ghoul2

The article doesn't cover this, but, a huge contrast is Torrent Power (Ahmedabad/Surat), which has consistently (over many decades, the ahmedabad power company is actually over 100 yrs old) maintained world-highest levels of electricity supply and service quality, distribution losses, billing efficient. Beating Tokyo, equaling Singapore.

Zero load shedding, on-site within 10 minutes typically in case of fault reports, ALL supply lines underground. Planned maintenance schedules communicated week+ in advance (rare, once every few years).

it was just

  10 PRINT "my first name"

on my zx spectrum, as a 6 year old. did me in. not even a loop. Just seeing that on our color TV (also new at the time, in my country).

I would say ALL of science, not just "much" of it. The difference between math and science is that in science, hard reality is the bottom line. No matter how good your model, how much data you have, what p values you get against simulations - hard reality wins everytime. Thats also why the so called soft "sciences" aren't sciences at all: they are not dealing with hard, objective reality at all.

Math + real world observation -> science

The traffic still has to flow from an india provider to the international leg and back via the domestic provider.

In anycase, then they should document it clearly that they have unacceptable insertion latency in india and the free plan is entirely unusable for india. Instead of advertising '10 pops in india!!'

I also believed that, but at least for India, this doesn't work in practice. Unless you atleast do the 25$ pro plan, cloudflare routes even india-to-india, hell, even mumbai-to-mumbai and aws_mumbai-to-cloudflare_mumbai traffic via Marsaille!! Not even Singapore. The unstated reason is that indian transit is expensive, though I fail to believe its cheaper to go from mumbai to marsaille and back to mumbai.

I am guessing the real reason (and at this point I am discounting incompetence - this has been true for years, so they are aware). You switch to the pro plan for the zone and everything now routes within india, 100s of milliseconds of latency saved.

Its even worse for workers and workers AI and embedding search. I found multiple seconds of latency, all vanishing the moment the zone is on pro plan (It seems R2, workers, workers AI - none of them are deployed in an India POP - unconfirmed, of course, cause there is no way to actually communicate with cloudflare).

Now 25$/month isn't much - though it does change calculations compared to "FREE!!" - but I would have liked to know this going in, instead of discovering this after having made the commitment. Seems like a deliberate dark pattern, to force people into the pro plan.

Shame, really - I love the CF stack(workers and DO are just so fantastic to build on), but these shenanigans, plus the utter refusal to provide ANY level of support, keep souring me on them.

This seems a worsening that seems to have started at Opus 4.8. When I (was forced to) switched from 4.7 to 4.7, I noticed that 4.8 had a very strong bias against viewing tool output or reading files. It ALWAYS wanted to 'grep' things. So for example, it would run the test suite and grep for 'PASS', thus completely missing the 4 tests that FAILed. Its bias is so strong, that it does this despite explicit instructions to the contrary, and continues to do so even after having noticed its own mistake, in the same (claude code) session.

For editing files, it again seems to prefer writing small python scripts, rather than read the file and rewrite it (even small files, like notes). This again causes it to constantly miss duplications, inconsistencies, etc.

And from 4.8 to 5, it doubled down on this. At this point not moderately-long-horizon task is getting done consistently, due to its pigeon-hole view of the workdir.

Anybody else notice this?

if sqldump isn't a backup, what is?

Exactly.

I am now looking at 9 to 10x the app code, for non trivial stuff. And this is not just the usual unit/integration/e2e suites too. Extensive deadcode, field use, code shape tests (ast walking) - god objects, 15-param helpers, wrapper-piled-upon-wrapper, badly named (a skill) functions/methods/objects, path-dependent artifacts, etc. There is no way I could have written all this manually.

It doesn't guarentee high quality OR reliability or readability or maintainability, but does reduce churn and makes me feel a bit more confident about deploying route A work product in prod.

> Lint rules, type errors, commit lint, anything

These are basic: they work for human coders and are typically quick to setup. But gates at this level are far from enough for agentic coding. You need an extensive test suite/e2e suite/benchmark suite. Typically much larger than the code base itself: This is what enables those "i ported bun to rust in a weekend" headlines. Where ever such extensive 'gates', high quality ones, already exist, the agent can do a good job of making things work, cause it can automatically iterate.

But: every hole, every gap in coverage, will be eventually found by the coding agent and 'exploited'. So you need full, extensive coverage.

