golol

Yes I don't think it should be allowed in Europe, yet the outrage I see, also from, Americans, as if this is some new never seen before level of danger on the road is unreasonable.

>And it may well help launch it, but that’s where its relevance ends.

You can expand in parallel. If Starship works you can launch 100 spacecraft instead of 1, or in other words, you can expand your system by a factor of 100 "in parallel" and hence increase whatever you get from that small payload margin by a factor of 100.

Besides that there are other approaches like laser array + photon sail which more directly benefit from mass to LEO.

It's basically the same size as an F-150. It has a crumple zone although it probabl only activates for hard car collisions. I've seen no evidence that it is significantly worse for pedestrian safety than an F-150. Which is not good at all, but it is still a bit pretentious to paint it as a horribly dangerous car when there is no evidence that it is worse than one of the most common cars in the US.

I have my fingers crossed for Starship. If it works as intended all these things will happen.

What I'm saying is that it might be the case that "capability c" (somehow measures as a positive scalar) of your fused sensor system is not the sum c = c_1 + ... + c_n of the capabilties of each sensor c_i, but more like c = (c_1^p + ... + c_n^p)^(1/p) where p is a large power. The case p = 1 is where every sensor contributes to the whole system with perfect efficiency. The case p = infinity, i.e. c = max (c_1, ..., c_n) is where really only the best sensor does all the work and the weaker sensors can not add anything useful. The case of p > 1 somewhat large represents that all sensors contribute but the real heavy work is being done by the best sensor and other sensors only have significant impact if they have roughly equal capability. So I believe that the team at Tesla has the intuition that p is too big for radar to be worth it.

I think if you have two sensor systems with one being significantly less capable than the other, the gain from having both quickly diminishes. I believe Tesla dropped the radar because it was simply outclassed by the camera in 99% of circumstances. So the radar really just becomes a hacky failsafe for bery rare circumstances, and Musk would like to avoid such a situation, he wants a "clean" solution.

o1 is much much much better at mathematics. I am am a phd student and sometimes give it small problems to play around with, just to get some inspiration. That's the thing: It actually plays around and tries stuff instead of just spitting out a remembered answer like GPT-4o.

I replied with some guesstimations to the other comment, I would be curious to hear what you think.

So Starship+Superheavy have about 4.6kt of propellant which should mean about 1kt of methane. That's about 2.5kt CO2 per flight which yields 2.5mt CO2 annually for 1000 launches. The US releases around 4.4Gt CO2 annually so you end up with around 0.05% of US emissions. I actually thought it would be more so I hope I didn't miss an order of magnitude somewhere. Quick search says US airliners release less than about 200mt annually so you're looking at less than 2.5% of that. Annually about 850 million people fly in the US so if you imagine a Starship flight having 100 passengers then in some weird way flying with Starship is maNe about 200 times more CO2 intensive than flying with a plane.

Now high-altidude effects would increase the greenhouse effect of these emissions by some factor I don't know. At the same time a large amount of fuel is actually burned by the booster at below airliner altitudes. Another mitigating effect is that some of this methane could be produced renewably.

I haven't heard about Ozone depletion from spaceflight except for the case of reentering Aluminum burning up. Well this wouldn't be an issue with Starship flights.

So you know overall, let's just give it a lot of conservative margin and imagine we are looking at about 10% of the US airline or 0.2% of total US greenhouse impact. That is on the scale of a whole new industry, yes, but if you can imagine a world which has attained a sustainable rquilibirum surely you can also imagine one which has a bit of margin left? And this is only for the US!

What do you gain? Three of the points I mentioned have potential direct benefit for people.on earth: asteroid mining, L1 solar shield and brilliant pebbles.

Asteroid mining "for profit" also means that the economy profits, unless you believe that capitalism doesn't work. I read that less than 200t platinum are produced annually. Say your asteroid mining crashes that economy. I found some figure of 20kt of CO2 for each ton, so 4mt in total. That is more than your 1000 launches emit! Now my calculations are of course wrong but it seems plausible that the orders of magnitude could match. So you should definetly consider that your asteroid mining operations might stop other mining operations from happening which reduces your greenhouse impact.

Then you have the solar shield. It is certainly possible, although of course difficult and has some risk. But it's good to have a backup plan, no?

Because you should consider the upside when you mention the downsides. 1000 Starships launching a year can mean exoplanet imaging, drilling on Europa, interstellar probes, an actual Mars colony, asteroid mining, brilliant pebbles, O'Neil cylinders, L1 solar shields, space junk removal and so on. Not all at the same but each of these is a project that is potentially feasible with enough mass in orbit. I'm not saying you have to like or care about any of them but you should consider what you actually gain for all that emitted CO2.

Take Optimus, teleoperate it to be a Bartender, collect the training data for your actual robotender. Repeat for any reasonably specific task.

"Everything comes at a cost" as a meme is really deeply integrated in our culture. Since thousands of years probably.

Why not? What would you rather be: Fat to your death, or healthy and dependent on modern technological society in yet another way? How is it any different to diabetes treatment?

What's wrong eith a medicine as a crutch? If you break your leg you use a damn crutch and that's good. If you suffer from an illness and we have a medicine that's worse than the illness and affordable - go for it. Phrasing it as a crutch suggests it is somehow only a temporary that prevents you from finding a "real" solution by changing your "lifestyle". It doesn't matter, only outcomes matter.

Yes which can not do a bunch of things that FSD can.

Tesla has the world's best autonomous vehicle offering by some way of measuring things. There are many ways to measure things, but at least in some "category" they are indeed leading: Level 2 systems for the US which you can privately buy.

