I’ve never heard about modern people doing serious persistence hunting, except for a stunt that I read about years ago. I think it was organized by like Outside or some running publication that got pro marathoners to try and they failed because they didn’t know anything about hunting

All of which I have done is legal, I do not hunt native animals, only introduced species. I apologise for anyone who may find the following details grotesque, the damage that these animals did was often quite nasty.

> Where were you doing this?

Central Queensland, Australia. I had to get permission from farmers and national parks and wildlife if I was to go on their property. I started this when i was around 5 years old , doing walk hunting which is just the same thing but for a full day. I think this prepared me as a child to 'long distance' the tracking. Knowing what tracks looks like, mud and fur on trees, how animals traverse rivers, where to start and stop looking for tracks so that I don't waste time looking at the wrong spots.

I have also done this on properties in Daintree, in far north Queensland (tropical). I found rainforest hunting much harder because finding tracks was a challenge and I had to spend time worrying about crocodiles and snakes and poisonous trees, being prey myself.

I might be doxing myself, however I don't think many of the kids I went to school with end up on HN, but I've been wrong before, if you know who I am, please stfu.

> Were you ever successful?

Regularly, almost every time that I found tracks I was able to catch at least 1-2 animals, largest take down was about 13 animals, I would say less than 10 times over the course of 5 years I came home with nothing. I would hunt almost every other weekend.

I followed the steps taught by my father, who is a australian bushman who seems to know every tree and animal, can see and hear animals hiding in the bush that I can only see after trying to look for 10 minutes, my mother is equally as good in the bush but with less hunting and more capable when it comes to the people side of things.

> How did you do it.

I hunted with my brother and father and sometimes mother.

Basic equipment: - Knife - Arm guard(s) - Water - Dried meat - Backpack - Matches and lighter. - Tourniquet - First aid kit (not always)

I did use a modern knife, I don't know if that is cheating or not, but I feel like strangling or bashing an animal to death was a bit cruel. I have hunted with dogs a few of times, but you can't bring them on national park land so this limits the success.

I sometimes wore leather guards (leather vambrace ? made by my mother) on my arms and ensured i had some kind of leather scarf around my neck because cats get scratchy, dogs get bitey and pigs will try to gore you with tusks. I once used a kickboxer arm guard but it had holes and I was bitten through the holes, so not doing that again.

I did not wear shoes when hunting in the central Queensland, it seemed safe enough and I didn't impale myself too often, feet adapt.

I mostly hunted pigs, dogs (not dingos), large cats, or deer. I have successfully only caught deer less than handful of times.

> like what were your tactics?

This is the 'ideal' situation, it doesn't always work this way but it's what the goal is, one needs to adapt to the changes as they happen.

Tracking phase:

Walk an area that had prints, track the prints, follow the freshest ones. These paths you can use later, because animals will frequently go back to an area they know if you lose them.

Usually the best place to start tracking is around crops and other large animals, pigs and cats will separate the young offspring from the group and kill them for food.

By paying attention, you can get a good idea of their behavior, the animals will repeat successful behavior that gives them food and water.

This usually, but not always means that they will be going for water at dusk and dawn. The first step is denying them that water, wait near the place they get water.

Hunting phase:

GOLDEN RULE: NEVER UNSHEATH THE KNIFE UNTIL YOU ARE CLOSE ENOUGH TO GET THE KILL. (I have had friends come hunting and cut themselves slipping down an embankment with the knife drawn, infuriating!)

SILVER RULE: DRINK WATER, ALWAYS HAVE ENOUGH FRESH WATER SOMEWHERE.

BRONZE RULE: If you get lost, do NOT just start wondering, you idiot, relax, don't panic and listen, drink some water, look for smoke and light.

You (the first) will need to be there before they get there, so this often means being there well before the sun rises. Stay downwind so that the animal doesn't smell you and not come to the water.

Ideally you want to be running them east so they are looking into the rising sun, not the biggest deal but if you have this option, take it. next best option is to have them running 'on the plain' , aka not in trees, this allows you to track them by sight.

If it is a herd animals (aka, everything in my list but cats), a small group will typically test the area first and the full group will join them when they consider it safe.

Wait till the full group appears, hopefully the sun is up enough that you will be able to see prints left behind.

You want there to be enough light that you can see the tracks, so sometimes this means letting them 'start to drink' before you begin the hunt.

You have to make yourself seen as the biggest threat possible, make noise, appear large, use a torch to make light, sometimes you can sneak up close enough and get the first kill by hand and then make a lot of noise (sneaky sneaky!) . This may mean shoulder checking the bigger animals, diving on them or booting them to get them moving.

Chase them (direct them if possible) towards the person hidden ~ 1000m away, be as aggressive as possible in the movement to keep the moving quickly. This -will- tire you out but the second person will continue giving chase.

Second must also be as 'aggressive' as possible while trying to keep the main group together.