This suite itself is a LOT of work, much more than what it would be if you only had to worry about human coders. So if YOU are writing the gates yourself, you are going to fall between A and B: you still have a convoluted, unreliable, impossible-to-reason-about prod app, AND it still took you eons to build it cause you spent that time writing the gates. This is no win at all, on either front.

Where gates, extensive ones, help in route A is they can significantly reduce the churn/spinning, fix-x-break-y, to just the parts that are gaps in coverage (which the agent will keep finding at a ridiculous rate). And your work would be to have it ALSO keep filing in those gaps with more tests/etc.

One of the reasons go coding with agents works better than other langs is cause I even include AST based enforcements in my governance suite. Its still route A, but with somewhat more confidence, somewhat less frustration.

The problem, in my experience is:

1. It writes SO MUCH code, in minutes, that theres no way I as a human can review it.

2. I am not very motivated to read/review this code anyway: it was cheaply written, by some _thing_ that is not going to improve from my feedback.

3. If you DO review it conscientiously , it becomes a never ending thing: you keep finding issue upon issue.

4. If you report the issues for fixes, the fixes normally fix that immediate issue, and typically add another parallel path/another option/another 100-odd lines of code, instead of a structural fix.

5. If the code base is large, and if you let AI write a meaningful amount of code in it, its no longer _your_ code. You lose the depth+width of understanding needed to reason thru things mentally, cause you no longer KNOW enough about the code.

6. If you do review seriously, and either fix things yourself, or have the ai/harness fix the issues for you, the rounds of fixes take so long, if you look back you realize it would have been better to just do it yourself in the first place.

My personal opinion there are really only two choices:

A. If you want to build FAST, using an agent, just let the agent write tests, validations, extensively, have it keep running them (it likes to call them gates), and just let it loose. Give up on the idea that its your code, and that you understand it. The bottom line becomes: does it work, and WHEN it breaks in weird ways, just use the agent to find and fix the issue (probably breaking something else in the process).

B. limit AI use for non trivial, prod-quality projects to limited research, very tiny targeted changes, write most of the code yourself still, and review every line. You won't get much speedup, maybe 20-30%, but it will still be YOUR code, and you will still be able to reason about it.

Isn't it an age old maxim: democratic governments should remember to not build tools/give them selves powers that they won't want the opposition to be able to use WHEN they come to power.

I just had claude write me a statusline:

https://gist.github.com/abgoyal/bde7af4ba0ba4d19094361667dac...

Shows the 5h and weekly usage %, when it resets, current model, estimated spend in the session, context used, cwd and project.

Save it in .claude/ and then add a snippet to settings.json:

  "statusLine": {
      "type": "command",
      "command": "bash $HOME/.claude/statusline-command.sh"
    }

India really has very deep penetration of 5g, and at very low cost. There might be a rare place that starlink might be needed but really I cannot image starlink having much consumer/retail uptake in india. Not needed, and too expensive. There might be commercial users - offshore rigs etc, but india is too densely populated for there to be many 'truly remote' locations.

India has still not permitted starlink to start ops.

There is a hilarious one in the Indian direct tax (Income tax) code right now:

If your (taxable) income is less than 7L (inr), no tax is assessed. If your income is greater than 7L, but upto 12L, tax is assessed, but a tax _rebate_ (a discount on payable tax) is granted for the exact amount of tax owed. This means net tax paid till 12L is zero. But if your income is 12L + 1, this rebate no longer applies, so the entire tax assessed (on the 7L+ income) is now due. This means that a person who earns 12L+1 INR makes ~70K inr less, post-tax, than the person who earns 12L INR.

After complaints, to correct this, they added another layer: If your taxable income is >12L but less than 12L+~70K (the tax thats due), then the maximum tax due is just the amount your income exceeds 12L bye. So you stagnate at the same post-tax salary from 12L to 12L+70K.

Recently, banks where also asked to put their official websites/netbanking on *.bank.in domains. I have wanted that for SO long.

Isn't that exactly what gvisor does?

Its a fantastic project, but seems almost inactive now. I have a tiny PR pending for weeks, even reviewed, but not merged. I have another patch I have not submitted as I want to first navigate the earlier PR to completion. Both were bugs that bit me, and I ended up wasting quite a lot of time trying to find it:

1. go:embed supports "all:<pattern" while tinygo silently ignore it. I ripped my hair out trying to figure out why my files were not showing up in embedfs. PR pending.