Just earlier today I wanted to check if exp(inx) is an orthonormal basis on L^2((0, 1)) or if it needs normalization. This is an extremely trivial one though. Less trivially I had an issue where a paper claimed that a certain white noise, a random series which diverges in a certain Hilbert space, is actually convergent in some L^infinity type space. I had tried to use a Sobolev embedding but that was too crude so it didn't work. o1 correctly realized that you have to use the decay of the L^infinity norm of the eigenbasis, a technique which I had used before but just didn't think of in the moment. It also gave me the eigenbasis and checked that everything works (again, standard but takes a while to find in YOUR setting). I wasn't sure about the normalization so again I asked it to calculate the integral.

This kind of adaptation to your specific setting instead of just spitting out memorized answers in commonn settings is what makes o1 useful for me. Now again, it is often wrong, but if I am completely clueless I like to watch it attempt things and I can get inspiration from that. That's much more useful than seeing a confident wrong answer like 4o would give it.

I'm a math phd student at the moment and I regularly use o1 to try some quick calculations I don't feel like doing. While I feel like GPT-4o is so distilled that it just tries to know the answer from memory, o1 actually works with what you gave it and tries to calculate. It's can be quite useful.

>Your argument does make sense for teaching an intro to programming type class, where it may not be obvious what is a function name versus variable name.

Yea but the truth is a huge of audience of people who are not super familiar with the syntax of a languague will read tutorials describing how to build stuff in that language hoping they can piece things together. So for them it would be helpful.

Also, isn't all of programming just playing with types in a way? I feel like Intro to types never stops lol.

The point is that tyoe is an EXTREMELY vaulable information and if you are explaining code to someone it is very helpful to clearly see what are keywords, what are arbitrary variable names, and what are the types of the variables. For example if you show me a programming language where there is a list object and you write list.one(), I don't know if list is a variable or a keyword, and I don't know if one is a variable or a keyword. Much better to write mylist.one() if one is a default function, or mylist.myfirstelement() otherwise etc.

I mean everyone knows using descriptive variable names is good practice, but then in a coding tutorial it is somehow fine to use foo, bar, a, b, c? That makes things clearer for someone who understands all the types and the syntax, and wants to see the structure algorithm more clearly. It hurts someone trying to learn the language.

foobar should die out. myvariable, mystring, myfunction etc. are better in every way.

I have a similar situation lol, I used a Facebook account almost a decade ago which I lost access to (email). Whenever I try to make a new account it instantly gets banned as a duplicate.

Yea sure it is a difference, but for me not outrageously immoral. I guess you can get in trouble though.

This seems less like a "manual for attack" and more like tweeting that your local storage unit rental never puts locks on their garages and gates and "anyone could just walk in and out".

> I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains.

Sure it's plausible just like many other forms of baseless speculation are, I still don't think to state it as a matter of fact like it is happening in this and other articles and in comments.

> I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains.

Yes but do you know of any examples with sequential data, and in particular transformers? I don't and I think it is a somewhat different thing. With pictures you just train an encoder into some low dimensional latent space and there you can interpolate, but working with sequences you don't want to give the whole sequence as an input to the model and generate the whole sequence as an output.

The natural way to do interpolation with an audio transformer model, which it can in principle indeed do, would be to somehow audio-prompt engineer it to speak in a voice that interpolates SJ and your voice actor. But that would be a pretty fragile method.

> AI is astonishingly good at laundering and remixing without exposing the training set

They just about manage to make a good multimodal transformer that can generate audio and you expect that right away they can also interpolate in latent space? How does that actually work? It's not so simple. What benefit do they have from training on Scarlett Johansson's data, because they sure as hell have a big risk. They clearly hired a voice actress and they clearly told her to sound like Scarlett Johansson in "Her", and the end result perfectly fits with that. The voice doesn't "uncannily sound like SJ", no it just vaguely resembles her voice and mostly just mimicks the mannerisms from the movie. For me this is a perfect example of Occam's razor. One explanation is simple and realistic. The other explanation requires significantly more advanced AI control than OpenAI has claim/demonstrated, and it requires Altman to be so obsessed with the SJ idea that he goes out of his way to secretly train on her voice, risking legal exposure, while still hiring a voice actress.

>I remember when OpenAI practically cloned Scarlett Johanssen's voice...

There is absolutely zero evidence for this. I find it infuriating that this keeps being stated as a fact. So they go and hire a voice actor and clearly use her voice to train, but then they also scrape Scarlett Johansson from youtube and splice it into the training data to make the voice a bit more like hers? Really does that sound realistic?

Why does the following obvious solution not work: - Homework is just voluntary. You have to force yourself to study anyways. Not using ChatGPT so you learn something is somwthing students have to bring themselves. - Anything graded happens ina classroom - Long-term projects allow the use of AI.

Exactly. As long as OpenAI keeps putting the world's most advanced AI into my hands I will accept that they are open - in some sense. Maybe things would have always been like this, maybe if OpenAI didn't exist Google and the other actors would still publish Claude, Gemini etc. But in this world it was OpenAI that really set the framework that this process of developing AI happens in the public eyes right now. GPT-2, GPT-3 with APIs, Dalle, ChatGPT, GPT-4, now o1 and soon to be voice mode. OpenAI ensured that these aren't just secret toys of Deepmind researchers or something.

Even if up to 30% of casualties were civilian this would still be quite a surgical strike in my opinion. At 50% I would not say so anymore. Nobody knows yet anyways.