The animals will USUALLY split, this is very common, but you need to make an educated decision on which animal to pursue. I've had most luck with the males (more reasons to follow).

You can usually hear the animals (and the human making noise) and catch up within 15/20 minutes at a moderate pace. We have 'woops' and 'aaahs' sounds which travel well and are clear over distance to signify left and right directions (if the animals are ahead / too quick). I believe that this is an older aboriginal hunting trick in some tribes.

Because the first person is trailing, it is usually a shorter path to catch up. Once caught up, the second take take a breather, give them some water and you keep going.

By this point you are usually at the 5.0/6.0 km mark. Your first animal will either collapse or stop in its tracks (easy kill). Some smarter animals will attempt to fight you when they realise they are tired.

If you find a log nearby, (some people carry a short staff/walking stick) you can hit it/throw it to spur it onwards, into further tiredness. You want them to be so tired they can't put up a fight. Pigs will often do this in an attempt to allow the sonder/mob to gain extra distance while you deal with it, but ideally you want to keep them moving together.

Large cats also will defend their group this way, usually this will give them time to get up a tree or hide, but you must keep on them so they don't have that luxury. Its quite hard to kill a large cat this way because they will try to claw you or get on top of you and attack you, bad times.

If you are lucky enough to have a third person, you hopefully position your third person in high ground nearby, so that whatever direction they need to go, the travel is downhill (it's easier, better visibility and you can adapt to changes when things change).

This should usually be the older person (or the newest) because they have the animals tired and need to continue to present a threat, but not too long.

By this time the animals will be out in 'new territory' and this is where risk happens, they will no longer be following the usual routes they know and can act erratically.

Here is where the persistence is, you kinda need to 'rotate' the front chaser, have someone who is has the energy to chase the main group, and the trail people shepard anyone who breaks free back to the main group. This can be anywhere from 10-20km. Keep someone at the front, continually giving chase, this person should make tracks as CLEAR as possible, footprints on mud and dirt, leaving arrows in the dirt, pointing to the direction they are going (when tracking humans foot position is NOT always the direction they are going).

This is where most injuries happen in people, do not ignore them. Rotate the people at the front, leave people behind that are too tired, hurt (but not dying), tell them to start a small fire, hydrate, and rest, do NOT keep going.

If animal group has any offspring, they usually can't walk by now and mum won't leave them alone, you need to make the decision on if you kill the mum + kids now (usually the first and second can do this) and the group continues on.

After a few rotations of slow jogging, backtracking and tracking, you will find even the most hardened of animals has tried to find cover/hide.

The cats go up trees, the pigs try to hide in logs/brush/wherever, deer will go to a thicket/grass and crouch and hide. You can usually just meet them where they are, let them try to take the first blow (on your armored parts) then go in for the kill. Most times though they are simply too tired, panting on the ground in a state of fatigue.

Always aim for a one cut kill, go for the jugular and be sure, it is cruel to have to go for two, sometimes you must, but don't aim for it.

Since the group has split, circle back.. the footprints will be very easy to follow now because they will be deeper (since they are panicked) fresher (because today), and if there was a sow with piglets, they will be noisy. Usually they will not be resting not too far from the group. The sun will be about 10/11AM and they'll be hot, tired and thirsty.

If you can't find any, wait downwind for a few hours and see if they come back to the water at dusk, if they haven't had any water they will be very thirsty and this chase usually won't go on for more than about 1.5km.

Dragging carcasses back to civilisation can be a real pain in the arse, if you know there are other feral animals in the area, bury it deep or burn it. The fire can be a good method of finding/giving directions your hunting mates if you are lost or new.

Most of the pigs had some kind of worms (only good after being cooked, and then only for dogs), so the meat was not so great. The dear meat was 'passable' for jerky and cats and dogs were not worth the trouble.

I respect the marathon runners for even trying, its very different as you need to both plan, think, run and pace yourself. The whole hunting isn't a sprint, its a slow methodical paced plan, I have many fond memories of spending time with my family in the Australian bush, hunting and camping.

I do find it interesting that they state that each task is done with a fine tuned model. I wonder if that’s a limitation of the current data set their foundation model is trained on (which is what I think they’re suggesting in the post) or if it reflects something more fundamental about robotics tasks. It does remind me of a few years ago in LLMs when fine tuning was more prevalent. I don’t follow LLM training methodology closely but my impression was that the bulk of recent improvements have come from better RL post training and inference time reasoning.

Obviously they’re pursuing RL and I’m not sure spending more tokens at inference would even help for fine manipulation like this, notwithstanding the latency problems with that.

So, maybe the need for fine tuning goes away with a better foundation model like they’re suggesting? I hope this doesn’t point towards more fundamental limitations on robotics learning with the current VLA foundation model architectures

But it seems like a degree of "RL in real life" is nigh-inevitable - imitation learning only gets you this far. Kind of like RLVR is nigh-inevitable for high LLM performance on agentic tasks, and for many of the same reasons.

