90% of TSLA income is from EV sales. Everything else is just Musk's grandiose and increasingly absurd predictions which have a long history of falling short of reality.
Meanwhile the story of Jensen and Musk continued onward with custom chips to support FSDv1, in which J's personal delivery of the DGX1 to OpenAI served as the catalyst to the relationship, iirc.
The problem for nvidia is ... where do you go from here and still have spectacular performance improvements?
Nvidia got extremely lucky again and again and again, and what specifically did it is that right in time non-Nvidia researchers learned to train on smaller floating point bit lengths, which Nvidia raced to support. And great, well, done! A list of ironies though ... for example it's Google Deepmind that made the Turing generation of cards viable for nvidia. However, the new floating point formats train has arrived at it's last station, the NXFP4 station. There is no FP3 and no FP2 to go to. Is there a new train to get on? I'm not aware of one.
Nvidia's argument is "Blackwell easily doubles Ada performance!" ... but that is deceptive. The actual improvement is that Blackwell NXFP4 (4-bits) is more than double Ada FP8 (8-bit) performance in ops. That's the train that's arrived at its last station. Go back further and the same is true, just with larger and larger FP formats, starting at FP32 (single precision). Aside from a small FP64 detour, and a few "oopses" in some of the format they chose turning out useless or unstable, all quickly abandoned that's the story of nvidia in ML.
Comparing, for example, FP32 you don't see big improvements: e.g. 4090: 83 FP32 TFLOPS, 5090: 104 FP32 TFLOPS. Given the power requirements involved that's actually a regression. If you're stuck at 8 bits, nvidia's story breaks down and Ada cards beat Blackwell cards in performance per watt: 4090: 5.44 Watt/FP32 TFLOP, 5090: 5.5 Watt/FP32 TFLOP. Or, FP8, same story: 4090 is 0.681 Watt/FP8 TFLOP, 5090 is 0.686 Watt/FP8 TFLOP. Now effectively the new memory still buys some improvement but not much.
Will the next generation after Blackwell, with the same floating format as the previous generation be a 10% improvement and subject to further diminishing returns and stuck there until ... well, until we find something better than silicon? I should point out 10% is generous, because for FP8, Blackwell is actually not an improvement at all over Ada, on a per-watt basis for equivalent floating point lengths.
Plus Blackwell is ahead of the competition ... but only 1 generation. If nvidia doesn't get on a new train, the next generation of AMD cards will match the current nvidia generation. Then the next TPU generation will match nvidia.
Yes. Most companies play these financial games to some extent.
I lumped government subsidies in with EV sales since they are related. Trump wiped these out.
Robotaxi and robots are the fantasy category. They are not currently income producers and may not be for years to come. His robot demos have been widely panned as fake.
Really bad AI code is no different than really bad sushi.
It can be a lot harder to tell if it's bad.
You can either vet it thoroughly or you can consume it and see what happens.
Business pressures often lead to the latter approach. Managers that would balk at bad sushi pressure developers to consume AI code that is known to often be "bad"--- and then blame developers for the effects.
Generally the law allows people to make mistakes, as long as a reasonable level of care is taken to avoid them (and also you can get away with carelessness if you don't owe any duty of care to the party). The law regarding what level of care is needed to verify genAI output is probably not very well defined, but it definitely isn't going to be strict liability.
The emotionally-driven hate for AI, in a tech-centric forum even, to the extent that so many commenters seem to be off-balance in their rational thinking, is kinda wild to me.
Computer code is highly deterministic. This allows it to be tested fairly easily. Unfortunately, code productionn is not the only use-case for AI.
Most things in life are not as well defined --- a matter of judgment.
AI is being applied in lots of real world cases where judgment is required to interpret results. For example, "Does this patient have cancer". And it is fairly easy to show that AI's judgment can be highly suspect. There are often legal implications for poor judgment --- i.e. medical malpractice.
Maybe you can argue that this is a mis-application of AI --- and I don't necessarily disagree --- but the point is, once the legal system makes this abundantly clear, the practical business case for AI is going to be severely reduced if humans still have to vet the results in every case.
Why do you think AI is inherently worse than humans in judging whether a patient has cancer, assuming they are given the same information as the human doctor? Is there some fundamental assumption that makes AI worse, or are you simply projecting your personal belief (trust) in human doctors? (Note that given the speed of progress of AI and that we're talking about what the law ought to be, not what it was in the past, the past performance of AI on cancer cases do not have much relevance unless a fundamental issue with AI is identified)
Note that whether a person has cancer is generally well-defined, although it may not be obvious at first. If you just let the patient go untreated, you'll know the answer quite definitely in a couple years.
