Mihai Ionescu’s Post

Sometimes I feel like the cheems of using LLMs. I see all this breathless hype from buff doge folks about how they're using LLMs to do their jobs in no time and they're uber productive and I get mad FOMO. Here's how I honestly use them daily: - Code autopilot: I use Supermaven for this and as a rusty old dog programmer I find it pretty useful. I use it basically all the time I'm coding. I don't find it gets in the way too much, and it's generally on point with 1-5 line suggestions. - Code generation: I use this once every couple of days, and I'm generally disappointed 😆 Maybe I'm just a prompt noob, but if I try to create a feature in an existing code base it'll generally hallucinate a few APIs and just not quite get it right. It works really well for isolated things that don't need to play well with others though. Yesterday I used it to whip up a simple HTML page to preview a bunch of test cases outputs I had in a file. Super good for those kind of quick throwaway things. - Code inspection: A couple of times a day I'll just ask an LLM "where's the bug in this code" and it'll come back with something useful. Maybe I'm just lazy, but I'm finding it a good first step. - Random brainstorming / question answering with chat interfaces: Once or twice a day I'll throw some amorphous problem into ChatGPT or Claude to try to get an answer. Sometimes it's useful, sometimes it's pretty generic. So I'm using LLMs multiple times a day, but not for absolutely everything. I feel like I'm maybe getting a 20% boost on productivity from them. Pretty epic still, but not the hype that some people are spinning. How often do you honestly use LLMs? Are you buff doge or cheems?

I've been coding daily with AI since early last year and recently made a notable change: GitHub Copilot to Sourcegraph Cody. Some time ago, I noticed a big change in Copilot Chat. It became way faster, but the quality of interaction went way down. I presume this was both a change of model (GPT 4 to 4o) and prompt. Suddenly, all it wanted to do was spit out code. It had no desire to explain things unless you explicitly said "no code, just explain". This was clearly a system prompt change. Moving to Cody, I can actually get conversations again. It will explain everything and acts a lot like ChatGPT 4. Autocomplete code quality is about the same, but code in chat is definitely better. And this is the free version using Claude 3.5 Sonnet. There are a few things I miss from Copilot: - Cody fonts are too small and you can't change them. - no shortcut to enable/disable autocomplete (or just disable in comments) - no Ctrl-right arrow shortcut to accept part of the suggestion when it gets it half right I plan to upgrade to paid to see if Claude 3 Opus does a better job. I'm reading that the big advantage of the slower, more expensive models is understanding prompts better, which is valuable.

Following the Claude 3.5 Sonnet hype over the past few weeks, I decided to give it a solid try - by asking it to build production ready feature sets. I don't say this lightly: Claude 5X'd my coding productivity. This was particularly surprising, given how bad ChatGPT was in performing the same tasks. More important than the productivity gain, was its ability to allow me to build features I wouldn't have otherwise considered building - given how much thinking I'd have to do, and how much research and documentation I'd have to read to make it possible. What's interesting here is that Software Engineers who have worked under the agile methodology have a natural bias to work on smaller feature sets that are easier to iterate on. Paired with an AI pair programmer like Github Copilot, what could take a full day to code*, can be done in an hour. The implications here are worth a discussion of its own. I don't recommend that developers blindly jump into AI. Generated code is almost always incomplete. Without experience, it's difficult to identify the inefficiencies in generated code. Typically, this experience has come from tenure: writing code, and getting your hands dirty with problems. In this way, the famous "calculator" argument for AI generated code doesn't apply - because calculators are always right; Generated code isn't. Nevertheless, given that these AI tools are here to stay - it's becoming increasingly clear to me that the new skill isn't "how to write code", but "how to generate the right code". The latter is aided by experience that comes with the former, but going forward - we're going to need to reimagine how more inexperienced developers can be equipped to generate the right code without the experience of writing code.

