Notes on Miscellaneous AI stuff

Last updated on 10 Apr 2026
Over past few days, I viewed most of this video: Beyond Vibe Coding: Essential Python Concepts for Effective AI-Assisted Programming, 53 mins, published Nov. 2025. It is by an academic (computational chemistry, IIRC) from San Jose state university.

It is aimed at entry level students. So most of the video is pretty basic stuff. However, it gives decent exposure to Python and using Python in Google Colab. It also shows how to pick up python code from an AI chat (ChatGPT IIRC) response that had some computation, and use that in Google Colab.

I very much liked the introductory part of the video where the author talks of how coding skills help even with AI assisted development. Some of this introduction is also mentioned in the description part of the Youtube video page.
---------

17 Mar 2026

Basic level video that "demonstrates step-by-step how to access and generate text using google.colab.ai, a built-in module in Colab, ideal for quick experiments, learning, and prototyping with AI." It does not need API key.
--------

8 Apr 2026

Gemini Code Assist (GCA) in VS Code struggled to do a task involving a long Markdown file but web Gemini was able to do same task quickly and well. Pro model was used in both cases. Details follow ...

I uploaded a large Markdown document - ~150 KB and ~1500 lines, to web Gemini which was a reorganized version of the main document covering the many topics discussed in a chat. I told it further ...

Please ignore the Summary section at top of document, as of now. They will be developed later.

I have now got the # Power Backup For Desktop PC Details section done, except for its Contents part (at top of section). The Contents part is out of sync with the actual details. Can you pick up the Contents entries upto level 3 from the Details section, and give me an updated Contents section?
...

I switched from Fast to Pro model and continued:
Can you provide it to me in Markdown with links to corresponding section?
---

Then Gemini did a great job of providing me that markdown TOC of ~30 entries. I tried a few of the links and they worked correctly.

Me: Terrific job! I was trying to get the same task done in VS Code using Gemini Code Assist (which is correctly configured to use my Google Pro plan). It took too long and so I stopped it. You were able to do it in much lesser time. I faced some other issues too with Gemini Code Assist (GCA) in VS Code for entire document related task. It is using the same 3.1 Pro model. HOw is it that you were able to do it so well and quickly whereas GCA seemed to be hung! I should also say that for smaller content task, GCA while taking some time, did do the job.

Is GCA not fully supported by Google? I mean, does Google expect users to use AntiGravity and not GCA?

[Gem said GCA is fully supported by Google but VS Code extension has limitations ... See Summary below.]
...

Me: Hmm. So the takeaway for me is that large document wide tasks is where GCA may struggle due to VS Code Extension limitations. It would be better for me to use Web Gemini for such tasks. Am I on right track here?

...

Me: Give me a summary of our chat here on topic of large document task being problematic with GCA in VS Code but which you (web Gemini) could handle it very quickly and very well. I will put it up as a blog post. That will help me as a reference for future and also may help others.

...

Gem: Here is the background of the exact task you gave me, detailing how you arrived at this realization:
  • The Initial Problem: You were deeply involved in reorganizing a massive, highly detailed Markdown document that maps out your custom PC power backup architecture (covering your Dual Parallel UPS strategy, inverter setups, and thermal management).
  • The IDE Bottleneck: You attempted to use Gemini Code Assist (GCA) directly inside VS Code to update the document's Table of Contents. However, because you were feeding it such a large amount of text, the VS Code extension choked on the data size, hung up, and you were forced to abort the process.
  • The Web Interface Success: You brought that exact same large-document task over to me here on the Gemini Web interface. I was able to process the entire document and output the correctly formatted Markdown ToC almost instantly.
  • The "Why" Behind the Tech: This stark contrast in performance prompted you to ask me why GCA struggled so much while I handled it effortlessly, and you questioned whether Google was phasing out GCA. I explained that GCA is still fully supported but is fundamentally bottlenecked by the VS Code Extension Host when it tries to process and render massive text blocks. In contrast, this web interface is an unencumbered pipeline built specifically for raw throughput and heavy lifting.
  • Your Workflow Epiphany: You correctly concluded that you need to adopt a "right tool for the right job" approach: keep using GCA for surgical, line-by-line coding and local context within your workspace, but bring massive document rewrites or large data dumps over to the Web interface.
  • The Final Task: Recognizing how valuable this operational insight is, you asked me to summarize this technical reality so you could publish it on your blog. Your goal was to document this boundary between the two tools, ensuring you—and others—wouldn't waste time watching an IDE freeze over a task it was never built to handle.
----------
09:53, 10 April 2026

