> Break down sessions into separate clear, actionable tasks.
What this misses, of course, is that you can just have the agent do this too. Agent's are great at making project plans, especially if you give them a template to follow.
It sounds to me like the goal there is to spell out everything you don't want the agent to make assumptions about. If you let the agent make the plan, it'll still make those assumptions for you.
I do something very similar. I have an "outside expert" script I tell my agent to use as the reviewer. It only bothers me when neither it OR the expert can figure out what the heck it is I actually wanted.
In my case I have Gemini CLI, so I tell Gemini to use the little python script called gatekeeper.py to validate it's plan before each phase with Qwen, Kimi, or (if nothing else is getting good results) ChatGPT 5.2 Thinking. Qwen & Kimi are via fireworks.ai so it's much cheaper than ChatGPT. The agent is not allowed to start work until one of the "experts" approves it via gatekeeper. Similarly it can't mark a phase as complete until the gatekeeper approves the code as bug free and up to standards and passes all unit tests & linting.
Lately Kimi is good enough, but when it's really stuck it will sometimes bother ChatGPT. Seldom does it get all the way to the bottom of the pile and need my input. Usually it's when my instructions turned out to be vague.
I also have it use those larger thinking models for "expert consultation" when it's spent more than 100 turns on any problem and hasn't made progress by it's own estimation.
> This just... softens attentions ability to attend?
I think this does soften, but not linearly. That is to say the fixed state size limitation means that it softens more as it gets further into the past.
This uses the Taylor approximation to approximate softmax, but that IS only an approximation. I wonder exactly how much that trade-off costs in terms of accuracy vs performance? I note that they say it's close to float16 with four Taylor terms.
My other concern would be that Taylor itself is fairly complex. I wonder how well GPU's handle this in comparison to good old fashioned softmax? The last time I used Taylor with a custom Triton kernel it was still very slow. That could just have been my own jank vibe-coded implementation though.
I guess that what I'm saying is I'd love to see an LLM actually have it's attention mechanism replaced with this and get benchmarked on real world tasks in comparison to quadratic attention. They don't seem to have done that here. They claim that's it's close to being the same, but my experience tells me that it needs to do better than get "pretty close."
They also haven't' tried to write a high performance kernel for triton yet. If it goes the way my last experiment with Taylor did they're in for some bad news.
I'm just a hobbyist though, it's certainly possible that people with more time/resources could outperform me without much effort. I just want to see it tested on something familiar and benchmark-able.
> Who out there is working on really new unique ways to deal with infinite history, other than me of course :)
I'm working on a novel (I think) linear attention mechanism in my personal lab that's O(L) for effectively infinite context. I haven't yet decided how much of it is going to be open source, but I agree with you that it's important to figure this out.
Was your work open? Is there some place I can read more about it? I'm trying to figure out what to do with my thing on the off-chance that it actually does turn out to work the way I want it to.
I'm trying to figure the same thing out for my stuff. I figured out a simple way to train location prediction so I'm using it for guided window prediction which is great for attn (predict a distance in the past to look at) and for memory (predict an x, y location for a 2d window into a memory store to look at that will be helpful). I suspect there are a lot of people out there that have found that one weird trick but haven't released it because they don't know how to capitalize on the idea. Why give OpenAI and others the keys to the future for free?
"The History of every major Galactic Civilization tends to pass through three distinct and recognizable phases, those of Survival, Inquiry and Sophistication, otherwise known as the How, Why, and Where phases. For instance, the first phase is characterized by the question 'How can we eat?' the second by the question 'Why do we eat?' and the third by the question 'Where shall we have lunch?" - Douglas Adams
Folks on HN are very much in the "Where" stage of life. No one here works 4 out of 8 hours just to pay for their food. Nobody should.
That said, you very much seem to be missing the point. Ultra processed food is far, far cheaper than whole foods. That is one reason they are more popular.
For example, it would cost me more just to buy the ingredients to make tacos at home than it does to go through a Taco Bell drive through and buy enough for the family already prepared.
We're not going to be moving to four hour workdays by feeding people food that costs twice as much and takes longer to prepare.
My brother and his wife began cooking pretty much every meal at home a couple years ago. Prior to that they ate out very regularly, especially once they had kids.
They started cooking because feeding the family of 5 at McDonalds cost close to $80.
There may have been a time where fast food was cheaper, but it seems we're past that.
As far as Taco Bell goes, a single crunchy taco is $2.19 and their fancier ones are closer to $5. When I used to eat there I'd usually get 3 tacos and a drink, so I'd be into that today for something like $10-$11. I cook tacos at home regularly for cheaper, and with homemade tortillas and grass fed beef no less.
> They started cooking because feeding the family of 5 at McDonalds cost close to $80.
