They chose BnB to suggest "bed and breakfast", a common term for a small hotel that was typically just an old house where the owners lived and would host, as we all know. So we know what the right result will be. But since it's a giant corporation versus a person with no resources, we may get the wrong result.
They seem to be mixing together things that are actually harmful to the public, with things that are merely harmful to their business model (which is their claim that they can grab whatever data that they want regardless of the wishes of the owners of the data and use it to improve their models, but competitors can't do that to them).
Even before LLMs were a thing, it wasn't this way: rapidly generating code was not the most valuable skill. As you say, it's much more important that the code can be confidently modified and extended, and reused, not just now, but then. In a mature product, the initial writing of the code will be the least of the work; maintenance is much more expensive. Ideally, design decisions should appear only once in the code when this can be achieved, because then there's one place to fix or one place to modify, instead of dependencies on some detail that appear all over the code. It's too easy with auto-generated code to wind up with redundancy and code duplication, resulting in a brittle mess.
Hopefully in this world, someone figures out how to deliver the performance equivalent of a B300 GPU for about 1/100th the power of a current B300 (which can be up to 1400 watts), or the world will bake.
Compared to 20-30kw spend on cruising the highway in a car (considerably higher for older ICEs) 1.4kw does not really seem to dent the energy consumption.
Especially if cognitive technologies mean that we need to travel less (eg communiting to work, or ineffecient supply chains).
2-3 hours at 20kW daily (assuming commute, etc) vs 20ish hours always on tasks really is close to equivalent
which is to say that the last thing our planet needs is another universalized technology that outputs as much total emissions as cars
in an ideal world, we'd keep LLMs/CNNs/etc specialized and academic until we are hitting diminishing returns on optimizing fundamental microprocessor tech like GAA. but the pursuit of market dominance and mass adoption is our current operating philosophy, and so we have things like this top graph: https://hai.stanford.edu/news/inside-the-ai-index-12-takeawa...
>Grok 4's estimated training emissions reached 72,816 tons of CO2 equivalent, or roughly the same amount of greenhouse gas emissions created from driving 17,000 cars for one year
current global average electricity production = approximately 3.6 terawatts
But then again these systems do not use 1400 watts all the time. We would probably have much more then 11.2 terawatts demand if all humans turn on all their electrical consumers at the same time.
Most of the energy consumption comes from moving data between memory and compute units rather than the math itself.
While HBM stacked on a silicon interposer was a signification efficiency improvement over traditional DDR, there's still room for improvement. The high energy costs will drive this further, e.g. near-memory/in-memory compute (PIM) or tighter 3D packaging and possibly optical interconnects.
(And of course training/model optimisations)
Right, if everyone in the world has a GPU we would have solved so many problems with power generation and... I'm not going to do the math but I think we'd be mining asteroids too? I'd probably use it as a bookend at the point.
I think personalized suriveillance and manipulation (thought control, manipulating people through filtering of content they receive) on a level any dictator would drool over can be gotten for a fraction of 1 GPU per person.
Specialized chips like what cereberus is building combined with small models can give great result in the long run, but at the pace AI is growing it's not very practical yet.
For traveling salesman that's more than good enough. But in many cases an O(n^3) algorithm can't be used because n is in the billions. I remember interviewing a candidate who asserted that register retiming in digital circuits was a non-problem, so they were surprised that we were still working on improvements, because they had learned that the Leiserson-Saxe algorithm gives an optimal solution in O(n^3) time. But because real circuits are so large that that approach can't be used. Polynomial time often isn't good enough; even quadratic time often isn't tolerable.
I spent my career in electronic design automation, where practically every interesting problem is NP-hard, but we have to solve them, or approximately solve them at least, and because real-life problems often have structure, with the right approach very large problems can be solved exactly despite the theoretical complexity, and when exact solutions can't be found a decent bound can often be found that is an acceptable solution.
Sales people still have to plan their trips even though finding the optimal solution is NP-hard (to give one example). No matter; there are decent heuristic methods.
It's been common in electronic design automation tools to have license terms like that (forbidding use to create a competing product). However, competing companies have often found workarounds, either by finding loopholes or just breaking rules and hoping not to get caught.
In many cases, I found that when colleagues said "we didn't change anything" they mean "we didn't change anything relevant" which winds up meaning "we didn't change anything we think could be relevant to this issue" and then "whoops, turns out it was relevant after all".
It was the preferred lab computer in the mid to late 1970s and into the 80s. I got my first job because I knew PDP-11 assembly language, and worked with both DEC's operating systems for them (RT-11 and RSX-11) and later Unix (the lab I worked with had some machines running Version 6, though Version 7 was the first that I used seriously. It had a very clean and symmetric instruction set that used the program counter as if it were another general purpose register. I had an LSI-11 board (the single-board version of the machine) with 4K 16-bit words of core memory and a paper tape punch with a tiny loader in ROM to read in the tape and peek and poke memory, and I'd sometimes initialize the core memory to a known state by running the one-instruction program
mov -(pc), -(pc)
or 014747 in octal. It would fill all of memory with 014747.
I built the PDP-11/70 emulator that controls the nuclear reactors in Ontario. That was 20 years ago and I'm probably still the youngest person who can read PDP-11 assembly (and the raw octal)
There's a decent chance you're younger than me, but I can still read it. Coincidentally about 20 years ago I got asked to fix a bug in some PDP11 code that ran on real hardware because it was tripping up an emulated system. Nothing wrong with the emulator as such, it was just a logic bug in the program that they no longer wanted to work around.
I made the fateful mistake of saying "Sure, sounds fun, how hard could it possibly be...?"
Not pdp-11 but once I was tasked with fixing a c program the offshore team had failed to fix for over a year. I basically took a month off (at my desk) after I fixed it the first day. They were happy, I was a legend at that company!
The FPGA emulator we made had to pass the xxdp test suite which was provided to me on punched tape and microfiche. The emulator had a specific test for FDIV overflow which even tested the accuracy of the partial result. None of the software emulators I tested did this. I emailed Gordon Bell who introduced me to Bob Supnkk, and we found the original flowcharts for it so I could replicate the divider logic precisely. Imagine a nuclear reactor dependent on this lol.
> It had a very clean and symmetric instruction set
Indeed. Motorola's 68000 CPU took so much inspiration from the PDP-11's ISA, it was almost a spiritual successor. The 68000's 8/16-bit little brother, the 6809, widely considered the most powerful 8-bit CPU ever - was also heavily inspired by the PDP-11.
reply