Sandip Bhattacharya

@sandipb.net

Opensource/food/other fleeting interests. Observability SRE. "Centrist" AI enthusiast. Rationalist/Humanist. Currently Toronto. ⛺️: 🇮🇳🇺🇸🇩🇪🇨🇦 @sandipb on Twitter

I did a side-by-side comparison of Codex using its internal web search and a web search mcp like Exa. The internal web search gives a more stale result. Not too stale - about 4 days

Using Exa gives you the latest actual release detailsUsing Codex's own search gives you a bit stale results.

While replacing my NAS drive, I wanted to double check using AI. Frontier model GPT 5.6 Sol got confused even after an utility point out the mapping, made a bad assumption and almost made me pull out a good drive from the NAS. AGI is pretty far out.

› how did you get it wrong

• I incorrectly assumed Linux /dev/sata4 meant DSM “Drive 4.” Synology’s kernel device names do not map directly to physical bay numbers.

  Worse, synodisk actually reported Disk ID 2 for /dev/sata4; I overlooked that and used the path number. I should have required DSM’s Locate Drive confirmation before naming a bay.

  Your screenshot confirms no harm occurred: all drives remain healthy.

Remember how Apple got so much hate for deprecating old hardware when new ones came out? Even though the older ones were good enough for their users? LLM providers are doing the same. Cheaper older opus and sonnet models would be good enough for a lot of ppl. But no, you need to pay more for new

People here standing in Church street, having ice cream, vibing to the dj at the pride event behind them and watching the World Cup, probably experiencing the most funmaxing I have seen.

church street toronto, with a dj playing dance music in the background, an ice cream truck, and a bunch of people at a cafe watching the Columbia vs Portugal match

Sad. It seems I just read an article on scroll.in that smelt mostly AI generated or at least extensively AI edited to the point you can make out common sentence structures. I expect most news outlets to at least run an AI scanner to remove smells

I don’t understand why some people don’t get it. The longer your AI post gets, the stinkier that AI smell emanates from it. Keep it short and you might get away with it. Even your fellow AI enthusiasts are unlikely to be with you on this one.

Pi + qwen 3.7 plus is such incredible superfast AI agent experience like no other. I just have to be careful to prompt pi (which is always in YOLO mode) so that changes always go through review.

Too much hype about GLM-5.2. I used it and it is good for coding, but very people are mentioning - *it is not a vision capable model*! If a model cannot see, there is a lot of feedback you cannot give to the model! I am back to Kimi and Qwen.

For people on ChatGPT Plus plans(~$20), stick to 5.4 instead of 5.5 to make your accounts run twice as long. I asked ChatGPT to do a cost benefit comparison and this is what it came up with.

A screenshot titled “Core comparison” showing a table of AI models with columns for Model, Type, Intelligence (relative), Speed, Input cost per 1M tokens, Output cost per 1M tokens, and Notes. Models listed include GPT-5.5 (highest intelligence, medium speed, $5 input / $30 output), GPT-5.4 (near-5.5 intelligence, medium-fast, $2.50 / $15), GPT-5.3 Codex (coding-focused, fast, $1.75 / $14), GPT-5.4 Mini (low-cost, very fast, $0.75 / $4.50), and GPT-5.2 (older, largely obsolete). Below is a second section titled “Intelligence level multiplier” describing settings (Low, Medium, High, Extra High) with increasing reasoning depth and cost (~1x to ~10–25x), along with suggested use cases. A note at the bottom explains that model choice and reasoning level are independent (e.g., a stronger model with low reasoning differs from a weaker model with high reasoning).

Tip: If ur Claude Code token usage is burning up ur quota but you still want to use Opus, Opus 4.6 uses less tokens than 4.7 but is still pretty good Use this to start a new session: claude --model "claude-opus-4-6[1m]" Or this to switch in the middle of a session: /model claude-opus-4-6[1m]

#Obsidian says that it does a lot of things to make sure that the actual #markdown files are free to be read via any other application, but that's not true in spirit. There are a lot of Obsidian features that result in what can only be called broken markdown.

I wanted to add mine & just two more blogs from ppl I knew to Kagi's small web: blog.kagi.com/small-web It was a ridiculous amount of effort to find people I who I knew to be bloggers, and still had a blog that has been updated in the last year. Most blogs are dead or moved to walled gardens

Kagi Blog - Kagi Small Web

As a part of our ongoing pursuit to humanize the web, we are pleased to announce the launch of the Kagi Small Web initiative. What is Kagi Small Web? To begin with, while there is no...

blog.kagi.com