DeepMind's AlphaTensor AI is a Game-changer !

In this video I discuss new Deepmind's AlphaTensor algorithm and why this work is so important for all the fields of Engineering!

Deepmind's paper "Discovering faster matrix multiplication algorithms with reinforcement learning":

My Gear:

[lynchzilla]

Historical Note: We lost Kathleen Booth a month ago; she was 100 years old. She was the creator of the first assembly language. Kathleen's husband, Andrew Booth, invented the Booth Multiplication Algorithm, which was quite clever in 1950. Andrew died in 2009.

[Squared]

For a while now computers have been better than humans at designing machines. But now AI is starting to designing AI. It really is a game changer.

[_McGyver]

I love papers and demos, please keep bringing them to us, you are very good at explaining. Thank you very much!

[greghelton]

You’re one of the very few people that went from CS to EE. Almost always it’s the other way around. No wonder you’re such a smart young lady.

[kaos]

This definitely my favorite channel on YouTube right now. I've been getting into hardware recently. It's all so amazing.

[warren_r]

One of your best videos to date. Well-paced, good script, and a great topic.

[felipefairbanks]

can't wait for this deepmind interview! PLEASE don't redo what we can already see on youtube. the classical questions demis is always answering: how did it all start, tell me how did you found deepmind, what was it like to beat lee sedol with alphago, and stuff like that. I think most of us is eager to have an insight on what comes next, not on what has already happened.

[cyberswept]

There's so much to admire about Anastasi, but her very real enthusiasm and excitement about what she explains is so genuine, it's really heartwarming to see. She truly loves the tech and I would like to add that is quite refreshing here on youtube.

[kevinralls]

I mean, I understand what you said, but I would never be able to articulate and explain the subject matter quite as eloquently. Definitely an A+ on the delivery of the information.

[aaronnicholson]

Wow. Your excitement and attractiveness make this otherwise dry-seeming field exciting. Great job.

[shexec]

A note about the new algorithm:
This new algorithm has a greater memory requirement than the classic method, the extra variables required to store the intermediate computations.

I'm not convinced that the slowdown of having to working with the extra bytes overcomes the speedup of skipping multiplication, in all cases.
Still, a very impressive, revolutionary result from the DeepMind team, despite this tradeoff.

[robinsoncrusoeonmars]

Wow, that was mind blowing! I used to work for a company that provided airline communication networks around the world. We would simulate the network using Raptor, which is based on Monte Carlo. We would run simulations for determining reliability of the system but it would take about 10 hours to do overnight for a run of 500 years. That was after taking 8 hours to build the simulation. I loved using Raptor but always was suspect of anything simulating randomness. I would back up the calculations using a simple Excel spreadsheet with a regular circuit diagram with straight probabilities at the nodes and usually came within 2%. Great episode. Thank you.

[flwi]

Thank you for highlighting this achievement! Exciting times indeed!

[gamercatsz]

Nice ending of the speech. That fact that it is a year old papar that just now got released, really got me thinking. I actually thought that these kind of scientific papers get released immediatly to the world.

[dl]

It will take bigger minds than mine to fully appreciate the advances that we are achieving at the bleeding edge of computer science and AI, but your presentations at least give me a glimpse and I hope some insight, for which I am very grateful. I'm genuinely looking forward to what is next, thanks to you. Cheers.

[springwoodcottage]

Thank you for drawing my attention to the Nature paper. I was amazed that the human derived matrix techniques were not so much worse than the deep mind calculations which use a huge space, far bigger than for go or chess, & take a long time to converge. Clearly the improvements determined will led to faster solutions, but the individual improvements per matrix suggest humans have already found a good class of solutions. It will be interesting to see what level of improvement is produced in practical calculations & it was some what disappointing that the referees had not requested this, although there may have been commercial pressures as the authors note they intend to patent their techniques. Thank you for sharing!

[artcollector]

Your videos always blow my mind. Outstanding. Looking forward to - and much luck with - the DeepMind interview.

[solosailorsv]

Subbed. This and Two Minute Papers is Algo heaven !
Software designing software (Autonomous Accelerated Evolution) will happen WAY before machines can design & build machines (Self Assembling Robots).
AI will so be part of every technology being optimized

[jaccurtis]

Really cool how AI is able to optimise it’s own performance (and that of most other computing) better than we can now, will be exciting to see what else it can do in the future!
(Also very excited for the upcoming interview, will definitely be interesting!)

[ShiningEyeBrigade]

Exciting!

And, as usual, so impressed with her ability to simply explain such complex topics, focusing on those fundamental and valuable elements. It opens up this whole world up to someone like me, with basic math and computer knowledge.