AI Buddy Coding: Why Your Best AI Coder Needs a Human Partner

“A little over a year ago, we copied and pasted AI code snippets.  Today, we have spec-driven development.  Yet the biggest lesson isn’t about the tools - it’s about how we use them.”

We’ve come remarkably far in twelve months.  From basic code completion to sophisticated integrated assistants.  From vibe coding to specification-driven development.  The technology has evolved fast, and it's continuing to improve.

But here’s what I’ve learned building systems with AI coding tools: the best results don’t come from better AI.  They come from treating AI as a coding partner, not a code generator.

Why AI Thinks Every Language is JavaScript

“Ask ChatGPT to write C code and it’ll give you Python with semicolons.”
AI coding tools have a dirty secret: they’re brilliant at some programming languages and barely competent at others. The difference isn’t random - it’s a direct result of what they were trained on.
And what they were trained on is overwhelmingly JavaScript and Python.

Tool Hopping: Why Switching AI Coders Destroys Your Project

I watched a team switch AI coding tools mid-project.  Three weeks later, their database architecture had quietly changed and half their API endpoints had lost authentication.

Switching AI coding assistants mid-project isn’t just inefficient.  It’s dangerous.

The Evolution of AI Coding: From Vibe Coding to Spec-Driven Development

First-generation AI coders let you chat your way to code. Second-generation tools are discovering that coffee-shop conversations don’t build production systems.

AI Coding: The Emperor’s New Code

Six months ago, everyone said AI would replace developers.  Today, most AI-generated code needs more debugging than writing from scratch.

The Context Window Swindle: Why Longer Isn't Better

AI vendors are racing to offer million-token context windows.  But... they're not doing it for your benefit.

Remember when 4K context windows felt limiting?  Then 8K arrived.  Then 32K.  Now we're being sold 128K, 200K and even million-token windows as the solution to all our AI problems.  Sadly, it's all a con.

Here's what they're not telling you.

How AI Conversations Actually Work

Every message you send doesn't just process your latest prompt.  The AI re-reads the entire conversation history - every single token - to generate each response.

Send a 100-word question in a fresh chat?  You're using roughly 100 tokens.  Send that same question after an hour of back-and-forth?  You might be burning through 50,000 tokens for the same answer.

The cost isn't linear - it's exponential.  Each message gets more expensive as your conversation grows, and those costs compound fast.  Think Cookie Monster but for tokens!

The Lazy Prompting Tax

Larger context windows enable sloppy habits. Why bother being precise when you can just dump everything into the conversation and let the AI "figure it out"?  Why manage your prompts when the context window is supposedly infinite?

This laziness has a price tag.  A substantial one.

A well-structured conversation with clear prompts might cost pennies.  That same interaction stretched across a rambling 50-message thread could cost pounds - delivering identical results while training you to be less effective.

The Platform Play

AI platform owners understand something most users don't: confusion is profitable.

Larger context windows sound like features.  They feel like generosity. In reality, they're revenue optimization strategies that shift costs to users who don't understand the pricing model.

The more context you use, the more you pay - but the pricing structure obscures this reality until your bill arrives.

What Actually Works

Effective AI use requires discipline:

• Start fresh conversations for new topics rather than extending existing threads
• Be precise with prompts instead of relying on accumulated context
• Extract and restate key information rather than referencing earlier conversation points
• Structure your approach before dumping content into AI tools

This isn't just about cost control.  Focused, well-managed conversations produce better results.  Shorter contexts mean the AI spends less computational effort tracking irrelevant history and more on your actual problem.

The Lanboss Perspective

At Lanboss AI, we help organisations implement AI without falling for vendor marketing.  Understanding how these tools actually work - and what they actually cost - is fundamental to safe, cost-effective AI adoption.

Larger context windows aren't inherently bad.  But treating them as an invitation to lazy prompting and unmanaged conversations is expensive and counterproductive.

TL;DR Million-token context windows are a feature designed to increase platform revenue, not improve your results.  Disciplined prompting and conversation management will save you money and deliver better outcomes.

GCS Leader Podcast Series

I have been most honoured to be featured in the GCS Leaders Podcast series.  David Bloxham and I have a fun and meandering conversation about the myths and realities of AI, where it is today and how to safely use AI to your advantage.

AI Action Figures, cute but not best use of AI

The recent viral fad to create cute, self styled action figure pictures of yourself or a favourite pet is not the best use of AI ever imagined.  Each image uses lots of GPU time and that means lots of energy.  An article on Techradar raises concerns about this massive misuse of AI fueled by this FOMO trend.  Even Sam Altman likes the idea, but he says it's melting the OpenAI GPUs.

AI Over Hyped?

Over the past month ro so we have seen a growing number of articles about AI market over hype and in particular, Microsoft which has pulled back from plans to build new data centres and also from letters of intent relating to purchase of 2 Gigawatts (GW) of energy, representing a significant scaling back.  The general feeling would appear to be that AI's energy consumption is already too high and delivering the promises will exasperate that further.

AI Literacy: A New Legal Requirement Under the EU AI Act – Are You Ready?

The EU AI Act is the world’s first comprehensive law governing AI, setting a precedent for how organizations across industries must handle AI responsibly.  Among its many provisions, a crucial but often overlooked requirement is AI literacy—a mandate that ensures employees working with AI have the necessary skills and understanding to assess AI risks and opportunities effectively.

