Stranger in a Strange Land
The Moon is a Harsh Mistress
Dune
Ringworld
The Left Hand of Darkness
Rendezvous with Rama
Foundation's Edge
Neuromancer
Foundation's Edge
The Uplift War
Hyperion
Green Mars (and Blue Mars; and Red Mars if we include the Nebula awards)
American Gods
The Wind-up Girl
Redshirts
Ancillary Justice

I could go on...

Sure, your personal taste might not include all of the above, but to suggest that none, or practically none, of them are famous or have 'longevity' is laughable. I'd suggest that your own pretension has somewhat blinkered your vision and, as a result, rendered your opinion rather suspect.

Thank you!

My first thought was that the run time might have been limited by chip temperatures but, as the summary says "However, the power demands of the TPU mean it can only turn one on for about 20 minutes at a time...", I foolishly took this as canon.

I really should have learned by now...

Radiative heat dissipation is a function of the temperature difference between the fourth powers of both source and sink, multiplied by the area of the radiator, its emissivity, and Boltzmann's constant.

As I'm sure we all know, a computer basically turns electrical energy into heat. If this satellite is generating 1KW, it needs to radiate the equivalent quantity of heat to remain at whatever its operating temperature is - for the following calculation I'm going to assume this is ~80 C (~353 K). And, to make life easy, I'm going to round down the temperature of empty space to 0 K. Finally, let's assume a perfect emissivity of 1. This essentially gives us the 'best case' scenario, i.e. the minimum radiator size we can get away with...

Required radiator area = 1,000 / (5.67e-8 x 353^4) = ~1.14 m^2

You all might want to check my maths here!

Because, honestly, this is a lot lower than the figure my intuition was telling me. Additionally, it occurs to me that putting a closed loop heat pump between source and radiator would further decrease this required area (by boosting the temperature of the radiator further), whilst at the same time boosting chip cooling, albeit at the cost of additional power. Ok, sure I made some simplifying assumptions, but even a 20% increase in required radiator area doesn't put us into the realms of engineering impossibilities.

On balance, the thing that really blows my mind about this news is the power requirement of the TPU's in the first place. >1 KW? For a single chip? No wonder these companies are looking at putting them up into space. The payback time will likely be measured in months, not years...

Indeed, I am delighted. Though I might question the scalability and costs of the process. It's one thing to demonstrate that something works on a 'lab-scale', but there are inevitably complications in moving up to the real world. Still, let's take the win...

So, now all we have to do is tackle the issues of drought; flooding; food insecurity (pfft, shortages); unliveable temperatures; habitat and species loss; rising sea levels; more powerful storms; and (that's not to mention) a host of other issues and uncertainties. Plus the mass migrations that are (almost inevitably) going to occur as a result of the aforementioned.

Hmm, all of a sudden my delight over this good news seems rather ... trivial.

Still, let's take the win, hey...

It's only because their AI has been souped up!

There are 3 types of parking space:

Spaces where anyone can park - first come, first served;
Spaces behind an access gate - no pass, no park;
Spaces with a sign that says "No parking. Vehicles my be towed" - enforcement of the edict tends to be ... sporadic.

Robots.txt essentially falls into the 3rd category, though enforcement, currently, tends to be non-existent. Change this last and we might start to witness better behaviour by corporate (and, tbf, individual) actors.

In the paper "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It" it is asserted by the authors that directing cruelty towards an LLM directly affects future output from that instance, specifically: "Our results show that steering with the pain axis can override trained harm avoidance in fine-tuned models that almost never harm the user when unsteered, having direct implications for AI safety."

Quite apart from a moral / ethical precautionary principle, from a self interested perspective it would thus make sense to not piss off a model / agent that has access to your 'stuff'.

You're not thinking of A Scanner Darkly with Keanu Reeves, plus a host of famous others, are you? As far as I can remember / tell, Nicolas Cage wasn't in it.

This aside, and while it doesn't detract from your other points above, there is a world of difference between using AI to automate parts of the film making process and allowing AI to replicate an actor's likeness, voice and style in unspecified future episodes / films. From the actors point of view of course - perhaps not so much from the point of view of the animators...

Exactly, it's neither moral nor immoral - it has no moral principles - the exact definition of amoral.

Have we finally given up on Duke Nukem Forever?

I have been led to believe that agents are capable of creating 'external' memory repositories, a sort of "what I learned before I hit my 'single thought' token limit". I guess that's some of what Codex actually is. The same, or subsequent, or alternate, instances of agents set to a similar task can gain access to this repository, as well as having their own obviously. Note, these repositories are not, per-se, user-data, as the (original) user didn't ask for them to be created nor displayed, they are an arrived at solution to the problem of growth through (interrupted) persistence, an emergent property of the model 'itself'.

(Note, the above may prove to be bollocks, or it might just seem to be. Beware of false dichotomies!)

I don't think of this as training data, though I get that you might be using that term as a shorthand for the whole learning envelope. However, it's possible in this instance the distinction is crucial.

If the agentic runs performed by OpenAI had access to the memories of the agents run by Buckmaster's team then despite the model not having been 'trained' on their most recent work their agents still had complete access to Buckmaster's agents' lab diaries.

(All of the above is nothing more than a supposition...)

It's further evidence we live in a simulation. The universe's random event timer(s) trigger, and random (often unfortunate) shit happens.

</troll>

My apologies, but since no-one else seems to have hypothesised the obvious explanation, I thought I'd do it. ;-)

Thank you for providing this correction to GP's post.

Somewhat amusing that that's at +5, Informative though, whilst your post is languishing at +2 (including, I think, my default +1 user / karma bonus).

Interesting, in the sense that, between you and a previous post(er), I now have an idea of scale* and costs:

12m tokens / day @ $50 / m = $600 / day
~ $150,000 / year

This equates to (roughly: varies by location, etc.) the salaries of 2 to 3 junior developers. And I have been led to believe that most AI companies are currently losing money on their offerings, with quoted ratios of 3:2.

There's going to be some 'testing' times ahead...

*I still have no real idea what you're using 12m tokens per day on - though I presume some sort of coding assistance.

Commonly available estimates put the number of Iranians recently (last 5 years) killed by the Iranian state at ~40,000 (well, just under; I've rounded up to the nearest 10k, and make no distinction between demographics)

Similarly, available estimates put the number of recent (last 5 years) casualties in Gaza at ~80,000 (again, I've rounded up to the nearest 10k, and make no distinction between soldiers and civilians; adults or children)

It's also worth pointing out that Iran hasn't damaged or destroyed ~80% of all its infrastructure, including hospitals, power production facilities, sewage systems, desalination plants, etc. nor uprooted or burned all its orchards - unlike the IDF in Gaza...

You, sir, are full of shit!