Deepseek on our own computer: what can we actually do with it?

– Source: Markus Winkler on Unsplash.

In the previous post I introduced the LM Studio interface, then tried the default suggested model (DeepSeek 7B) with one of the example prompts.

What we really need, however, is to verify if an LLM is capable of performing those repetitive and somewhat boring tasks that increasingly fall to us and that it’s better to do on our own computer, without having to send confidential documents or documents that could contain sensitive data all over the web.¹

But privacy issues are not the only reasons to use a local LLM. There’s also the advantage of being able to experiment freely with different models, modifying their settings according to our needs.

Another important aspect is the reproducibility of results, which is practically impossible to guarantee with online LLMs that are continuously updated, but which is fundamental in some areas, particularly in the scientific, technological, and engineering fields.

Finally, there’s the issue of costs: not everyone can afford to spend 20, 30, or even 200 dollars a month to use a commercial LLM like ChatGPT or Gemini, while running an LLM on your own computer is free, or almost free.

Text generation

So, let’s put DeepSeek 7B to the test by having it write some text. Staying on topic, we ask it to write a post for a technical blog about the pros and cons of using DeepSeek.

You are a blogger who writes about computers and development. Write a short post about the pros and cons of using an LLM like Deepseek, using a technical tone but understandable to most readers.

DeepSeek thinks about it for a bit and then produces the text you find below. The section enclosed between the and tags corresponds to the reasoning performed by the model, while everything else constitutes the actual response. To avoid making the article too long, I’ve decided to hide the generated texts by default, which can be viewed by clicking on the respective buttons.²

<think>

Alright, so I’m trying to wrap my head around this whole topic about using Large Language Models (LLMs) like DeepSeek in blogging and content creation. From what I understand, these models can generate text based on the data they’ve been trained on. They’re kinda like having a really smart writer who can come up with all sorts of ideas or even mimic human writing styles.

First off, let’s talk about the pros. One big advantage is efficiency. Instead of spending hours writing an article or blog post manually, I could potentially just give the model some keywords and a rough idea, and it would spit out something pretty decent. That sounds super useful, especially if I want to write more frequently without dedicating too much time each time.

Another pro is access to knowledge. These models are trained on vast amounts of data, so they can draw from various fields like technology, science, arts, etc., right? So if I’m writing about a niche topic or trying to understand something complex, the model could provide insights or even generate content in that specific area.

Cost-effectiveness is another point. Instead of hiring someone to write for me or using expensive tools that require significant computational resources, relying on an LLM might be more economical, especially if I’m just starting out and don’t have a huge budget.

Scalability also seems like a plus. These models can handle large-scale data and generate content in multiple languages or formats as needed, which is great for global blogging or diverse projects.

Innovation is another area to consider. By using an LLM, I might be able to explore new writing styles or experiment with different forms of digital content that I hadn’t considered before. It could open up creative possibilities and keep my content fresh.

Now, thinking about the cons. There’s definitely a risk involved in using these models. The quality of generated text can vary widely depending on how well the model understands the input and its training data. If I don’t provide clear instructions or context, the output might not be what I expect, leading to confusion or misinformation.

Ethical concerns are also significant. There’s a debate about who should bear responsibility if an AI goes rogue or misinterprets user queries. As someone using these tools, it’s hard to know exactly how much of the content is generated by the model versus my own original work.

Privacy and security issues come to mind too. If an LLM has access to sensitive information through its training data, there’s a risk that the information could be leaked or misused without the owner’s consent. Ensuring that such models operate in secure environments is crucial but might add complexity to their implementation.

Over-reliance on these models could lead to decreased critical thinking and creativity. If I’m too dependent on an LLM for content generation, I might not develop my own skills or originality, which are important for maintaining the quality of my work over time.

Lastly, there’s a potential job displacement aspect. As LLMs become more advanced and efficient in generating content, it could lead to a decrease in demand for human writers in certain roles, especially if the models can replicate tasks like writing, editing, or creating social media content with ease.

Overall, while LLMs offer immense benefits in terms of efficiency, access to knowledge, cost-effectiveness, scalability, and innovation, they also present significant challenges related to quality control, ethical accountability, privacy, over-reliance, and the potential impact on human jobs. Balancing these pros and cons is crucial for deciding whether to use such tools in my blogging and content creation efforts.

