Full transcript
0:00David Silver just raised $1.1 billion at
0:03a $5.1 billion valuation, the largest
0:06seed round in Europe so far, to prove
0:08that the AI industry is building the
0:10wrong thing. The thesis is that every
0:12frontier model right now, whether it's
0:14GPT, Claude, Gemini, is all bounded by
0:17human knowledge. Because they learn from
0:19human data, they're able to retrieve,
0:22compress, approximate, or even remix
0:24human intelligence. But, they claim that
0:26by design, they cannot discover things
0:28that humans don't already know. The
0:30company is only a few months old, but as
0:32some of you know, Silver is one of the
0:34most important people at DeepMind and
0:36the person behind AlphaGo and Alpha
0:38Zero, both of which made AI history.
0:40AlphaGo in 2016 when it defeated the
0:43world Go champion, and Alpha Zero, which
0:45was released a year later in 2017 when
0:48it discarded human data entirely and
0:51learned only by playing itself, starting
0:53with the rules of the game. Using this
0:55method, Alpha Zero was able to surpass
0:57AlphaGo in just a few hours, and even
0:59went on to master chess and shogi using
1:01the same algorithm. What was super
1:03radical was that Alpha Zero stunned all
1:05of these grandmasters because it was
1:07able to rediscover centuries of
1:09established theory on its own, and even
1:12went beyond to do super unconventional
1:14moves that were superior and ultimately
1:16won. The big difference was that AlphaGo
1:18learned from humans and Alpha Zero
1:20learned from first principles alone. So,
1:22based on this theory, Silver wants to
1:24build what he calls a super learner, an
1:26AI that learns entirely from its own
1:29experience without any human examples or
1:31input. AKA the way that Alpha Zero
1:33learned chess, but generalized to
1:35everything. Of course, reinforcement
1:37learning itself is already pretty common
1:40today and used by most labs. So, Silver
1:42isn't trying to pioneer RL here, but
1:44rather he's trying to remove the human
1:46data ceiling entirely. The idea that if
1:48we really want to build
1:49superintelligence and go beyond what
1:51humans know, we have to start without
1:53human data or human priors. Right now,
1:55Ineffable Labs doesn't have a product or
1:58any published benchmarks, but the
1:59investment here is really in the talent
2:02and the thesis.