2. go allows setting some global vars at the build cli (think build version/tag etc). In the code, one can define a default, and then the value provided (if any) on the cli can override it at build time. Tinygo fails to override the value at build-time, silently, if a default value is provided for the var in code. This PR I have not submitted yet, as its more intrusive.

I hope it picks up steam again soon. I love using go for embedded and CF worker use and tinygo makes both of these use cases much more viable than regular go. Honestly, I hope tinygo can be rolled over into the main go toolchain as "target arch".

Honestly, I am vastly more interested in how they built the model: its not backprop-trained, it was done manually. And how did that bug find its way it?

I hope they next blog about that.

I long time ago, I was operating a "social network" which allowed image uploads. (India local, didn't amount to much as was shutdown.).

Immediately at launch, we started having a huge amount of (image) pron being uploaded into the pages. We put in some rate limits etc, but did not want to put any major restrictions of user signups etc as that would hurt signup figures (important to the ceo!).

We already had some content review people thru a temp agency on site, so we checked with them and they were fine doing this manual filtering of these images for us. All young (early 20s) women. While my team built a quick "dashboard" for them to be able to do this image filtering quickly and conveniently, I had a detailed conversation with them as I was very concerned about having them review this kinda stuff for 8 hours a day, 5 days a week. Truly nasty stuff.

They were _perfectly_ clear that they had no issues with it, and they told me in so many words to not give it a second thought and to let them get back to work.

It was a surprise, but it was a point of realization: there are much worse things they could be doing. And looking a porn has shock value only the first time. I was under-estimating these women and assuming they were some "snowflakes" who could not deal with something this silly and non-threatening.

Just my own person anec-data.

As a RA-ship student in the US, I still had taxed withheld on my meager grad student "salary", and was required to file a 1040NR-EZ. I obviously did not want to spent 50$ at H&R block, to claim like 200$ back. Fortunately the local library had sessions where you could grab a blank form (the library had boxes full of them), and someone walked a bunch of us (i think there were like 40 in the room), to filling it out. The session mailer had already listed the stuff we were supposed to bring with us at the session. Even if someone forgot, they could leave a place holder and take care of it later.

The session was an hour long, and I was done with my return by the end of the hour. I dropped it off in the mail on my way back from the library and that was the end of it. From the subsequent year onwards, a pre-printed form arrived in the mail with like 90% of the stuff prefilled, and it took like 5 minutes to fill out the rest and drop it in the post box.

I honestly didn't think it could have been any easier - of course, not having taxes withheld from a far-below-minimum wage salary would have been nicer.

When I lived in the US(5 yrs), early 2000s, from my second year onwards, I used to receive a pre-filled 1040NR-EZ with my W2 info already printed/filled-in on it. Typically, I would just add a deduction, and mailed it back. Does that program not exist? Or was it only for NR?

I have always counted to 20 on one hand. even as a kid. base, lower joint, upper joint, top. times 5 - including the thumb: my motor memory is trained so that i switch seamlessly from keeping the curse on top of the finger using my thumb, and then, once i cross 16, switch to using the index finger to "cursor" the thumb.

I asked this question on another post and was downvoted, trying again: don't we lose the "contextualization" that LLM embeddings do (embedding on Token X contains not just information about X, but also of all tokens that came before X in the context, causing different embedding for "flies" in "time flies like an arrow" vs "fruit flies like a banana")?

The image embeddings, as I currently understand, are just pixel values of a block of pixels.

What am I missing?

But does this not miss the "context" that the embeddings of the text tokens carry? An LLM embedding of a text token has a compressed version of the entire preceding set of tokens that came before it in the context. While the image embeddings are just representations of pixel values.

Sort of at the level of word2vec, where the representation of "flies" in "fruit flies like a banana" vs "time flies like an arrow" would be the same.

Isn't this "as intended" in the westminster-style system? The govt is formed by MPs from the majority party (or alliance). By definition they MUST be able to pass ALL money bills, which only require a simple majority. Any failure to pass a money bill is equivalent to the govt no longer holding a majority support in parliament. And that means either the king/president/govgen invites someone else from the current parliament who they have good reason to believe DOES (potentially) have support of majority of the parliament, or dissolve the parliament and call fresh elections if there is no such majority.