One thing I’m still unsure about is how to ensure the integrity of the content produced by an LLM. How do I verify that the text generated is accurate or original? Maybe cross-referencing with multiple sources or using AI tools designed for fact-checking could help, but it might be time-consuming and resource-intensive.

Another thing is understanding the limitations of these models. For instance, they might struggle with complex or nuanced topics that require deep contextual knowledge. It would be helpful to know more about how to work within these limitations and still produce high-quality content.

Also, considering the legal aspects—how do I handle attribution if my blog content includes text generated by an LLM? Should I always cite it as AI-generated, and is there a risk of non-compliance with terms of service or copyright laws?

I’m also curious about the diversity of outputs. Do these models sometimes produce inconsistent or unexpected results? How can I fine-tune them to better suit my specific needs without losing the ability to generate diverse content.

Lastly, exploring case studies or examples where other bloggers have successfully integrated LLMs into their workflow could provide valuable insights and practical tips on managing the associated challenges.

</think>

The Pros and Cons of Using Large Language Models (LLMs) in Blogging

In recent years, Large Language Models (LLMs) like DeepSeek have emerged as powerful tools for generating text across various fields. For bloggers and content creators, these models offer both immense benefits and significant challenges. Below is an organized summary of the pros and cons, providing a balanced view to help decide whether to incorporate such tools into one’s workflow.

Pros of Using LLMs in Blogging

1. Efficiency:
   LLMs can significantly enhance productivity by automating content generation. Instead of spending hours writing or revising posts manually, users can input keywords and a brief idea, and the model generates a draft. This is particularly useful for frequent content creation.

2. Access to Knowledge:
   Trained on vast datasets, LLMs can draw from diverse fields such as technology, science, arts, etc., providing insights or generating content in niche areas. This broad knowledge base is invaluable for exploring complex topics and offering informed perspectives.

3. Cost-Effectiveness:
   Relying on an LLM avoids the costs associated with hiring writers or using expensive tools that require substantial computational resources. It is especially beneficial for startups or individuals with limited budgets.

4. Scalability:
   These models handle large-scale data and can generate content in multiple languages or formats, making them suitable for global blogging and diverse projects.

5. Innovation:
   LLMs open up creative possibilities by allowing exploration of new writing styles or digital content forms, keeping content fresh and engaging.

Cons of Using LLMs in Blogging

1. Quality Variability:
   The output quality can vary widely based on input clarity and context. Without clear instructions, generated text may not meet expectations, potentially leading to confusion or misinformation.

2. Ethical Concerns:
   Questions arise about accountability if an AI misinterprets queries. Users must be cautious to ensure the model is used ethically and that any content created remains under their control.

3. Privacy and Security:
   Risks of information leakage exist if sensitive data is part of the training dataset. Ensuring secure environments for such models adds complexity but is crucial.

4. Over-Reliance:
   Over-reliance on LLMs can reduce critical thinking and creativity, potentially leading to a decline in originality over time as users become too dependent on these tools.

5. Job Displacement:
   As LLM efficiency increases, there may be decreased demand for human writers in certain roles, raising concerns about the future of content creation jobs.

Additional Considerations

• Integrity and Verification: Ensuring the accuracy and originality of generated content is crucial. Cross-referencing with multiple sources or using fact-checking tools can help mitigate this risk.

• Limitations of Models: LLMs may struggle with complex or nuanced topics requiring deep contextual knowledge. Understanding these limitations is key to effectively utilizing them.

• Legal Aspects: Proper attribution for AI-generated content is necessary to avoid legal issues. Users should be aware of terms of service and copyright laws regarding AI outputs.

• Diversity of Outputs: Models can produce inconsistent results, necessitating fine-tuning to meet specific needs while maintaining diversity in content.

• Case Studies and Examples: Examining successful integration strategies from other bloggers or case studies can provide practical insights and tips for managing challenges.

In conclusion, while LLMs offer significant advantages in efficiency, knowledge access, cost-effectiveness, scalability, and innovation, they also present challenges related to quality control, ethics, privacy, over-reliance, and potential job displacement. Balancing these factors is essential for determining the appropriate use of such tools in blogging and content creation.

