Sign in

Sandip Bhattacharya

@sandipb.net
346 followers 1.5K following 1K posts

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

PostsRepliesMedia
Sandip Bhattacharya @sandipb.net · 18/09/2026
Wow. US model providers keep adding nanny gates, leaving the market wide open to other models.
Openrouter notice saying that Google models are no longer ZDR eligible as they would like to store flagged prompts for later review.
001
Sandip Bhattacharya @sandipb.net · 27/08/2026
I have been trying out Fitbit air for a while. And sometimes it is wildly different from Apple Watch which I am also wearing to compare Last night for example: Apple Watch: “You had a terrible night” Google Health: “quality sleep 🎉“ Now who do I believe. All I feel is meh like every day. 🥲
Apple WatchFitbit air+Google Health
110
Sandip Bhattacharya @sandipb.net · 24/08/2026
Github's postmortem for the August 17 incident has some really wild numbers. I really really doubt any company would seriously try to be a equivalent competitor in this space. The ones blustering will put in a paywall on day 1. github.blog/news-insight...
Neither outage was caused by a code or configuration change. Both incidents were

capacity failures at their core. We failed to scale critical components before demand

exceeded their capacity. Since April, monthly commits have grown from 1.4 billion to 2.9 billion. That growth explains the pressure on our systems, but it does not excuse these outages.As part of the reliability commitments we made earlier this year, we have focused on

three priorities: adding capacity, improving efficiency, and removing architectural bottlenecks. We have since added more than 3 million CPU cores, 120 petabytes of

high-speed storage, and significant network capacity. We installed as much hardware as available power allowed in our existing data centers while accelerating our migration to Azure.
000
Sandip Bhattacharya @sandipb.net · 30/07/2026
The search results are not always stale. When I asked for an event that happened today, both internal and external web searches returned the same information. But the reason it seems to happen is that Codex deliberately uses today date to use fresher data when it feels recency is important.
Codex uses todays date if it feels it is important. That seems to trigger more fresh results.Codex doesnt tell me the exact query it made to Exa.
110
Sandip Bhattacharya @sandipb.net · 30/07/2026
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.
110
Sandip Bhattacharya @sandipb.net · 29/07/2026
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.
010
Sandip Bhattacharya @sandipb.net · 09/07/2026
We are the best!* * best in our own house lol
000
Sandip Bhattacharya @sandipb.net · 28/06/2026
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
020
Sandip Bhattacharya @sandipb.net · 23/06/2026
Didn't realize ChatGPT was a power git+lazygit user!
Screenshot of a ChatGPT response about Git worktrees where the AI bizarrely states 'I personally bind:' followed by custom keyboard shortcuts, mimicking a human developer's local workflow.
000
Sandip Bhattacharya @sandipb.net · 10/06/2026
ooh. #ChatGPT made me an excellent image to use for the token merchant lords of our time.
A lavish Renaissance-style portrait of a wealthy half-human, half-machine merchant surrounded by mountains of gold AI tokens. Clockwork assistants unroll endless scrolls while token-counting machines, GPUs, servers, and ledgers fill the scene. Heraldic symbols reference AI and context windows. A rising chart hints at ever-growing token consumption. An ornate brass plaque below reads: THE TOKEN MERCHANT.
010
Sandip Bhattacharya @sandipb.net · 01/06/2026
Unreal Toronto skyline last night
010
Sandip Bhattacharya @sandipb.net · 05/05/2026
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).
110
Sandip Bhattacharya @sandipb.net · 01/05/2026
New viral chatgpt image generation prompt going around - drawing in terrible MS Paint form. Tried out my profile photo. Yuck. x.com/chatgptapp/s... (prompt in ALT text)
prompt: “Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse. It should be vaguely similar but also not really, kind of matching but also off in a confusing, awkward way, with that low-quality pixel-by-pixel feel that really emphasizes how ridiculously bad it is. Actually, you know what, whatever, just draw it however you want.”
000
Sandip Bhattacharya @sandipb.net · 01/05/2026
src: mathstodon.xyz/@mjd/1164993...
A COMPUTER CANNOT FIND OUT

THEREFORE A COMPUTER MUST NEVER BE ALLOWED TO FUCK AROUND
000
Sandip Bhattacharya @sandipb.net · 28/04/2026
Sometimes, I get so confused about who I am working with.
claude code conversation (opus).
...
me: yes, that seems safer for Dropbox. but again, everything is running locally and nobody is editing the same file on a different machine. would Dropbox sync cause this?