I am not quite sure why an action with such a clear established precedent be considered foreign interference? or was it the case that there WAS a suitable candidate with a possible majority but they were NOT invited by the govgen to try and win a trust vote in parliament?

well, Jio (Indias largest ISP, with ~300M customers) has cloudflare r2 blocked. And for a while even. I am sure its due to some bucket serving out pirated movies, but this is a bit insane. And neither Jio nor Cloudflare seem to want to figure this out.

I have stuff on R2, I don't personally use Jio, and the office has multiple ISPs but not Jio. Customers had randomly complained that some of the icons were broken/some files could not accessed, but never followed up our request for browser console/network snapshots, and our testing always showed everything working fine (we were very small with a handful of customers, at the time).

Finally, purely by chance, office network was having issues, so one of my QA people switched to their mobile hotspot, and they were on Jio. And then they could see all the broken stuff, but weren't sure why. Stuff escalated to me, and it finally clicked!

Easy to work-around by using a custom domain, though painful if you want to do access control/signed urls/etc as CF still only supports them on the r2.cloudflare urls. Had to put a worker in front.

Father passed away today morning. At least he is not suffering anymore. I don't know why I am posting this here. Just a tribute, maybe.

He was born to poor parents, and the eldest of 7 children. His father could never have paid for his education, but he has brilliant from the very start, and they lived in a village town, where the school had been built by a kind, local rich man, and my father studied his entire life on scholarship. He was the first person from his district to earn a masters degree (in electronics), and in his era, his bachelors and masters were each earned from the very top university of the country. Again, all on scholarship. Started a job as a professor, but quickly switched, first to a premier research labs, working on the very first satellites/payloads built in my country, and then later to being a technocrat in the state government: his qualification and ability were so high that he joined as the ranking technocrat of the state, and remained that till his retirement. Right around his job change, he got married and had three children - me and my two sisters.

During all this, he also earned an executive MBA.

After retirement, he didn't feel like sitting idle, and started a new career as the head of department (for two departments: computer engineering and electrical engineering) at a good, well regarded engineering college/university. He did this for 12 years, when he had a brain stroke, and then I asked him to not work anymore.

He recovered from the brain stroke, but more strokes/falls happened at a steady pace. He passed away at 79. The end was pretty painful, and I wish he had not had to go thru all that.

He lived a simple, harsh life - He never cared for comforts, or money, or luxuries. But he was genuinely content. In the end, all he really asked for was to "lets go home" (from the hospital). Fortunately, we did manage to get him discharged and his final few days were at home.

be at peace, pop. I love you.

Meh. Rocks are a feature, not a product.

I don't agree. At all. My father is right now in a very bad state: he can't recognize me, doesn't know if he is at home or in the hospital, is being fed thru a tube, and entirely dependent. He no longer has the mental ability left to decide on this on his own, but if he were, I am sure he would choose to go rather than live like this. He is not in excessive pain - he just has nothing left that can be called "life".

Unfortunately, in the place we live, there is no legal framework for choosing this anyway. So he is unable to chose, and I am unable to chose for him either. I absolutely hate to have him be in this state - such a brilliant man, reduced to this.

He had a (brain)stroke, and my grandfather on the mothers side also died of a stroke. So genetically speaking, my chances look (relatively) bleak. I know for a fact I won't want to go-on if I ever reach anywhere as bad as his current condition. But there's no option available to me either. I will have to suffer similarly, and the people I end up being dependent on will have to suffer similarly and watch me suffer similarly.

I think being able to decide when you want to "go" is and should be a fundamental right - a natural right - every person should have. And we would, if it weren't for this false religious crap being foisted upon the world.

Thats how it works in India. All authorized repeating charges ("mandates") are listed on a portal maintained by the card issuer. you can go in anytime and simply cancel the mandate from there. This is mandatory under banking regulations.

Credit cards are also required to be "tokenized" when stored at a merchant or payment aggregator - the user authorizes the bank to allow the merchant or the aggregator to "store" the card details for use later, and the bank then issues a card token, tied to the specific merchant/aggregator. They are not allowed to store the original card info at all - just this token. This makes the token not worth stealing, as it can be only used by that merchant, and is trivial to de-auth if needed, with or without merchant cooperation.