What if anything do you think is wrong with my analogy?
I think what is clearly wrong with your analogy is assuming that AI applies mostly to software and code production. This is actually a minor use-case for AI.
Government and businesses of all types ---doctors, lawyers, airlines, delivery companies, etc. are attempting to apply AI to uses and situations that can't be tested in advance the same way "vibe" code can. And some of the adverse results have already been ruled on in court.
This is a particular meme that I really don't like. I've used em-dashes routinely for years. Do I need to stop using them because various people assume they're an AI flag?
It is still just a collection of inanimate parts. At no point does it suddenly come to possess any properties that can not be explained as such.
Now, apply the same logic to a computer and explain how AGI will suddenly just "emerge" from a box of inanimate binary switches --- aka a "computer" as we know it.
Regardless of the number of binary switches, how fast they operate or how much power is consumed in it's operation, this inanimate box we call a "computer" will never be anything more than what it was designed to be --- a binary logic playback device.
Thinking otherwise is not based on logic or physics.
I think we might be using "emergence" differently, possibly due to different philosophical traditions muddying the waters.
I'm going to stick purely to a workable definition of emergence for now.
Also, let me try a purely empirical approach:
You said the car "never possesses any properties that can't be explained as a collection of parts." But consider: can that pile of parts on the workshop floor transport me over the Autobahn to Munich at 200 km/h?
We can try sitting on top of the pile while holding the loose steering wheel up in the air, making "vroom vroom, beep beep" noises, but I don't think we'll get very far.
On the other hand, once it's put (back) together, the assembled car most certainly can!
That's a measurable, testable difference.
That (the ability of the assembled car to move and go places) is what I call an emergent property. Not because it's inexplicable or magical, but simply because it exists at one level of organization and not another. The capability is fully reducible to physics, yet it's not present in the pile.
parts × organization → new properties
That's all I mean by emergence. No magic, no strong metaphysical claims. Just the observation that organization matters and creates new causal powers.
Or, here's another way to see it: Compare Differentiation and Integration. When you differentiate a formula, you lose terms on the right hand side. Integration brings them back in the form of integration constants. No one considers integration constants to be magical. It was merely information that was lost when we differentiated.
86 billion neurons, 100 trillion connections, and each connection modulated by dozens of different neurotransmitters and action potential levels and uncounted timing sequences (and that's just what I remember off the top of my head from undergrad neuroscience courses decades ago).
It hasn't even been done for a single pair of neurons because all the variables are not yet understood. All the neural nets use only the most oversimplified version of what a neuron does — merely a binary fire/don't fire algo with training-adjusted weights.
Even assuming all the neurotransmitters, action potentials, and timing sequences, and internal biochemistry of each neuron type (and all the neuron-supporting cells) were understood and simulate-able, using all 250 billion GPUs shipped in 2024 [0] to each simulate a neuron and all its connections, neurotransmitters and timings, it'd take 344 years to accumulate 86 billion of them to simulate one brain.
Even if the average connection between neurons is one foot long, to simulate 100 trillion connections is 18 billion miles of wire. Even if the average connection is 0.3mm, that's 18 million miles of wire.
I'm not even going to bother back-of-the-envelope calculating the power to run all that.
The point is it is not even close to happening until we achieve many orders of magnitude greater computation density.
Will many useful things be achieved before that level of integration? Absolutely, just these oversimplified neural nets are producing useful things.
But just as we can conceptually imagine faster-than-light travel, imagining full-fidelity human brain simulation (which is not the same as good-enough-to-be-useful or good-enough-to-fool-many-people) is only maybe a bit closer to reality.
Well, the amount of compute is certainly finite in this era. 250 million GPUs in a year is a big number, but clearly insufficient even for current demand from LLM companies, who are also buying up all available memory chips increasing general prices rapidly, so the current situation is definitely finite and even limited in very practical ways.
And, considering the visible universe is also finite, with finite amounts of matter and energy, it would follow ultimate compute quantity is also finite, unless there is an argument for compute without energy or matter, and/or unlimited compute being made available from outside the visible universe or our light cone. I don't know of any such valid arguments, but perhaps you can point to some?
There's also no proof that intelligence can not be produced by an algorithm. Given the evidence so far, like LLMs seem to be able to beat average humans at most tests and exams it seems quite likely.
reply