The question of 'can it fly' is clearly a 'yes, given a little bit of effort'. Flying isn't hard, autopilots have been around a long time. It is recognizing and dealing with things you didn't anticipate that is hard. I think it is more interesting to have 99% of flying done with automated systems but have an LLM focus on recognizing unanticipated situations and recovering or mitigating them.
>I think it is more interesting to have 99% of flying done with automated systems but have an LLM focus on recognizing unanticipated situations and recovering or mitigating them.
Seeing how Claude (or any current LLM) perform in even the most low-stake coding scenario I dont think I would ever set foot on a plane where the 1% of most risky scenarios are decided by one.
That sounds like a solution looking for a problem though, i see plenty of arguments against throwing critical safety information that are in charge of peoples lives into an LLM "just in case the result is better than the result that the current battle-hardened systems already provide"
To properly test an LLM based emergency system against the current as-is system there needs to be a way of verifying whether the LLM detected emergency is classed as an emergency as-is. If this information was available publicaly it could enable bad actors things like stress-testing the EMP-tolerance of the current systems or what level of malware infiltration is detected.
General LLMs I would say are uniquely bad at this sort of thing.
I mean if you have a stable plane, then it'll do alright, as it'll mostly fly straight and level (assuming correct trim) reacting to turbulence however, the sampling rate would probably too slow, so you'd end up with oscillations.
For recognising that you're in a shit situation, yeah, it'll probably do that fine, but won't be able to give the correct control inputs at the right time.
>For recognising that you're in a shit situation, yeah, it'll probably do that fine, but won't be able to give the correct control inputs at the right time.
Even that im not sure of, I know relatively little about aviation safety but I can imagine that there are all kinds of 0.0000000001% percent corner cases that no plane has ever encountered that still need some sort of reaction, who knows how easy an llm can distinguish those from the 0.000000001% corner cases that no plane has ever encountered that are completely fine and can be ignored.
I agree with your intuition, There are lots of corner cases, but there are also a fucktonne of checklists: https://www.aviationhunt.com/boeing-737-normal-checklists/ (this is just a small "normal" one) but for loads of situations there are check lists, thats something the LLM can probably do very well.
However its as far as I know the check list volume scales with how "airline-y" the plane is. so for a one seater, the checklist is small and only handles a few things. For a 777 its a binder.
I've been watching the math of batteries and cargo ships and we may not be too far from shipping electrons generated in the Sahara to the UK and Europe at a reasonable price. That totally changes the game if you have cargo ships moving to where the power will be needed. I can imagine these ships going to where the weather is predicted to cause an issue to help even out the grid and just in general creating a responsive base load for the world. It sounds like sci-fi, but with the direction batteries have gone it isn't that crazy anymore.
Does it still work out if you take into account the insurance premiums for a cargo ship stacked with batteries? Can't imagine the fire hazard is pretty.
Oh, there's precedent for shuttling freighter size metal fire hazards intercontinentally to top up charge, is there?
How does that work out in cost per kWh? Profitable operation anywhere close?
Crisis relief (as suggested by jmward01 here) may be another matter, but setting up the ability to do this on scale, and maintaining it, can't be anything like easy economically.
We are at a moment where we are finding more and more ways to integrate solar in. It is likely we will go 'too far' in some ways but hopefully over the next few decades we will see a lot more well integrated solutions like vertical panels complementing farming and solar integrated, potentially with lower efficiency but also less impact, into things like building surfaces and other non-traditional places. Getting a diversity of options out there, and iterating on them, is key to the next phase where solar is everywhere reasonable by default and well integrated in to daily life.
I live in the UK in a town of 10,000 people, so say 4,000 houses (probably far higher than there are). If every house had a 10kWp (way more than most installs) that would be 40MW generation.
On the outskirts of town we have a 40MW solar farm about the same size as the golf course. Most people have no idea it's there, it uses barely any land compared to the rest of farmland around here. That generates about 40GWh a year.
The cost of renting the land it's on each year is about £20k a year, or 50p per MWh, basically nothing. Land is effectively free compared to the value from "farming the sun", it's far cheaper than the scaffolding to put 8kWp on a roof
I wonder if this has implications for custom home chips/prototyping. I'm sure a big issue is vibrations but something like this could remove the need for masks at least. (again, not my area so I am clobbering terminology I am sure). It may open up home fab capabilities.
I think half the fun for people that do things like this is figuring out how to out innovate a multi-billion dollar company so that they can make something 1/4th as good but at 1/10000th the price. I bet there are some -really- innovative people out there that would figure alternatives to a lot of the expensive parts of the process and figure out how to be able to produce 2000's level chips at home. I'm not one of them though :)
The problem is that they aren't (yet) 1/4th as good for 1/10000th of the price. Patterning is just one part of the process - and not nearly the most difficult one.
I think abusing a write-off electron microscope to side step the need for masks is also an interesting idea, however, I believe acquiring wafers of sufficient quality and depositing layers to be etched could be the bigger challenge here.
Hold on, if I had an electron microscope, can I just put in a decapped cheap large format photodiode under it, jack the beam current way up, and start etching trenches on it?
I don't think so: it's a microscope, not a synchrotron. :D
I meant "drawing" on a photoresist layer with a SEM and then wet-etching it. Also all silicon in a photodiode is doped, so the etched parts would be of little use, I believe.
