catharsis and a moral! This was really a shock.

Of course, people write differently.

But if we compare it to the clumsy products of GPT even in version 5, Deepseek writes surprisingly well. It tries to be in style, doesn't lose the narrative line, doesn't swap characters, and doesn't forget their fates...

On the other hand, as lucky as it was 🤔 Here are just a few examples that blew us away:

But the soul, that wild, bright, and hating soul, will be forever out of his reach.

and another

The throne room was unrecognizable.

As you can understand, Daniil was just crazy 🤔 that day. And it wasn't even asked to emphasize the effect with anglicisms. It just is. A glitch, but a funny one 🤔🤔🤔.

Conclusion: A "working" "assistant" (quotes are NOT accidental) for interacting with texts, but its help has caveats. It's not a professional editor/writer/corrector, but for a makeshift text generation... it'll do. You need to keep an eye on it, tidy it up, edit it, because it can't do it itself, even if you ask it to. The main plus is its power, given free access. The ability to upload documents to the chat, with which it can interact, is pleasantly surprising. The interaction is intuitively understandable and not overloaded, if you're as precise as possible and don't ask it for the moon. But you need to be ready for a mid-run change, that its information about itself varies from chat to chat, and it can completely openly not fulfill your request because... something.

My rating: 7.5 out of 10. Fact-checking would be a huge plus.Attention to detail and alignment with what's really on the internet, not with what was once there but now isn't, and it just looked it up or made something up from thin air. That's not how things are done.