Past few days, I noticed that my 4G data consumption was high. Today I noted that VS Code was doing heavy network download. I turned off Gemini Code Assist Auto Completion. That reduced network download to zero or almost zero as I was not doing any other network download related activity.

Hmm. I should dig in later into this topic to understand whether GCA really does do lot of network download for Code Completion. 

I enabled CoPilot extension and saw that even CoPilot extension is resulting in lot of network download (while I am in VS Code). I disabled CoPilot extension. After some time, the network download traffic came down to zero or almost zero.

I asked Gemini about the above.

Gem: That is a very sharp observation. Since you are working on a 4G connection, you've essentially discovered the "hidden cost" of the modern AI developer experience.

You are correct: both Gemini Code Assist (GCA) and GitHub Copilot are remarkably "chatty" on the network. They don't just send a few bytes of text; they are constantly streaming data back and forth to maintain the "magic" of real-time suggestions.

Why the Data Consumption is So High

It isn't just the final suggestion you see on the screen that consumes data. The background activity is much more intensive:

  • Continuous Context Uploading: For a tool like Gemini to give a smart suggestion, it needs to know what you’ve already written. Every few keystrokes or every time you pause, the extension may upload hundreds of lines of "surrounding context" from your current file (and sometimes open tabs) to the cloud model so it understands the logic.
  • Frequent "Ghost" Requests: These extensions often trigger requests even if they don't end up showing you a suggestion. If you type a space or a common keyword, a request is fired off. On a fast 4G connection, these hundreds of tiny JSON payloads add up to megabytes very quickly.
  • Telemetry and Syncing: Beyond code completion, these extensions are constantly sending telemetry (usage data) and syncing your preferences or indexed "workspace symbols" to improve their internal logic.
  • Model "Polling": Some versions of these tools use streaming responses (where the code appears character by character). This keeps a TCP/HTTP connection open and active, which creates more overhead than a simple one-off request.

My View on Your Situation

Given that you are a retired consultant and a spiritual aspirant often working from Puttaparthi (where network stability or data caps might be more of a concern than in a high-speed corporate office), these AI tools can be a "data vampire."

  1. The "Tax" on 4G: If you are using a mobile hotspot, the overhead of the SSL/TLS handshakes for every completion request is significant.
  2. Productivity vs. Cost: For a physicist and software veteran like yourself, the boilerplate code these tools provide might not always be worth the massive data drain, especially if you already have a clear mental model of your React Native or PowerShell logic.
  3. The "Search" Alternative: You might find it more data-efficient to keep these "auto-completion" features off and only use the Chat feature (Ctrl+Shift+I or the sidebar) when you are actually stuck. A single chat query uses far less data than 500 auto-complete checks performed while you're just typing a simple for loop.

Recommendation

If you want the benefits of AI without the 4G bill:

  • Keep Auto-Complete OFF: As you've already done.
  • Use Manual Triggers: In VS Code settings, you can often set these extensions to only provide suggestions when you manually trigger them (e.g., pressing Alt + \ or Ctrl + Space) rather than "as you type."
  • Local Alternatives: If you ever find this data usage becomes a dealbreaker, you might look into "Local LLMs" (like Ollama with the Continue.dev extension) which run entirely on your machine's hardware and use zero network data.

Do you find that the "ghost text" suggestions actually speed up your React Native work, or do you prefer writing the logic yourself and just using AI for debugging?

**New Div for next entry**

Comments

Post a Comment