From BPR to AI: Embracing a Pragmatic Journey of Transformation

In today's fast-paced business landscape, AI adoption is emerging as a transformative strategy that mirrors the revolutionary approaches of the past. Much like the 1990s Business Process Re-engineering (BPR) and the 2000s Kanban and Lean methodologies, integrating AI into your operations is about optimizing processes, enhancing decision-making and securing a competitive edge.  However, the journey to successful AI implementation is neither instantaneous nor magical.

Learning from the Past: A Legacy of Transformation

Elon Musk and OpenAI: Analyzing a Potential Reunion's Impact on AI

The AI industry found itself at another potential inflection point with Elon Musk's expressed interest in acquiring OpenAI, the company he co-founded in 2015 before departing in 2018.  This development is particularly noteworthy given the complex history between Musk and OpenAI, as well as the current state of the AI industry.

DeepSeek's Efficient AI Breakthrough

DeepSeek has made waves by delivering high-performing AI models with significantly lower resource requirements than industry giants like OpenAI and Anthropic.  Here's how they've done it:

Resource Efficiency at Its Best

• Innovative Architecture: Using a mixture-of-experts (MoE) and FP8 precision, DeepSeek optimizes hardware use, making training more efficient

Optimizing Your AI Prompts using AI

In today's fast-paced digital landscape, AI isn't just a tool for generating content - it can also help you fine-tune your prompts themselves for better results.  By using AI to analyze and refine your queries, you will recieve clearer, more creative and targeted responses.  If you're counting tokens, it will also save on usage costs.

So try asking your AI system of choice:

"Please optimise the following prompt and add placeholders for any additional information I might need to provide:  

---

Please write me a process to deploy nginx as a docker image to act as a reverse proxy for a web server

---"

Different AI systems will definitely yield varying results, so try posting this on a few.  This helps in so many ways to organise your thoughts and information gathering, its a real time saver, especially when you run the resultant prompt!  

We've almost stopped using search engines in favour of well written prompts.  Integrating AI into your prompt-writing process is a smart way to save even more time fault finding and prompt refining...

Happy optimizing!

ps: Please pardon my manners when writing prompts, you really don't need to do that, but do read our other blog posts about prompt manners!

What are the Security Risks of running DeepSeek R1 locally?

Tools like Ollama and llama.cpp run models locally, offering users unprecedented control and privacy over AI operations. However, this control comes with its own set of security considerations.  This blog post delves into the potential security threats and provides insights into how to manage these risks effectively, especially with questions over the provenence of models like DeepSeek R1.

Understanding DeepSeek R1 and Ollama

DeepSeek R1 is a cutting-edge AI model known for its reasoning capabilities, developed by DeepSeek, a Chinese AI research lab. It's designed for tasks ranging from data exploration to complex problem-solving.

Ollama, on the other hand, is a platform that allows users to run large language models like DeepSeek R1 on local machines, enhancing data privacy and reducing dependency on cloud services. While this setup offers numerous benefits, including lower latency and cost efficiency, it's not without risks.

Potential Security Threats

1. Data Poisoning:
   • Risk: Models like DeepSeek R1 can be manipulated through data poisoning, where malicious actors introduce corrupt data into the training set. This could lead to the model producing harmful or incorrect outputs inadvertently.
   • Mitigation: Regularly audit and update the training data. Use verification tools to check the integrity of the data before and after model updates.
2. Model Jailbreaking:
   • Risk: There have been reports of security vulnerabilities where DeepSeek R1 could be tricked into generating harmful content, like ransomware code or instructions for dangerous activities. This vulnerability can be more pronounced when running models locally without regular security patches or updates.
   • Mitigation: Stay informed about known vulnerabilities. Implement additional layers of security checks or filters on model outputs. Regularly update the model or use patches as they become available from the developers.
3. Privacy and Data Exposure:
   • Risk: While running DeepSeek R1 locally reduces data transmission to third parties, there's still a risk if the model or its outputs are not properly secured. Malware or unauthorized access to the local system could expose sensitive data processed by the model
   • Mitigation: Ensure robust local security measures like strong antivirus software, firewalls and access controls. Encrypt sensitive data before processing and be cautious about what data you feed into the model.
4. Virus or Malware from Model Execution:
   • Risk: Some posts on social platforms have raised concerns about AI models potentially carrying viruses or other forms of malware, although the likelihood is debated. The risk is more significant with models from less scrutinized sources
   • Mitigation: Download models only from reputable sources. Use sandbox environments for initial testing of new models or updates. Keep your system and all software up to date to protect against known vulnerabilities.
5. Backdoor Threats:
   • Risk: Given the model's origin, there's scepticism regarding potential backdoors, especially in contexts involving sensitive or strategic data processing
   • Mitigation: Use code auditing tools to check for any suspicious code or behavior. Consider using models from different origins or combining models to cross-check outputs.

Best Practices for Safe Local Execution

• Regular Updates: Keep both Ollama and DeepSeek R1 updated to leverage the latest security patches
• Secure Environment: Run the model within a secure, isolated environment or a virtual machine to limit system-wide risks
• Educate Yourself: Understand the model's capabilities and limitations to avoid misuse or misinterpretation of outputs
• Community and Forums: Engage with developer communities for insights on security practices or known issues with the model.

Conclusion

Running DeepSeek R1 locally on Ollama or similar can be a powerful way to harness AI capabilities with enhanced privacy and control. However, it is crucial to approach this with a security-first mindset. By understanding the potential threats and implementing robust security practices, you can minimize risks while maximizing the benefits of local AI model execution. Stay vigilant, keep informed and ensure your AI operations are as secure as they are innovative.