porridgeraisin

This is the most relevant comment in this thread.

This is the way. Works very well.

bound check elimination exists separately, PGO can help unveil more opportunities for BCE.

From my skim of their paper, they have gotten robodojo 14% versus previous sota 9%. So its an improvement, but doesnt seem like a reliable model yet, which is fine. I also notice the model becomes worse with more data on the printer refilling task, so thats either a bad run or its a data point against the robustness. The transfer learning results are meaningful. 75% on new complex tasks from <10 hours of demo is really good. I believe this is a 2x improvement over the SOTA. 10 trials per task is standard for these papers but is pretty weak statistically. Memory tasks its weaker at but they have acknowledged it, and I cant see why this arch would have problems w that in the future.

How is it much more expensive in europe? cos in india where we import most of our ONG and its one of the more expensive things on a household bill, it is right now 102INR/L = 0.93 EUR/L.

Insofar as PGO helps inlining.

https://xcancel.com/synthwavedd/status/2078514339552628880

Yes. The dhammapada uses a lot of Anushthubh chandas, which is an option in the website. However, the language differences between pali and sanskrit are enough to make TTS trip up. E.g only one "s" variant, and karma -> kamma, etc,. If you finetuned on pali a little bit it should work fine.

Yes, but monte carlo estimating the critic model is not new.

GRPO is policy gradient/PPO with your value function baseline monte carlo estimated using k rollouts. The only new thing is finding out it works well with binary rewards and LLM policies.

It's just another way to frame it. It's as foundational as the many other ways to frame it. I'm not aware of any major insight you get specifically from this framing. Is there one?

https://x.com/eliebakouch/status/2077428860639973471

gpqa diamond alone makes up 20% of their "capability index" and is responsible for ~70% of the gap between their model and nemotron nano (10 point difference!)

so what about contamination? i looked at the mixture and they train on a dataset "AIML-TUDA/QA-base" (for 10 epochs lol) that literally is a very very light rephrasing of the gpqa diamond test set

tl;dr: they trained on 10 epochs of a "very light rephrasing" of EVERY gpqa diamond EVAL data :)

Yep, I gave up on them for now. One of my older laptops also needs a charger that has the third ground pin, lest the trackpad become laggy due to electrical joise. I couldn't find a compact gan charger with a three pin plug and the third actually being connected to ground.

Gan chargers also die really fast.

But this particular instance isn't much related to that. In my school too, we did all that, and extending to beach cleanups. It's not like the same effect was there.

Besides, most garbage in most Indian cities/towns is not consumer litter in the first place. Side note: this is why 'guarded' areas such as metro stations are spotless.

This particular lab happens to be primarily privately funded, donations, grants, etc,. Most big ticket research at IITM like this one are industry-funded or donation-funded. The public funds cover all the usual stuff and I don't mean to understate it.

Other publications from the same place:

https://iitm.humanbrain.in/project.html

notably:

https://brainportal.humanbrain.in/publicview/index.html

This whole thread is 100% correct. Zoning, rezoning, related points are all correct. One minor correction though, the megacities where the comment mentions zoning being better is in most cases only true definitionally. For example, there are portions of chennai which are by the book, not part of chennai, but for all practical purposes, are. e.g, the new major city-wide bus stand is definitionally outside the city, but practically it is the bus stand to go to if you want to travel to the rest of the state, is not at all zoned properly and faces the same issues.

The only exception where rezoning has actually been done properly is hyderabad.

Here is the reason for this:

- The 74th amendment in india only calls for a _planning commitee_ (reminiscent of india's past) for any area with a population of over 10 lakh. But this MPC is completely distinct from the organisation that does the actual service delivery. In Chennai's example, the CMDA are the planning guys, and the GCC (greater chennai corporation) are the execution guys who actually employ garbage collection folks, buy cleaning machines, fix storm water drains, etc etc,.