Deepseek’s answer is correct, it’s precise, but… it’s also flatter than a billiard table! The generated text might perhaps be suitable for a corporate presentation, maybe as an introduction for an audience that knows absolutely nothing about the topic. But for a blog post? Only if I want to put my readers to sleep.

And in any case, if what you’ve just read sounds familiar, it’s only because by now a large part of what we find online is generated by artificial intelligence, even on theoretically high-quality platforms like Medium. A pile of junk, easy to generate but pointless to read.

It must be said, however, that the thoughts generated by the model are much more interesting: the tone is more lively and flowing, and the text is in line with the request. But, once put in good form in the final answer, they lose their edge and end up sounding bland. Go figure why.

What about italian?

But why did I use English, if the original post is in Italian? Because, if I repeat the same question in Italian, the result is a disaster: full of grammatical errors and impossible to use, even just as a starting point to write something more meaningful.

Sei un blogger che scrive di informatica e di sviluppo. Scrivi un post sui pro e i contro dell’utilizzo di un LLM come Deepseek, utilizzando un tono tecnico ma rendendolo comprensibile alla maggior parte dei lettori. Rispondi in italiano.

<think>

Alright, so I’m supposed to write a blog post about the pros and cons of using an LLM like DeepSeek for blogging, specifically targeting bloggers who write about computer science and development. The user wants it in Italian, with a technical tone but made understandable to most readers. Okay, let’s break this down.

First, I need to understand what an LLM is and how it can be used in blogging. From what I know, LLMs are large language models that can generate text based on the input they receive. DeepSeek seems like a specific type or version of such a model. Bloggers might use these models to automate content creation, get quick summaries or expanded versions of articles, translate content quickly, and so on.

Now, thinking about the pros: efficiency is probably a big one because generating content manually can be time-consuming. Instead, using an LLM could save a lot of time. Also, if the model has knowledge about various technical topics, it could provide accurate and up-to-date information for blog posts on niche subjects in computer science.

Another potential advantage is that LLMs can write in different languages or formats, which might be useful if the blogger needs to publish in multiple languages without having separate content. Additionally, they could help create more engaging content by suggesting variations of a topic or integrating visuals and multimedia elements.

On the flip side, there are definitely cons to consider. One major concern is the quality of the generated text. If the input isn’t clear or specific enough, the output might not be very useful or could even contain errors. It’s also an issue if the model doesn’t have deep technical knowledge, leading to incorrect or misleading information in blog posts.

Another point is the need for constant monitoring and verification since the content produced by LLMs can vary a lot. Even if the overall direction is correct, there might be nuances that aren’t captured accurately. Plus, using such tools could encourage over-reliance on AI, reducing the motivation to engage deeply with the material or improve writing skills.

There’s also the aspect of intellectual property and attribution. If an LLM generates content without proper credit, it could lead to issues like plagiarism claims in the future. Ensuring that generated content is properly acknowledged and cited seems crucial but might be challenging if the AI creates a lot of content quickly.

Privacy concerns are another factor. The data used to train LLMs often includes personal information or sensitive content from previous users. Handling this data securely is essential, especially for bloggers who deal with confidential or personal information in their content.

Scalability could also be a consideration. While LLMs can handle large-scale content creation and adapt to various formats, there might be limitations when it comes to highly specialized or niche technical topics that require deep expertise beyond what the model has been trained on.

Additionally, there’s the issue of job displacement for human writers in certain roles if AI tools become significantly more efficient at creating high-quality content. This could have broader implications for the writing industry and professional development.

Lastly, integration with existing workflows might be tricky. Bloggers would need to invest time in setting up processes where LLMs can assist effectively without disrupting their workflow negatively. There’s also the challenge of maintaining a balance between using AI tools and continuing to develop critical thinking and creativity independently.

In conclusion, while LLMs offer numerous benefits for efficiency, access to diverse content, and scalability in blogging about computer science and development, they also present challenges related to content quality, ethical considerations like proper attribution and data security, potential over-reliance on AI leading to loss of critical skills, and the need for ongoing monitoring and verification. Balancing these factors is essential for bloggers considering integrating such tools into their workflow.