I really want to get to the point that I am looking online for a GPU and Nvidia isn't the requirement. I think we are really close to there. Maybe we are there and my level of trust just needs to bump up.
Problem is, NVIDIA has so many quality of life features for developers. It's not easy getting especially smaller scale developers and academia to use other vendors that are 1) much more difficult to use while 2) also being slower and not as rich in features.
Personally I opted in to being NVIDIA-vendor-locked a couple of years ago because I just couldn't stand the insanely bonkers and pointless complexity of APIs like Vulkan. I used OpenGL before which supported all vendors, but because newer features weren't added to OpenGL I eventually had to make the switch.
I tried both Vulkan and CUDA, and after not getting shit done in Vulkan for a week I tried CUDA, and got the same stuff done in less than a day that I could not do in a whole week in Vulkan. At that moment I thought, screw it, I'm going to go NV-only now.
I did my thesis porting my supervisor's project from NeXTSTEP into Windows, was an OpenGL fanboy up to the whole Long Peaks disaster.
Additionally Vulkan has proven to be yet another extension mess (to the point now are actions try to steer it back on track), Khronos is like the C++ of API design, while expecting vendors to come up with the tools.
However, as great as CUDA, Metal and DirectX are to play around with, we might be stuck with Khronos APIs, if geopolitcs keep going as bad or worse, as they have been thus far.
The basic premise, try to be lean, is a good one. The implementation will clearly be debated with everyone having their own opinion on it but the core point is sound. I'd argue a different version of this though: keeping things lean forces simplicity and focus which is incredibly important early on. I have stepped into several startups and seen a mess of old/broken/I don't know what it does so leave it/etc etc. All of that, beyond the cost, slows you down because of the complexity. Regular gardening of your tech stack matters and has a lot of benefits.
Not really on the topic, but I have wondered if we need a different type of test to help find model architecture potential. Standardized training sets followed by testing to see the potential curves of a model. train on x, test, add y, test, add z, test. At each increment you see how well the model is absorbing the information and extrapolate how well that architecture may do if more fully trained.
"The U.S. needs a coherent protection and survivability strategy across the DoW and all sectors of our economy. This conversation needs to be not only about how we do it, but how we organize to do it, how we budget and pay for it and how we rapidly deploy it."
This is all predicated on creating thousands of drones which is a state actor level threat. The first line of defense at this level should be diplomacy. Digging tunnels and the like is unreasonable in peace time and likely not that effective in reality. Standing defenses become well planned targets. The real answer here is to spend the time and effort on diplomacy before there are issues and to stop appeasing countries like the US, Israel and Russia when they act badly. 'Special relationships' that are abused should be abandoned and trust should matter.
> This is all predicated on creating thousands of drones which is a state actor level threat.
You can do a tremendous amount of damage with off the shelf consumer drones, and a minimal budget. Ukraine did an billions of dollars of damage to Russia's airfleet with a couple million dollars of drones hidden in trucks. Well in the range of cartels and terrorist groups.
> The first line of defense at this level should be diplomacy.
You are very much correct that the way to not get into this situation is to not start a war.
It is an interesting take. I think this is mainly early adoption pains though. This stuff is moving so fast that if you say 'it isn't useful because X isn't good enough' then just wait a month and X will be good enough to find Y as the blocker (or no blockers are left and it truly does become useful). Soon we will see this hooked into the home assistant world well combined with local and remote compute and then we are likely to see real movement.
Conventional LLM's are moving fast too. The argument is that OpenClaw isn't any more useful than conventional LLM's, and I suspect it will always be true because the conventional LLM's will gain any useful capabilities.
I think openclaw provides a unique feature of a standardized host environment for a persistent assistant. This is different than the chat interfaces that are presented by anthropic/openai/others that give you a 'while you are here' assistant interface and is very different from the idea of trained llm weights and ways of serving them up like llama.cpp and others. There really is something unique here that will evolve over time I think.
Agree. It is like just having one continuous chat session with ChapGPT forever. Of course they do have memory already (at least ChatGPT does). I ended up turning it off however, because it kept bringing irrelevant stuff into the current convo.
People like to think the pendulum will always swing back. That is because of survivor bias. They have always seen it swing back. Every fallen civilization believed in the pendulum theory too, until the last one. You can't magically remake our forests. We are just stupid.
This is kind of one of the points I try to make. All of the damage that is being done right now is very hard to fix when we have sane people back in power (if that happens). It’s 100x easier to destroy than build, and we are seeing it happen in real time.
Also, attempts to undo the damage will be painted as extreme, whether it’s rehiring / recreating the agencies that were destroyed or prosecuting the crazy corruption and seizing funds.
Not only that, but when you do fix it, voters will forget how bad things were last term and re-elect the people that caused the damage. It is extraordinarily rare in modern times for the party in power to not change when the incumbent president isn't running. It's also very rare for a party that controls the presidency to keep the house/senate after midterms.
It's easier to deny bad things than it is to face problems with only bad solutions. And anyone who tries to point out the reality will be called a defeatist.
I have had many discussions over the decades, presenting hypotheticals with no good answer--and I have found very few people who would even try to address the scenario as presented. They're always "misunderstanding" it in some fashion which allows a good outcome--and such "misunderstanding" can never be corrected because they are unwilling to contemplate the hard choice even when the choice is the best for all parties. (The assault team has taken the terrorist base--but there's a little kid reaching for his bottle that's holding down the deadman on the nuke. Nobody can reach him in time, do you shoot?)
reply