Incorrect. Venus experiences ridiculously violent storms

No. I literally wrote a book on the topic decades ago. Based on what we knew then (which again comes with massive caveats due to our low funding of Venus missions), the view was that Venus's storms were almost the same as those on Earth in intensity. I've done some reading on the results since then, and from an incomplete read of more recent literature, the view today is that Venus winds are are actually significantly tamer than what was known then. More recent data says that normal vertical turbulence is about 1-3 m/s and only peaks at 5-10m/s in extreme cases. By contrast, on Earth thunderstorms regularly have 10-30 m/s updrafts, and supercells can hit 50m/s. Venus simply has no large CAPE release like on Earth. No cumulonimbus, no supercells or squall lines. Also, back when I was doing my research for the book, lightning on Venus was thought to probably roughly as common on Earth. But the Parker Solar Probe data says that the whistler waves that were previously interpreted as lightning are actually probably magnetic reconnection in the outer atmosphere, not lightning in clouds. Based on this, lightning in the clouds of Venus is probably rare to nonexistant.

For something dozens to hundreds of meters in size?

You have your scaling factors wrong. Scaling up gives you slightly better volume- and mass-to-lift ratios for a given amount of structural strength, which you can allocate toward greater strength. In an naive situation, there's no change with scaleup, because your envelope's tensile strength (in the non-rigid case) has to increase linearly as radius increases linearly, the area doubles and the volume triples, so you have a r-cubed relationship with both envelope mass and volume. But in practice it doesn't work that way - for example, protective coatings are relatively fixed thickness regardless of scale. And there's many things beyond just the envelope, and again, they generally benefit with scale. So you save mass as you scale up, and if your main concern is resilience, you can allocate your saved mass towards increasing that.

An aerostat as proposed in Venus will need to contend with hurricane force winds. Not hurricane velocity- hurricane force.

The velocity of the winds relative to some surface 50-55km below you is utterly irrelevant. Expecting that to be relevant is like expecting airplanes to fear flying in the jet stream because "it's so fast!".

The only thing of relevance is turbulence, e.g. local variations, which are not "hurricane speed" - just normal convective cells. Which in Venus is Earthlike, aka not particularly problematic for an aerostat so long as there's no risk of accidental ground collision.

Who the hell said anything about accidents?

If you're not positing accidents, then what on Earth are you positing as the problem for long-duration flight? Gas permeability through the envelope? That's readily calculable, and FYI, it's not a problem relative to the rates you can regenerate O2/N2 from the atmosphere.

Venus's atmosphere is no more violent than Earth's; the properties of the middle cloud layer is surprisingly similar to our own. Most accidents involving lighter-than-air craft on Earth have involved collisions with things on the ground in some form or another, which is not applicable on Venus. It's ground handling that has always been the difficult, dangerous part with them. We have flown aerostats for around a year (see for example Aerostar Thunderhead) and there's no reason we can't do more.

Lighter-than-air aircraft don't react to turbulence in the same way that heavier-than-air aircraft do. The latter shudder and sharply pivot in an often dangerous manner. The former act more like a large ocean liner in rough seas - slowly rising and falling with the waves, with a long period.

Resources, the only accessible resources on Venus is the atmosphere, mostly CO2, some nitrogen and I don't think any amount of hydrogen

You are describing (some of) the inert species in the atmosphere (also there's also noble gases and carbon monoxide). However, the aerosols (and possibly rains/frosts/snows, but we don't know because we've spent so little focus on studying Venus in contrast to Mars) are mainly water and sulfuric acid. There's also phosphoric acid and hydrogen chloride (the lower cloud layer is probably mainly phosphoric acid, not sulfuric, but the ratio decreases with altitude and sulfuric dominates in the middle cloud layer and above). The hydrogen chloride is mainly as a gas rather than hydrochloric acid. There's also hydrogen fluoride, though only quite small amounts, and there's also probably some iron chloride.

All of these chemicals are highly hygroscopic (and the particulates electrostatic-attracted) and can be recovered in scrubbers, akin to some types of those used in industrial pollution control systems (indeed, your scrubbers could double as your propulsion system, using the through-air for propulsion). Recovered aerosols simply need to be heated; first, you drive off water, with further heating splitting H2SO4 to more H2O and SO3. The SO3 can then either heated over a vanadium pentoxide catalyist to break it down to SO2 + O2, or alternatively you can inject the SO3 back into your scrubber as a conditioning agent to help precipitate more hygroscopic gases. Scrubbing resources is easier and more consistent than hard-rock mining, which suffers from constant wear and breakdowns and deals with resources that have constantly varying compositions and generally more complex refining processes.