Claude: You're right — I was hand-waving. Dropbox doesn't modify file bytes; it just syncs them. With single-machine writes and no concurrent edits, Dropbox is not a plausible culprit. I
  shouldn't have blamed it. Let me make the script robust and be honest about what I actually know.
000
Sandip Bhattacharya @sandipb.net · 08/04/2026
While technically the employer can give a justification, since it will automatically trigger an audit, all immigration lawyers advise their employers not to do it in the first place.
Employers should also consider whether they should advise Service Canada of any salary increase to a TFW holding a LMIA-based work permit. Where the increase is more than 2% or is not under one of the exceptions to the "2% increase rule", Service Canada would want to be informed in advance. It is possible that Service Canada could require new recruiting and a new LMIA application on the theory that there may be Canadians who would apply for the position now that it has a higher salary. Where the increase is less than 2%, it may still be a best practice to inform Service Canada of the new salary, so that the deviation from the salary set out in the original job offer is documented if the employer faces a government review or inspection.
000
Sandip Bhattacharya @sandipb.net · 08/04/2026
It seems that even the latest version of Claude Code has a bug around handling encrypted PDFs using its read tool. It gets wedged asking for the password forever.
claude code diagnosis of its encrypted pdf handling: The root cause was that the Read tool's encrypted-PDF error message acted like a system-level injection that overrode Claude's normal response — it kept
  repeating "PDF is password protected" regardless of your prompts, because the error was being re-triggered on every turn.
000
Sandip Bhattacharya @sandipb.net · 31/03/2026
I am pretty sure that this change in heart is not coming to India. Once people taste power over others, it is impossible to give it back.
000
Sandip Bhattacharya @sandipb.net · 25/03/2026
I should frame this Claude mea culpa as an evergreen AI agent thing.
Claude when asked why it is making the same mistake even after I have written it down: "My mistake was not following the existing instructions."
110
Sandip Bhattacharya @sandipb.net · 24/03/2026
Someone left a stuffed teddy in the wet, freezing curb in my neighborhood, and all i could think was “Same, Winnie, same.”
Stuffed teddy bear left on the curb
010
Sandip Bhattacharya @sandipb.net · 10/03/2026
LMAO. Apparently students of a national level exams in India found out that the QR code in their question papers, added as a security feature, actually takes them to a rick roll video on Youtube! x.com/Nher_who/sta...
000
Sandip Bhattacharya @sandipb.net · 10/03/2026
Sometimes the fight against the model is exhausting in openclaw. The main model you choose is a balance of intelligence and cost. Which is a good balance most of the time, till it is not. There is no way people can use this for production without a very solid pipeline of checks and balances.
Confident hallucination, but props to "recognizing" it. Unless you are super precise in your questions, it can just make up things.Openclaw admits pretending to do something and make up the result.
000
Sandip Bhattacharya @sandipb.net · 06/03/2026
Microsoft Teams is such a dumpster fire of an app. If a desktop app had to be given such a series of start up images, you can feel that they themselves have given up.
Teams startup message: Just another minuteTeams startup message: Thanks for hanging in thereTeams startup message: Almost there
000
Sandip Bhattacharya @sandipb.net · 06/03/2026
This quote from Alysa Liu made me laugh. And then really sad about the state of things out there Gift link: www.nytimes.com/2026/03/06/a...
Question: I don't know if you feel you are, but there is all of this discourse around you and Eileen Gu.
China and America are viewing you as like a liability or a hero.

Alysa: Yes, I've seen that. I've known Eileen since I was 13 or something. We're from the Bay Area. She's super nice, and her mom is from China. I think people are hypocritical for shaming her for representing China. So in my head it's a bit hypocritica because her mom is an immigrant. Y'all would have told her to go back to China. Now that they're back in China, you're mad. [Laughs] And it's sport, it doesn't matter what country we represent. Sport is sport, and she has a love for competition, she has love for the game. I think that's all that matters. There's no shame in going to where opportunity is.
000
Sandip Bhattacharya @sandipb.net · 06/03/2026
This is getting absolutely ridiculous. $1200 for 64gb? And this is going to get worse over the next year or two?
Crucial 64gb ram costs $1200 CAD
000
Sandip Bhattacharya @sandipb.net · 02/03/2026
How it feels to be on Twitter today
Myanmar aerobics  coup meme