Re. not expecting it for ten years at least, current progress is pretty much in line with Moravec's predictions from 35 years ago. (https://jetpress.org/volume1/moravec.htm)

I wonder if he still follows this stuff?

What fascinates me is we could probably make self-folding clothes. We also already have non wrinkle clothes where folding is minimally needed. I wager we could go a lot further if we invested a tad more into the matter.

But the first image people seem to have of super advanced multi-thousand dollar robots is still folding the laundry.

Its both. A colorblind person will admit their shortcomings and, if compelled to be helpful like an LLM is, will reason their way to finding a solution that works around their limitations.

But as LLMs lack a way to reason, you get nonsense instead.

This assumes the colorblind person both believes it is true that they are colorblind, in a world where that can be verified, and possesses tools to overcome these limitations.

You have to be much more clever to 'see' an atom before the invention of a microscope, if the tool doesn't exist: most of the time you are SOL.

This makes me wonder if LLMs works better in Chinese.

It's an example of a simple task. How often are you relying on LLMs to complete simple tasks?

https://chatgpt.com/share/6941df90-789c-8005-8783-6e1c76cdfc...

This isn’t totally true, group/kin selection are important.

Regarding LLM’s I think METR is a decent metric. However you have to consider the cost of achieving each additional hour or day of task horizon. I’m open to correction here, but I would bet that the cost curves are more exponential than the improvement curves. That would be fundamentally unsustainable and point to a limitation of LLM training/architecture for reasoning and world modeling.

Basically I think the focus on recursive self improvement is not really important in the real world. The actual question is how long and how expensive the learning process is. I think the answer is that it will be long and expensive, just like our current world. No doubt having many more intelligent agents will help speed up parts of the loop but there are physical constraints you can’t get past no matter how smart you are.

At some point learning can occur with "self-play", and I believe this is already happening with LLMs to some extent. Then you're not limited by imitating human-made data.

If learning something like software development or mathematical proofs, it is easier to verify whether a solution is correct than to come up with the solution in the first place, many domains are like this. Anything like that is amenable to learning on synthetic data or self-play like AlphaGo did.

I can understand that people who think of LLMs as human-imitation machines, limited to training on human-made data, would think they'd be capped at human-level intelligence. However I don't think that's the case, and we have at least one example of superhuman AI in one domain (Go) showing this.

Regarding cost, I'd have to look into it, but I'm under the impression costs have been up and down over time as models have grown but there have also been efficiency improvements.

I think I'd hazard a guess that end-user costs have not grown exponentially like time horizon capabilities, even though investment in training probably has. Though that's tricky to reason about because training costs are amortised and it's not obvious whether end user costs are at a loss or what profit margin for any given model.

On the fast-slow takeoff - Yud does seem to beleive in a fast takeoff yes, but it's also one of the the oldest disagreements in rationality circles, on which he disagreed with his main co-blogger on the orignal rationalist blog, Overcoming Bias, some discussion of this and more recent disagreements here [1].

[1] https://www.astralcodexten.com/p/yudkowsky-contra-christiano...

That said, nearly every other task in the world is not easily verified, including things we really care about. How do you know if an AI is superhuman at designing fusion reactors? The most important step there is building a fusion reactor.

I think a better reference point than AlphaGo is AlphaFold. Deepmind found some really clever algorithmic improvements, but they didn’t know whether they actually worked until the CASP competition. CASP evaluated their model on new Xray crystal structures of proteins. Needless to say getting Xray protein structures is a difficult and complex process. Also, they trained AlphaFold on thousands of existing structures that were accumulated over decades and required millenia of graduate-student-hours hours to find. It’s worth noting that we have very good theories for all the basic physics underlying protein folding but none of the physics based methods work. We had to rely on painstakingly collected data to learn the emergent phenomena that govern folding. I suspect that this will be the case for many other tasks.

Go is entirely unlike reality in that the rules are fully known and it can be perfectly simulated by a computer. AlphaGo worked because it could run millions of tests in a short time frame, because it is all simulated. It doesn't seem to answer the question of how an AI improves its general intelligence without real-world interaction and data gathering at all. If anything it points to the importance of doing many experiments and gathering data - and this becomes a bottleneck when you can't simply make the experiment run faster, because the experiment is limited by physics.

[1] https://www.lesswrong.com/posts/8viKzSrYhb6EFk6wg/why-yudkow...

I'm reminded of a silly social science article I read, quite a long time ago. It suggested that physicists only like to study condensed matter crystals because physics is a male-dominated field, and crystals are hard rocks, and, um ... men like to think about their rock-hard penises, I guess. Now, this hypothesis obviously does not survive cursory inspection - if we're gendering natural phenomena studied by physicists, are waves male? Are fluid dynamics male?

However, Mr. Yudowsky's weird hangups here around rigidity and hardness have me adjusting my priors.