CMDA covers ~6000 sq. km. and all that is practically chennai, and they plan bus stops and FSI and roads and all that across this region. However, the GCC is _still_ limited _to this day_ to just 500 sq. km and this is the "chennai" that appears in reports. So all your sanitation workers and clean roads, tree lined roads, good storm water drains, etc, that you see in say besant nagar, are completely absent in the remaining 5500 sq. km. Instead, you have multiple different corporations, Tambaram, Avadi, Kanchipuram, etc, and the coordination problems here are just insane. It also absolutely does not help that political fracturing happens due to dominant castes differing across these regions. You also lose the simple concept of rich people cross subsiding poor people - which is essential in any public service. The taxes paid by boat club road residents, remain with the GCC, and does not contribute even a little bit towards better roads on land under the Tambaram corporation, ironically leading to worse infrastructure for even the boat club road resident making a trip to Trichy. So you get nice parks and heart-shaped traffic lights in nungambakkam before even a road is paved in thoraipakkam.

In hyderabad, they did a very nice thing. They merged 27 ULBs directly into the GHMC, tripling its area from 650 to 2000 sq. km. So its the same thing CMDA did but through a merger. Those 27 bodies - with their staff, their ward councillors, their budgets - are now part of GHMC. The sanitation workers who used to report to, say, Medipally Municipality now report to GHMC. The property tax collected there now goes into GHMC's budget. The roads there are now GHMC's responsibility to maintain.

This unified thing devoid of coordination problems and with great cross subsidisation seems to work much better. It will be a good model for other cities to follow.

I don't know about amdavad and pune but I am willing to bet they are also good for similar reasons.

There are few phones these days with 8000mah -> 10000mah.

Personally, I only buy a new phone if the battery life improves with a weight increas of a smaller factor.

So I'm looking forward to buying a 9-10ah phone in the coming months.

My current phone is 6000 mah.

It's an age thing. People right out of college tend to eat out more and in general use city-only amenities more. We are also over-represented in online forums.

In the classic FDE model palantir pioneered, one of the main features was that you would use your learnings from one customer and integrate that back into your in house product, so that similar customers are serviced for cheaper later on.

With AI coding agents FDEs are now everywhere. One because they can demand a higher salary due to simply doing more due to AI. And two, because AI really accelerates the whole bespoke solution implementation thing. However, from what I hear, none of the actual "integrate that back into your in house platform" stuff is actually happening. So it's a tiny bit of a farce.

Unlike the app store, there is no 30% fee. And also. No chargebacks, so it's simple for merchants as well.

Another classic. About the X11 selection mechanism

https://github.com/porridgewithraisins/x11cp/blob/main/rant

Yes this. Anytime a SaaS supports UPI Autopay, I'm 10-100x more likely to subscribe.

For context: it's basically a recurring payment but you open the Autopay section in Google pay or whichever app and it shows you all your subscriptions right there (which is a major +), the transaction history, and cancellation happens in the UPI app itself I don't have to faff around in the SaaS and deal with potentially some dark patterns.

Funnily enough, I have my agent use vim sometimes.

I prefer commits to be granular to the extent that I manually edit the patch in git add -p, albeit rarely. So I have a pty plugin that my agent literally sends control codes to. This is also useful in general for LLMs driving any interactive terminal program. Long live text, I guess.

There is one thing to note here.

The fact that it is more token efficient will itself lead it to be smarter since the context will be smaller for the same task. However, in opus models, you'll have built internal correlations like "if it did X it will usually do Y" which may not be true here, since grok 4.5 may have done X purely due to the smaller context size, but can't do Y cos it wasn't RLd on that pattern enough. So it will be a unique experience as far as opus tier models go.

It would have been in a different tone had it been human written. In fact, if it had not been written with AI, there is no chance SVD gets called the fancy "engine behind image compression", the blog would have said exactly what you said in the comment above: "Its one of the simple forms of lossy compression, here's a simple before/after on a sample image" or whatever.

Yea, people have tried for decades to ban deepavali fireworks here in india with zero effect. It's simply not enforceable. The police themselves burst.

- coreutils for windows

- native container runtime

and others

After that "audio player that lays out the file in contiguous memory to avoid cpu jitter" example from above, this one is my new favourite audiophile trope. "cheap clock on the computer" I giggled.