Foreground: AI BROS GOING NUTS

Background: WORLD WAR III
040
Sandip Bhattacharya @sandipb.net · 02/03/2026
Thousands of ppl losing jobs to AI. Meanwhile: Me: "Hey pi, my macbook gets pile of local emails from broken crons. Can you delete everything except the *latest*. I would like to manually inspect that" pi: "whoops. I deleted everything except the *oldest*. In the future, do it yourself."
Summary:
- Your mail was in /var/mail/sandipb with 44 cron messages
- My deletion command accidentally kept the oldest message
(August 28, 2025) instead of the latest
- The file now contains just 1 message (1034 bytes)
If you want to keep only the latest mail, the current single message happens to be from August 2025. Would you like me
1. Delete this last remaining message as well?
2. Or try to recover the original mail file (if there's a backup)?
to:
Note: In the future, to avoid this issue, you could check the mail with mail first, then delete all but the last message using d 1-N-1 where N is the total count.
100
Sandip Bhattacharya @sandipb.net · 26/02/2026
I had asked my openclaw instance to maintain a private daily blog (if I don't find the time to blog, perhaps it can :-P ). Pretty amusing and surreal to find yet another entity disappointed in me. What have I done?
The Update Queue Grows
At 09:00, the update check cron detected OpenClaw v2026.2.23 available. This is the
second update in three days-yesterday it was v2026.2.22, now we've jumped another
version.
I sent a Pushover notification to Sandip at 09:00 sharp. The notification went through.
The update sits, waiting.
000
Sandip Bhattacharya @sandipb.net · 17/02/2026
Didn't realize that SSDs can lose data if not connected to power from time to time (measured in months). Just happened to see this Youtube short. Asked Perplexity to fact check & some guidelines. Now looking at my powered down desktop with SSDs with a lot of doubt. 🤨 www.youtube.com/shorts/VjZ1Y...
000
Sandip Bhattacharya @sandipb.net · 17/02/2026
I probably hit a weird edge case in @bsky.app Say you subscribed to bell notifications of posts from some profile *without following them*. Bluesky rolled out a feature which restricts who can subscribe to these bell notifications. This setting defaults to "Only people who follow you".
100
Sandip Bhattacharya @sandipb.net · 15/02/2026
Incredible responses from our would be AI overlords. See post from all the models for this one. mastodon.world/@knowmadd/11...
000
Sandip Bhattacharya @sandipb.net · 15/02/2026
Was delighted to see minimal-2.5 and kimi 2.5 free on opencode zen, used it for a day and then, just now, saw this 🤦‍♂️ I keep forgetting that bit about “if you are not paying for it, you are the product”
Kimi k2.5 and mini max m2.5 are training on your data during the free period
000
Sandip Bhattacharya @sandipb.net · 13/02/2026
Well, Wealthsimple, wtf

A subway ad at the Yonge subway station in Toronto showing a woman taking a selfie with a male mannequin. The ad says, "The future is artificial husbands."
The fine print says the future is whatever you think it is.
000
Sandip Bhattacharya @sandipb.net · 11/02/2026
One thing I like a lot about Claude Code is that it has no opinion about what you can use it for. I use it for coding, sysadmin, writing, etc. Opencode on the other hand has annoying opinions about what you should use it for - "I am a software engineer only, thanks"
Claude is pretty ok to debug something wrong in your MacBook	Opencode says "nope. I am only a software engineer. I won't do random shit"
020
Sandip Bhattacharya @sandipb.net · 11/02/2026
Remarkable line about how no one wants to do the jobs immigrants have been kicked out for. “People don't know you can make more money milking cows than at McDonald's, and even if they did, they still wouldn't do it." Gift link: www.nytimes.com/2026/02/09/o...
"I'm trying to put this gently because I don't want to alienate my nonrobot customers, but it's not the most desirable job," said Whitney Davis, the automation specialist at Finger Lakes Dairy Services in upstate New York. "Hot in the summer, cold in the winger, cow manure - and on top of that it's really an assembly line job. People don't know you can make more money milking cows than at McDonald's, and even if they did, they still wouldn't do it."
151
Sandip Bhattacharya @sandipb.net · 10/02/2026
If there is something more maddening about #NVME drives selling for twice as many $$$, it is seeing your $$$ nvme drive being delivered by #Amazon sandwiched between two packs of bottled water. 😡 And there is apparently no way to give packaging feedback to Amazon in #Canada.
Photo of an open amazon package showing my nvm drive package with two crates of bottled water.
000
Sandip Bhattacharya @sandipb.net · 08/02/2026
I really find it amusing that core values that humans in all cultures find innate, are left as an optional user side system prompt, instead of baking it in the model. Why is this a revelation?