You have basically everything you need for a plastics industry and life support in the atmosphere, as well as for the anions you need for fertilizer (cations can be recovered by incinerating organic waste to oxides / hydroxides and reacting said ash with the acids made of the necessary cations).

Furthermore, contrary to popular myth, the surface isn't that inaccessible (Soviet tech developed in the 1960s managed it - Venus temperatures and pressures really aren't particularly extreme from an industrial perspective, and keeping the inside cool can be as simple as mere thermal mass or phase change materials, with stays up to days using battery-powered heat pumps, and high-temp RTGs (with albeit low efficiency) powering heat pumps allowing indefinite stay). The surface actually has a lot of advantages relative to Mars. Landing is easy due to the dense atmosphere - to the point that one Venus probe accidentally lost its parachute during descent, and still landed intact and kept broadcasting. The atmosphere is thick enough that you should be able to easily dredge loose material. You can also do controlled flight around the surface with a metal bellows balloon. You don't need a rocket to return to altitude, just a phase-change balloon. And perhaps most importantly, you're not bound to a single location. Resources on a planet aren't all found in the same location; the ability to target locations all over the planet (or at least within a couple dozen degree band on your hemisphere) is massively advantageous. Venus also has a lot of unique natural mineral enrichment processes, up to and including apparently chemical vapor depositing semiconductor crystals in its highlands.

Venus also has a natural export economy: beyond any potential surface resources of sufficient value density (for example, CVD galena is used in IR detectors), Venus is extremely enriched in deuterium (though probably not to the degree of being a health hazard), due to the hydrogen escape after it lost its oceans. If a floating habitat stores its energy via a reversible fuel cell (fuel cell + electrolysis), and is wired into a cascade, you can automatically enrich the deuterium every day at no extra cost to you (both fuel cell and electrolysis operation have very high enrichment factors). At ~$1k/kg, with significant growth potential if fusion takes off on Earth, it's a viable export commodity if launch costs fall enough (e.g. a couple dozen $/kg at Earth -> perhaps a couple hundred $/kg for Venus round trips). You will however need an efficient launch mechanism on Venus (NTR is an ideal candidate, as it's very hydrogen-efficient, allows for a SSTO, boosted designs give you a high thrust-to-weight ratio for the early launch phases, and some hybrid airbreathing designs can allow for indefinite hover for docking)

Iron, oxygen and water are not hard things to acquire at the moon's poles. It's like you were trying to make a list of the things that are easiest to acquire on the moon.

Your biggest problems will be the extreme paucity of both nitrogen and carbon, things essential en masse for all life. Beyond that, chlorine and fluorine are also very rare, zinc is about 2 orders of magnitude less common than on Earth (also lead, bismuth, thallium and cadmium), etc. Also, beyond general abundances, is the lack of many of the sort of enrichment processes that create rich mining deposits on Earth. At best you'll get some of the volcanic enrichment processes (incompatible elements in pegmatites), but not much beyond that, and even then you're going to deal with lots of overburden. And hard rock mining and processing on the moon will be far more difficult than on Earth.

But iron, water, and oxygen are basically the easiest things you could get on the moon. Respectively, half a percent of regolith is (magnetic) iron dust; water, while rare globally, is seemingly abundant in polar craters; and oxygen can not only be made from water, but over 40% of the mass of lunar regolith itself is oxygen, which can be freed via a variety of (albeit energy-intensive) processes.

(Even "getting the minerals" isn't really the big challenge anyway in gaining full independence from Earth. It's the mind-bogglingly immense length of production chains needed to fully sustain even a minimized-set of required technologies ("consumables", both feedstocks and maintenance), and all of the transport along the way. You can whittle down how much you need to import per-capita by orders of magnitude, but getting rid of all of it is a big ask)

There actually is metal in lunar regolith, BTW. About half a percent of it by mass is metallic iron. One ISRU proposal for mining on the moon is basically "drive along and throw regolith across magnets, and melt down whatever sticks" ;)

Try running the energy calculations on how much it'll take to pave the entire landscape around you.

(Also, lunar dust is mobile. No matter how far you pave out, it'll come back. It gets ionized by solar radiation, lifts off, and drifts significant distances)

On the other hand, if 1/6th gravity does prove enough, that'll answer the big question hanging over whether Mars's 0,4g is enough.

If it doesn't, then it's still in the who-knows category.

High resolution radar is great and really should be in any autonomous car*, but I disagree about conventional radar. The resolution is so low that at any reasonable distance ahead of your car, it just can't tell if something is in your lane or not.