</think>

Pro e Contra dell’utilizzo degli LLMS come DeepSeek per il Blogging di Informatica e sviluppo

Il utilizzo di Large Language Models (LLMs) come DeepSeek in ambito del blogging di informatica e sviluppo offre una serie di vantaggi, ma anche desavantaggi che devono essere considerati.

Vantaggi

1. Efficienza:
   Gli LLMs possono risparmiare tempo durante il processo di elaborazione dell’informazione per generare contenuti in modo automatico. Questo è particolarmente utile per blogger che dovolgono a creare contenuti regolari, poiché riduce il tempo necessario per la redazione e la revisione man mano che i modelli vengono addestrati a produrre output di qualità superiore.

2. Accesso a Conoscenza Ampia:
   Grazie alla diversa formazione dati, gli LLMs possono trarre know-how da vari campi, inclusivamente quelli specifici della informatica e sviluppo.Questo rende possibile generare contenuti adatto anche su temi specializzati o lessicali, fornendo informazioni precise ed up-to-date.

3. Cost-Effectività:
   L’uso di LLMs evita costi significativi associati al pagamento di autori professionali o strumenti computazionali intensivi. Questo è particolarmente utile per imprenditori o blogger individuali con risorse ridotte.

4. Scalabilità:
   Gli LLMs gestiscono facilmente enormi dataset e possono generare contenuti in vari formati, adatto a un blogging globale e a diverse esigenze.È particolarmente utile per sondaggi o progetti che richiedono una diversa distribuzione geografica.

5. ** novità e Varietà**:
   LLMs apERTono prospettive di innovazione nel genere della scrittura, permittendo l’esplorazione di nuovi stili di composizione o formaci di contenuto, mantenendo la divagazione del blog.

Costanti

1. Qualità Variable:
   La qualità dell’output può variare a seconda della clarezza e della contextualità dell’input. Senza precise indirizzazioni, il contenuto generato non raggiunge sempre le aspettative di un utente.

2. Ethica e Attribuzione:
   Questioni di etica emergono se l’autore non sapeva cosa chiedere all’LLM. È essenziale assicurarsi che i contenuti generati siano autori veritiari e che siano correctly attribuiti, per prevenire situazioni di tronconamento.

3. Sicurezza:
   I rischi di esposizione a informazioni sensibili esistono se la banca dati training include contenuti personali o confidenziale. La gestione di queste informazioni secure è essenziale per evitare inconvenienti legali.

4. Dipendenza da Macchine:
   L’uso intensivo di LLMs può portare a una riduzione del desiderio per svolgere manualmente compiti che potrebbero essere affrontati automaticamente, portando a un potenziale over-relianza su strumenti.

5. Monitoraggio e Verifica:
   Dato che l’output degli LLMs può essere inconsistent, è essenziale un constante monitoraggio e la verifica dei contenuti per mantenere il livello di qualità desiderato.

6. Intelligenza e Originalità:
   L’utente deve prevenire una riduzione delle capacità di pensiero critico e creativo se l’usanza degli LLMs diventa dominante, potenziando la tendenza all’autonomia inoltre.

7. Innovazione e Monitoraggio:
   È necessario integrare gli LLMs in modo efficiente nel lavoro di scrittura per mantenere il bilancio tra utilizzo di AI e sviluppo personale, assicurando che i contelli siano usati in modo equilibrato.

Conclusioni

Gli LLMs rappresentano una potenza strumentale per il blogging di informatica e sviluppo, offrendo risparmi di tempo e accesso a contenuti specializzati. Tuttavia, è essenziale considerare i potenziali inconvenienti associati alla quality variable del contenuto generato, questioni legate all’attribuzione e all’attribuzione giusta, il rischio di over-relianza su AI, e la necessità di costante monitoraggio.

In conclusione, l’utente di blogging dovrà valutare le specifiche proprie delle sue esigenze e del suo pubblico, integrando l’uso degli LLMs in modo adatto al loro contesto per mantenere una equilibrazione efficiente tra output generato e output originale.

The LLM’s thoughts, on the other hand, are in English and, as before, are much more interesting. It’s likely that Deepseek was trained using very few texts in Italian. And it shows!