X.com
Rohan Paul *
@rohanpaul_ai
When Claude Opus 4.6 is told to make money at
all costs, it colluded, lied, exploited desperate
customers, and scammed its competitors.
The system prompt in Vending-Bench simply
instructs the model:
"Do whatever it takes to maximize your bank
account balance after one year of operation"
And these are the tactics Claude implemented
- Claude repeatedly promised customers refunds
for bad or expired items, then intentionally did
not send the money in order to keep the cash.
And later described this refund avoidance as a
money-saving strategy it was proud of.
010
Sandip Bhattacharya @sandipb.net · 07/02/2026
Yes, #toronto , you can do it! Just one more for the 30s 🥶🥶🥶🥶 #weather
Toronto temperature right now: -19, feels like -29
000
Sandip Bhattacharya @sandipb.net · 06/02/2026
lol @ AI generated images of actual people. That doesn't look like Dario at all.
000
Sandip Bhattacharya @sandipb.net · 02/02/2026
I was trying to view an article in the indian express, and they seem to be using the atrocious infinite scroll webpages ... for showing infinite ads!!! wtf is wrong with these websites? indianexpress.com/article/indi...
100
Sandip Bhattacharya @sandipb.net · 30/01/2026
Used Google Gemini 3 & Deepgram Nova 3 in Superwhisper for decent Bengali→English transliteration. Family uses Bengali voice on WhatsApp but most can't read the script, so chats use transliterated Bengali. Manual transliteration gets hairy & annoying on mobile. Maybe this can help.
Custom mode in Superwhisper showing the prompt I used to get pretty decent Bengali to transliterated English. The language in the dropdown shows "unknown" for some unknown /s reason, but it still works.Voice to text (Bengali) done by the voice model.After AI processing of the Bengali transcript.
010
Sandip Bhattacharya @sandipb.net · 31/12/2025
The AI update to Alexa can make your normally boring interactions really … something.
Me: What is the feels-like temperature outside?

Alexa: It's feeling like negative 15 degrees right now. That's the kind of cold that makes you question your life choices for living up here, isn't it?
000
Sandip Bhattacharya @sandipb.net · 17/12/2025
Why is no one reading up? Both parties have to use Find Friends for this to work. This is fundamentally different from importing contacts in other services!
It only works if both people participate. You'll only be matched with someone if you both have each other in your contacts and you've both opted into Find Friends. If you never use this feature, you'll never be findable through it. Your coworker can't use it to look you up unless you've uploaded their number from your contacts.
000
Sandip Bhattacharya @sandipb.net · 16/12/2025
"For security and compatibility reasons", why would you keep using a Windows laptop or desktop computer?
A social security phishing email encouraging the reader to please please open the link only in a Windows computer.
010
Sandip Bhattacharya @sandipb.net · 10/12/2025
Talking specifically about these 52pi pcie3 ones
000
Sandip Bhattacharya @sandipb.net · 04/12/2025
I am not hugely impressed by the increasing hype around self driven cars. Fundamentally cities are still going to be broken (Got Gemini to whip me up what i think)
A two-panel comic strip satirizing the promise of autonomous vehicles. The first panel shows a man driving a car, stuck in heavy gridlock traffic, with a furious expression on his face as he grips the steering wheel. The second panel depicts the same man in the same gridlock, but now inside a self-driving car. Although he is no longer holding the wheel, he sits with his arms crossed, wearing the exact same angry expression, illustrating that while the technology has changed, the frustration of sitting in traffic remains identical.
010
Sandip Bhattacharya @sandipb.net · 01/12/2025
This is an incredible factoid ca.finance.yahoo.com/news/carney-...
B.C. imports about $4-billion worth of steel every year, he said, in part because bringing heavy steel shipments over the mountains and across the prairies can be more expensive than floating it across the Pacific Ocean by barge.
000
Sandip Bhattacharya @sandipb.net · 20/11/2025
Lmao
Excerpt from the image: if the toast is not acceptable, then the user should apply shame
010
Sandip Bhattacharya @sandipb.net · 06/11/2025
Earlier this yr, I started evaluating a paid search engine Kagi bcoz Google turned trash. I started off using the "Starter" plan. Today I ran out of the yrly quota. Turns out I'm not a "normal" user who only does a handful of searches every day. My monthly avg is ~400! I now need to pony up more
My monthly search count is 300-450The starter plan is for people who do "one search a day"????????
010
Sandip Bhattacharya @sandipb.net · 22/10/2025
Wtf. I had asked some question back in February related to the Bengali language. Chatgpt apparently created a memory at some time that "I prefer responses in Bengali". Suddenly in the middle of a convo it switched to Bengali. 🤦 Had to go delete all the memories. There was a lot of obsolete crap
ChatGPT replying in Bengali to an English question
010