looks like they did? https://capybara.info/

funny, I like it for the opposite reason. If I am buying a lot of groceries and want to pack them into my shopping trolley nicely, but there’s a line of 10 people behind me at the cash register. So I need to rush to put everything in as the cashier will just put more and more stuff on the conveyor to the point where other peoples groceries end up mixed with my stuff, because there’s only two lanes and the person right behind me only had 5 things, so the cashier is already handling customer 3 while I’m still packing.

With self checkout, I can decide the speed I want to pack at, make sure I don’t put fragile stuff at the bottom, can take it nice and relaxed. there’s 6 other terminals, so it doesn’t feel like I’m the one blocking the line. Just an overall less stressful experience.

also Gergely Orosz/pragmatic engineer has a pro AI bias and got VERY focused on AI recently, to the point where like 90% of posts are AI focused.

I hate this timeline

I think you misunderstood, i thought it funny that meta is putting expensive GPUs into what amounts to chicken coops. Not that the article was wrong and there’s actual chickens inside

Are these just industrial chicken coops?

You know, this kind of thing:

This is one of the big issues i have in my job. I’m on a dev team but a majority of the people around here are research data scientists. And a shitty 1000 line python script with zero regard for quality, that only really produces a bar chart as output and is never used again after 6 months is all they ever work with. And LLMs are amazing for that. But these data scientists don’t seem to understand that not all software is like that. So it’s really hard to have a real conversation about it.

And wouldn’t you know it, these are some of the same kind of people building this stuff and pushing the vibe coding narrative.

Even for that, 4 legs like the robot dogs is easier/more stable. Plus purpose built robots that don’t need to move are generally cheaper and make more sense, like a dishwasher.

and ten 500 bucks purpose built robots are cheaper than one 20-50k humanoid robot that does tasks slowly and needs a human remote operator for most of them

That’s true of most of the tech bros, though. They’re all losers that haven’t built anything themselves and that have been successful despite themselves, not because of themselves, just because they have money and fail upwards

It can be done but doesn’t work that well: https://www.youtube.com/watch?v=0rliFQ0qyAM

Actually anyone can post using the API, and many people did. Lots of the heavily publicized posts were written by humans

Yes. And that requires the participants to self-regulate and be educated, at least beyond a certain number of participants (and this number is more in the 100’s than 1000’s of devices in a city size area).

In case anyone is interested, in meshtastic, for a regular handheld device, set it to client if in a less populated area, client_mute in densely populated areas. router/repeater only for really well positioned nodes (think top of a mountain, NOT top of a tall building), router_late for ok positioned nodes where there’s other, better router nodes around, but they can cover some dead zone that the better router ones don’t cover. And then there’s other like client_base for e.g. a rooftop node to help handheld devices inside the house reach outside. And follow the guidelines of your local area group/forum/whatever on number of hops, band to use etc.

None of that is really complicated, but a lot of users will just buy a device, turn it on excitedly with default settings, see that messaging works and not dig deeper, and then worsen then network for everyone.

With meshcore, there’s a repeater firmware and a client firmware, it’s not just a setting that can be selected in the app. Default is client, and switching to repeater mean reflashing the firmware. And more clients don’t really matter for the network, since they don’t repeat. So much less risk of an unwitting user worsening the network.

And if anyone wants to play with this, all the devices should support both meshtastic and meshcore firmware, so you can play with both and see which you prefer.

And if you’re curious why I prefer meshcore, these are all the nodes I was able to hear on meshtastic and meshcore, respectively, in the same location, with meshcore I can reach twice as far in each direction

I’m definitely not an expert radio operator or anything, but this is how I understood it.

With Meshtastic, no, it actually gets worse beyond a certain point, especially in dense areas. Each radio rebroadcasts messages, so more message fill the air, causing more congestion. setting devices to client_mute (no rebroadcasting of messages) alleviates this to some extent, but most people don’t do that. In general, a few well placed nodes are better than a lot of clients.

Meshcore with dedicated Repeater nodes alleviates that a bit, as a node is either a client that can originate messages, or a Repeater that can only rebroadcast. And it has a bit better routing protocol, reducing unnecessary duplicates in some cases.

The radio band is license free in many countries, but there’s still some restriction, e.g. 10% duty cycle in my country, which means no radio can be sending more than 10% of the time (so 1 minute in every 10 minute window). That combined with (afaik) nodes only being able so send one message at a time limits how many messages can be in flight at an instant. And more clients increases background chatter from e.g. node announcements and duplicates, leading to further congestion. And in meshtastic, since packets “flood” out in all directions, and each node rebroadcasts again, and it’s not directional, all the packets travel all over the place when you send a message. You could have a packet traveling in circles between new nodes for a long time.

E.g. in my region, originally the default for meshtastic was 7 hops on long range fast preset. That led to congestion so they switched to medium range fast, then recommended limiting hops to 5, and now ideally to 3, just so there’s less chatter and less utilisation of the channel. 3 hops is barely enough for me to get out of the city I live in. Meshcore also does floods, but also can keep track of paths so you can have a relayed connection along a specific path that does not flood out in all directions, along with putting up repeaters being a conscious and strategic choice by users, so you get something more like a dedicated highway of repeaters that are placed high up and can relay messages much better than some client down at street level could. And the hop limit is more like 64 (I think it can be higher but that’s what is set somewhere in the firmware). And repeaters are stationary, so they don’t need to announce themselves to the network that often (e.g. every 3 hours in meshtastic vs every 24-48 hours in mesh core). All of that means less channel utilization and longer reach.

I agree, that would be amazing. I also hope it will help with some truly local community building (no troll farms from halfway across the planet spamming shit). Weather stations are already possible with sensor nodes, and most big repeaters have weather data. Though not like weather forecasts or anything.

The main issue would probably just be congestion, not even bandwidth. Once it’s used a lot, some packets will just be dropped due to congestion and you don’t get a reply at all.

A bit less of a problem with meshcore, with meshtastic in densely populated areas most users still don’t set their devices to client_mute, causing unnecessary rebroadcasts and even more congestion. Though with enough adoptions maybe governments might lower their restrictions on duty cycle, allowing for more traffic.

Meshcore might be a bit better suited for this, if you want to reach a forum further than 50-100km away reliably.

With the room servers it almost supports this use case already

I’m not talking about the quality of LLMs (they suck, in so many different ways…).

I’m criticizing the experiment setup, it is not really statistically sound. Doing 10 tests each with 52 different models is almost bound to have one model be correct 100% of the time (even if the true probability is closer to 50%), by pure chance. Doing 100 tests each might yield very different results with none of them answering correct 100% of the time. Or put another way, the p-values of the tests performed are pretty high, not <0.05, so the results don’t really say what they purport to say.

10 tests per model seems like way too little and they should give confidence intervals…

the 10/10 vs. 8/10 is just as likely due chance than any real difference. But some people will definitely use this to justify model choice.

I think >50% of my up votes are me swiping right to go back but missing the screen border by a tiny bit, causing the jerboa swipe to vote to kick in.

so who knows, maybe they just accidentally hit the downvote arrow while scrolling?

I’d prefer 1.0 school, but hopefully they’ll work out most the kinks in the beta version already. Alpha should just be for internal testing, aka the CEOs kids…

Plus firefox mostly exists because google pays them, probably so there’s no anti-trust action against them.