Are we all wrong about AI?

The use of AI for the benefit of mankind in the fields of science, medicine, nature is still in it's infancy but can not be stated enough. The people that dislike AI specifically because of entertainment sector such as art, music is SO SHORTSIGHTED and actually selfish that I just ignore their general complaints....and I'm an artist myself. But I'm also a student of science and technology. AI MUST grow for a better future for humans and life in general.

The humor of a video about AI innovations and the promotion of a job searching website is not lost on me.

"If corporate greed doesn't get to it" If my grandmother had wheels, she would have been a bike.

The chart at @5:45 is misleading in that the points for Hopper and Blackwell are FP8 and FP4. To make the comparison more fair everything should be in FP16.

Finally getting that cyberpunk with our dystopia. Nice.

Previous video on the negative side of the AI: https://youtu.be/vQChW_jgMMM?si=oFrosTUKA4saTpEA Also correction on the number of amputees, it should be 50 million not 500 million. The 500 million figure included smaller amputees like fingers and toes.

Been a subsriber since the channel had 30k subs. Still my favorite channel on Youtube. Please keep up the amazing work!

The problem when discussing the risk/reward of AI is that it's such a dramatic shift. You can analyze it's impact as much as you want but the future will not judge the success of AI on the risks and benefits we see now, but on the risks and benefits we've not even conceived of yet.

Little expansion on the chip foundry section. Structures aren't really 2nm or lower in size as that can cause issues with quantum behavior of electrons at that scale, which is why the development of more efficient structures is portrayed in the media as "on the edge of physics" and the numbers are cranked down for marketing reasons. Usually the developments boil down to developing more densely packed designs or slightly better lithographic processes that are resulting in an efficiency jump (usually measured in transistor density or power consumption/compute power) that otherwise only would be achieved by downscaling classic ~40nm structure processes to e.g. 14nm, 12nm, 8nm or 2nm. Now the new software by Nvidia is used to compensate for the diffraction of UV light that is sent through a mask and "carves out" the circuits from the silicon. The mask essentially is a (partial) blueprint for all the circuits to be etched into the silicon wafer. But at these small scales the photons of the UV light are diffracted a lot by the masks they're sent through which causes blurry edges and in the worst case results in separate circuits overlapping, circuits with gaps or other faulty areas in a chip. To make these masks as precise as possible so one can increase transistor density and reduce faults, software simulates how to achieve the best design and also generates the masks for these designs, with those simulations traditionally taking up thousands of CPU hours (the mentioned 2 weeks). The AI by Nvidia is used to improve the mask software's prediction as to how the UV light will pass through the mask and increases the precision of the lithographic process while also being a lot faster (about 8h). This allows for a much quicker change of a process as you don't have to wait weeks for new masks. ASML, as the world's leading lithographic machine produce develops the machines which TSMC is using to produce Nvidia's GPU designs. TSMC uses Nvidia's new software on existing, modified machines, so it can only speed up the design process for the machines they currently have available. ASML joining in is crucial for Synopsis and TSMC to in future have machines that are both on the edge of modern engineering and capable of using Nvidia's software, else if they don't have machines that are precise enough supplied by ASML, TSMC and Synopsis will only get so far with Nvidia's AI based software optimization before running into limits again.

Back in the early 90s, I investigated early AI and how it was being developed. I found there were two different types that were radically different. One was Neural Networks, which devised electronics to replicate a human brain. It was difficult and time-consuming, so it was basically no better than a fly's brain. The other was Expert Systems, which replicated it using software. This, of course, was simpler and easily manageable. The problem was that it wasn't very expandable or replicated the thought process very well. Years later, someone told me that the current AI tech we were using (about the time GPT came out) was the Neural Network version amplified with our new technology. Now, I'm having serious doubts about the reality of that. If it was truly Neural Network hardware designs, then current AI would be at the level of actual brains, many times over. Obviously, hardware is involved in AI to some extent, but I think it's more like a cross between NN & Expert, just boosted with better hardware. I feel that we may never actually get a good quality AI going because we didn't stick with the true Neural Networks design. Our brains don't work similarly to computer systems, so nVidia can't really be making real Neural Networks designed graphics chips. Of course, I could be wrong.