I took a short break from the vehicle tracking project to revisit a codebase to formulate an optimal strategy in a game of Connect 4. The rules are simple: the first player to connect four dots wins. I first heard about this coding exercise in 2018 when my friend from the library (then working on a master's thesis in economics) said they had to program it within a week. (Wow!) Now that I have Gemini, Colab, and ChatGPT to help me, this problem has become much easier, and allowed me to focus more on strategies and less on debugging. The strategies in the video: 1. Completely random game. Neither player blocks the other. Neither player makes decisions based on long-term outcomes. Both players blind to the rest of the board. Just random. 2. Hyper defensive strategy for blue only. Fixates on immediately blocking the other player, but otherwise selects a move at random. (I messed up this code and gave blue no long-term strategy.) 3. Hyper defensive (otherwise random) strategies for both red and blue players. 3. Very witty, well-matched players. Both players seek the shortest winning strategies, five moves deep. This creates a decision tree with 5^7 = 16807 possible outcomes stored in memory and selects the smallest (Thanks, Colab 👍). Here's the link to the Colab notebook, in my Github repository: https://lnkd.in/gvCTQyeM

Weekend well spent in a hackathon of building software solutions using no/low code applications. Our team's topic of choice was improving trading front-end application features. Our aim was to classify stocks based on their daily volatility and see if it fits a trader's most recent trading preference based on volatility. If a stock is classified as aggressive and it does not match with the trader's appetite then it will not be 'suggested' to the trader in the recommendation list. Given more time, we would add a feature of matching market risk profile of traders based on their historical trading pattern. Another solution prepared was to have a feature to enable traders to Q&A with plain language which suits average traders who want to explore the world of stock trading. For the prototype data architecture, we decided to use Dataiku to preprocess data and train machine learning model. Data was obtained from open source like Yahoo finance and Kaggle and was retrieved using python codes entirely generated by ChatGPT. ChatGPT and Bing were also incorporated into our proposed solution to provide opinions on a stock we would like to enquire. Our solution was visualized using Streamlit. We did not win the prize money but so many things learned in the past 2 days especially I am coming from business-investment background. From machine learning deployment, teamwork, exploring new tools and most importantly to developing a problem solving mindset ! #hackathon #investment #software #application

This might sound very basic, but for someone like me who doesn't know coding at all, it was surprising how a few prompts helped me get the code of a new application I have been working on. Now, I'm not sayings it perfect, its obviously going to be very basic considering that I provided very generic prompts. But it gets us somewhere. Chat GPT is a blessing. Even if I don't end up creating this application on my own, in less than 5 minutes, I have learnt what goes on in the backend and how frontend workflows need to be developed. I might not have enough knowledge to talk code in-depth but I might just be able to partake in a discussion and ask relevant questions to learn more.

#claude-3.5-sonnet. Saw lots of raves on LinkedIn (I think someone even posted a 98% score on coding) on how good claude sonnet latest is with python coding. I think I may have even seen a post where it could run the python code with external data. Git hub copilot is driving me mad and #chatgpt4o wasn't really justifying itself at £20 a month, so I thought I would give it a go. Silly me (but I quite like it) Basic #langchain prompt template and chain, it used deprecated .run() and still hasn't solved the error after 3 iterations. Annoying as I now have to go read the docs again. I haven't tested enough to prove it, but ChatGPT4o may have the edge if able to read up to date docs and with a lot of packages in #GenAI changing rapidly being up to date is important. I definately don't want to have to build some RAG with docs retrievers, after all this is just play time stuff not work. I haven't empirically tested yet (nor will I) but I think I will stay with it, I found the workflow of discussing and overall app, breaking it up defining functions and how things will work enjoyable and interactive. ChatGPT4o was pretty good at that too, but I found it effortless and was able to challenge approach and adjust and build up code gradually. Oh how I wish it could run the code and feel silly I believed a post that it could, but in reality, I'm going to go with claude for a bit. Sorry ChatGPT4o voice mode was fun, but until I can just talk to you through the day and you write my code, I am going to be unfaithful. P.S. Yipee LI doesn't even allow re-write with AI when you cancel your premium, what a relief.

Interesting survey of Software Developers with varying experience levels and companies sizes by The Pragmatic Engineer ⦿ ChatGPT and GitHub Copilot are the most popular; other tools include Google’s Gemini, Antrophic’s Claude, Jetbrains AI, and Codeium. ⦿ Common uses of AI-assisted workflows are code-gen in IDEs, AI chatbots for searching, AI-assisted debugging, and learning new languages/frameworks. ✓ The good: AI tools significantly help in completing projects, increasing test coverage, and facilitating experimentation. ✗ The bad: poor output, hallucinations, and over-reliance on AI tools. ⦿ Small changes compared to the survey run last year; increased interactive “rubber-ducking” and more team experiments with AI agents. Full post: link in comment