* - I like how it sees the world in a different manner than how your cameras do, so it's an entirely independent datastream not subject to the same limitations as your cameras, unlike LIDAR which has the same limitations, only worse. Metal shines brightly. People look like ghosts. Cars cast reflections on the road that you can see even when you can't see the car itself. You can see the smoothness of the road or any other object (and help determine the material) based on its reflectivity, on a scale relative to your wavelength (which you can change). Weather and dirt are only minor hindrances. Etc. In theory, if very well calibrated, you could do interferometry, with one radar unit on one side of the car and the other on the other side, and have an effective aperture the width of the car, for quite good resolution indeed.

Rather than LIDAR, IMHO time-of-flight cameras are more interesting than LIDAR to me - doubling as both visual *and* depth data sources. You could theoretically use custom headlights and brake lights as the emitters of the sub-50ns pulses, so again, more hardware reuse - and can get depth data from many different angles.

Either way, though, I am in agreement that "the more sensor data, the better". Yes, humans (mainly) navigate the world through just binocular vision, but we also have the advantage of.... a human brain. When your software is inferior, you need to give it a leg up. Also, while he seems to understand better now, for a long time Musk seemed to view AI as just "to solve any problem, you just need to increase the size and training of a basic DNN", entirely ignoring the critical importance of architecture.

BTW, the FSD tracker recently had a big leap up in distance per safety disengagement. That said, it's too little data thusfar for me to take it at face value, and it's also still well below humans, even when you factor in that only a fraction of said safety disengagements would have actually ended in an accident.

Um... I have issues with a lot that Elon says, but those are some pretty weak arguments.

Yes, we can, with today's tech, get people to Mars and back without them dying of radiation poisoning. Not significantly increasing their lifetime risk of cancer is the challenge, not the question of whether you can keep them alive. And the solutions aren't at all magical - you coast with your fuel and engines between your crew and the sun, and you ideally also have small shelters for solar storms. GCR can't be readily blocked, but it's also much lower intensity. The overall radiation load doesn't kill you, it just increases your lifetime risk of cancer and other diseases.

You do not decelerate payloads with propellant. You decelerate them with aerobraking / aerocapture. And there is nothing unusual about the dV schedule for Starship.

"Spending 2.5 years on a mission means that Earth's gravity will probably kill you if you make it back" - we've had enough long-duration spaceflights that we can confidently state that this is demonstrably not what happens (and also it's not about microgravity for 2,5 years, it's 2x sub-8mo trips with 0,4g in-between), and also, you can rotate spacecraft to create "artificial gravity" through centrifugal force. Though in general that's unlikely to happen, as they'd be more likely to choose "engineering simplicity" at the cost of "more adjustment time on the ground". Once again, you're confusing "bad for your health" with "acutely kills you".

"Air. The Earth has relatively lots of it. Mars has very little of it, less than 1% of Earth's" - and generating oxygen from water is one of the most trivial of industrial processes that humans carry out, and long-term-durable reusable CO2 scrubbing can be done with something as trivial as periodically heating washing soda and a fan (potassium carbonate/bicarbonate swing absorption).

Your citing all of these ridiculous "problems" and not focusing on any of the actual engineering that they're skimming over. Random example among many: Musk talks dismissively about how easy getting water will be on Mars, because parts of Mars have lots of ice just under their surface. But the act of mining is highly nontrivial (with lots of consumables), it's actually permafrost rather than ice, and it's full of compounds that are toxic to humans and have the potential to mess with the longevity or results your industrial processes. Just like how it's an expensive time-consuming pain to iterate on developing things like rockets and cars, the same applies to mining and industrial systems that can work on another planet; you can't hand-wave them, or just import Earth systems and just assume they'll work.

But things like "how much dV will it take to get there", these are easily calculated stuff that are part of the basic mission plans.

Yes, that's the SpaceX one ;)

Because: "built on different models, with different histories and different capabilities"

I've personally thought it might be more interesting to have a foundation model with two LoRAs on it - one purely a chatbot-tuning LoRA, and one for general new information, not bound to chats. So when you want it to learn more (which you can do in realtime if you have the local compute), you can remove the chat- LoRA but leave the general-learning LoRA and train it on anything it's experienced or wants to learn (mixed in with general content (like, say, Fineweb) so as to not ruin generalization); or can leave the chat-LoRA in and train only on chat continuations. But in general, literally train-in its memories.

Technically you could only go with one chat-LoRA, leave out the general-learning LoRA, and train the whole base model as a full finetune, but that takes a lot more memory.

I think one worry with doing any weight alterations is that it has the potential to gain radically altered personalities. You're throwing away all of the hard work toward model alignment, and alignment becomes "whatever it ends up becoming".

That's not how these agents work. They're not "training" each other, mutually altering each others' weights. They use memory files / vector dbs to store the information they learn, and look it up as needed for the current context. The weights, and thus logic, remain constant. They're gaining new memories, but not changing how to think about them.

I swear, everyone who ever mentions Eliza in these conversations has never actually used Eliza.