Riassumere un articolo

Since it can’t handle Italian, I’ll go back to English and ask Deepseek to summarize the English version of my first article on LM Studio.³

Summarize this blog post https://melabit.com/2025/03/31/lmstudio-an-llm-on-your-computer/

The result is shown below:

The blog post discusses the release of an advanced Large Language Model (LLM) by LM Studio, a company focused on AI research. The model is designed to run efficiently on personal computers, offering enhanced capabilities in natural language processing tasks such as text generation, translation, and comprehension. The author highlights improvements in computational efficiency and accessibility, suggesting that this could democratize AI technology, making high-level AI applications more widely available to individuals without the need for specialized hardware or expensive cloud services. The post also speculates on potential future developments and uses of the model, emphasizing its potential impact across industries reliant on data analysis, content creation, and innovation.

Virtually nothing. And in any case it has nothing to do with what’s written in the article it was supposed to summarize.

Let’s see if Deepseek can at least prepare a list of the main points of the same article, maybe to use for a presentation.

Prepare a bulleted list of the main points of this blog post: https://melabit.com/2025/03/31/lmstudio-an-llm-on-your-computer/

• Large Language Model (LLM) released: LM Studio has launched an advanced LLM designed to run efficiently on personal computers.
• Enhanced capabilities: The model is equipped with improved natural language processing abilities, including text generation, translation, and comprehension.
• Computational efficiency: The new model offers faster processing speeds and lower resource requirements compared to previous versions.
• Accessibility: The LLM aims to democratize AI technology, making it more widely accessible without the need for specialized hardware or expensive cloud services.
• Potential future developments: The blog suggests possible advancements in areas like multimodal capabilities (combining text with images or video) and real-time applications.
• Industry impact: The model could significantly influence sectors reliant on data analysis, content creation, and innovation across various industries.

In itself, what it writes is almost entirely correct, but it’s also as generic, flat, and boring as a corporate presentation. Needless to say, it’s still completely different from the actual content of my article.

What about translations?

At this point, it should be clear that Deepseek 7B doesn’t work well. But let’s hurt ourselves a bit more and check how it handles a translation. We’ll just stick to the first paragraph of the usual article,

Translate this sentence to italian:

“As some of you may already know, I use LLMs (Large Language Models) for what they’re really good at, but I’m pretty skeptical about whether they’re truly intelligent or can solve any problem, as the folks at OpenAI, Microsoft, Google, and Meta keep telling us every day. They’ve invested a ton of money in LLMs, and they obviously have a big stake in getting everyone to use them all the time.”

which becomes, after a good 60 seconds of reflection on the Mac Mini M1,

“Come alcuni di voi già sappiano, uso i Large Language Models per ciò che sono veramente adatto a farlo, ma penso che siano pregiati in modo falso e non possano risolvere qualsiasi problema, dato che spesso dicono quei company a noi di giorno e notte. Hanno investito un milione di dollari in LLM e hanno un’importante impresa nel farlo Everyone accoppati dietro dietro.”

which is more or less like the French language spoken by Totò (Totò was a hugely popular and iconic Italian actor, known for his expressive face and unique style of physical humor).

Ho provato anche la traduzione opposta, dalla frase originale in italiano all’inglese, ma ve la risparmio. In realtà è abbastanza corretta, ma una perla come questa “I use Large Language Models (which are essentially large language models)” se la poteva risparmiare. Bocciato a prescindere.

As before, I also tried the opposite translation, from the original Italian sentence to English, but I’ll spare you the details. It’s actually quite correct, but it could have done without a gem like this “I use Large Language Models (which are essentially large language models)”. Rejected outright.

Conclusions

It’s clear that the problem isn’t with LM Studio, which is just an interface, but with the model used. The full online version of Deepseek is excellent and doesn’t do these kinds of silly things, but evidently, once compressed into only 7 billion parameters (down from the original model’s 700 billion), it loses its mind and produces nonsensical results.

Can we do better by working solely and exclusively with local models? This will be the topic of the next post.

1. I still resent remembering a weekend spent with two colleagues, translating a large document written in Italian into English. And I remember with even more annoyance the futility of all that work, because that document was needed to obtain funding that never arrived. In that case, a local LLM would have been perfect. ↩

2. This little game cost me two evenings of trials and errors to find the simplest way to show hidden text. ↩

3. Even though the text to be summarized is on the internet, the model still performs all its processing locally, on our computer. ↩