How much would they eat from McDonald’s? And what size appetite are the kids?
Fast food has definitely gone up in price, but if you’re spending $80 at McDonalds you’re either a glutton or you don’t know what to order.
A “Big Mac Bundle Box” is $15-20 depending on region. It has two Big Macs, two Cheeseburgers, two fries, and a 10-piece nuggets.
If three of the five are kids (vs say 16+ boys lifting weights), I’d be curious how two of those wouldn’t feed the entire family for $30-40.
I’m not suggesting cooking at home is a bad thing nor that eating McD is a good one. But the details matter when you’re spending 2x more than it could be.
Oh I'm sure some of the cost is because both my brother and their teenage son can eat some food. They're both in good shape, they just exercise quite a bit and have always had an appetite.
I also thought $80 for 5 was high, but that was his anecdotal number. I would have expected $50-60 pretty reasonably, and still st that point a family of 5 could eat for a good bit cheaper at home.
You are right, I stand corrected. It's been about 10 years since I last did the math and it's changed dramatically since then.
I'm sure it varies by region, but my grubhub app and the 12 pack of tacos (hard or soft) is $24.99 here so about the same as the $2.19 you found.
I had perplexity pro figure out the cost of purchasing the ingredients for comparable homemade tacos. With great value (Walmart store brand) ingredients it came to $20.04. $6.49 of that would be "left over" ingredients you don't use (mostly half a pound of beef you could use for something else later).
So you save $0.96 cents per taco by doing all the work yourself and using generic ingredients. Plus you get an extra half pound of beef for later.
So if your time is worth less than $12/hr it's a net gain.
I'm assuming it takes you only half an hour to travel, shop, and bring home the ingredients then half an hour to cook. If you live further away, factor in gas etc, the time it takes to do dishes, or are a slower cook then the break-even might come out closer to $6-$7/hr.
When we make tacos it takes around 30-45 minutes, including making fresh flour tortillas.
Tortillas themselves use very little, a cup of flour and a couple tablespoons of butter so maybe $1-$2? The beef we use is around $12/lb and we use 1/2lb to feed two of us. I don't have a cost on the seasoning, we mix it fresh as well so its negligible.
I'd assume we end up around $10 to feed two adults and spend around 45 minutes on the high end. We'd spend about that long to get to taco bell, though we live in a more rural area so that may be an over estimate for most.
> Ultra processed food is far, far cheaper than whole foods.
I think this is mostly true in the US and a cultural thing.
In EU and SA for example I can buy “whole” food - just called food here - for a fraction of the price it would cost me to buy a bunch of cheeseburgers or some other junk food every day.
Anyone who believes something like this you can be 100% sure doesn't know how to cook even the most basic of staple foods. Cooking your own food is nearly an order of magnitude cheaper and, with a few cheap spices and seasonings, almost always tastier. The only valid argument is prep time here, and that too even only applies to certain kinds of foods.
Premature optimization is a thing in life AND in programming. Many folks make it far more complicated than it needs to be.
I regularly see folks agonizing about every decision and new study, but the thing is.... the tips on OP's very basic list are responsible for like 80% of the value one gets from "living healthy".
All the rest of the organic whole grain horseshit and panicking about microplastics MIGHT net you another 10%, but at double the cost to your happiness.
The last 10% is basically impossible to achieve without completely sacrificing your quality of life.
The TLDR is that until version 8.8.7 of Notepad++, the developer used a self-signed certificate, which was available in the Github source code. The author enabled this by not following best practices.
The "good news" is that the attacks were very targeted and seemed to involve hands on keyboard attacks against folks in Asia.
Blaming the hosting company is kind of shady, as the author should own at least some level of the blame for this.
If the attackers did limit themselves to a small number of Asian machines they gave up an absolute goldmine. I would venture to say a lot of technical people use notepad++ at work in jobs that would be very lucrative for an attacker to exploit. I know I definitely had an 'oh shit' moment when I read this and thought about where I have notepad++ installed.
out of curiosity, why is a self signed cert bad for this case? Can't the updater check the validity of the cert just as well regardless? Or did the attackers get access to the signing key as well?
> Until version 8.8.7 of Notepad++, the developer used a self-signed certificate, which is available in the Github source code. This made it possible to create manipulated updates and push them onto victims, as binaries signed this way cause a warning „Unknown Publisher“
It also mentions "installing a root certificate". I suspect that it means that users who installed the root cert could check that a downloaded binary was legit but everyone else (i.e. the majority of users) were trained to blindly click through the warning.
Notepad++ has way too many updates for a text editor. I purposely decline most of the nags to update for precisely this reason. It is too juicy of a target and was bound to get compromised.
What this misses, of course, is that you can just have the agent do this too. Agent's are great at making project plans, especially if you give them a template to follow.
reply