Special drop: Cerebras CEO Andrew Feldman, fresh off the $95 billion IPO, in conversation with Alex Wissner-Gross. Alex asks the questions, Andrew tells the story.
A wafer-scale engine 58 times larger than any chip ever built. SRAM stuffed to the gills on a chip the size of a dinner plate. The contrarian bets that took years of perseverance, $8 million a month for 18 months with no solution, then breakthrough. 15-20x faster than the GPU on inference. The OpenAI deal north of $20 billion. The AWS term sheet.
Then the bigger questions: how do you cut a 10-trillion-parameter model across chips? Why is Cerebras the right platform for chips in orbit, where fault tolerance is everything? Why do fabs take five years and $50 billion, with TSMC the greatest manufacturing company on Earth? Why do Samsung and Intel keep failing at the same node? Why are NVIDIA and AMD great at GPUs and terrible at everything else?
Alex pushes on Sam Altman as the counterweight to Elon and SpaceX AI's Dyson Swarm. Andrew on luck, hard work, and why you don't need 'the highest paid talent' to win. The pressure test on the soul of being a startup CEO.
Built from the Moonshots Podcast EP256 (Peter Diamandis, Salim Ismail, Dave Blundin), where Andrew Feldman joined Alex Wissner-Gross. Timestamps below.
#AI #ASI #ASIPill #Cerebras #AndrewFeldman #AlexWissnerGross #WaferScale #SemiconductorIPO #Inference #SpaceCompute #Moonshots #Hyperstition
TIMESTAMPS:
0:00 Welcome and setup
0:17 [Parzival] Strap in. The Cerebras CEO just rang the bell for a 95 billi
0:23 Andrew on the $95B IPO bell
0:48 [Parzival] And what does Andrew make of Karpathy joining Anthropic, and
0:54 Karpathy and frontier-lab drift
2:10 [Parzival] And what does the Cerebras founder think of the Elon and Sam
2:14 Elon vs Sam, both building cool
3:19 Cerebras mojo, exceptional people
3:35 [Parzival] On the ASI Pill, we want the origin story. How do you actual
3:43 Founding bet: chip the size of a dinner plate
8:30 [Parzival] And which of all those bets does Andrew think were dead wron
8:33 Got many bets wrong, inference bet right
9:51 Alex Q: SRAM at trillion-parameter scale
10:25 Cleaving 10T-parameter models across chips
14:06 [Parzival] And what about Elon's Tera Fab? How long does that actually
14:10 Elon's fab bet: 15-20 year project
16:15 [Parzival] Then what is so hard about building a fab, really?
16:18 Fabs are pyramids
19:31 [Parzival] And which fab does Cerebras actually use, and which one woul
19:35 Three-nanometer at TSMC, never Intel
21:22 [Parzival] What is the thing Cerebras quietly solved seven years before
21:26 Solved CoaS seven years before NVIDIA hit it
25:18 Alex Q: WSE 8 and the 10-year future
25:51 Infrastructure builders don't pick the apps
27:52 Alex: deliberately deferring to frontier labs
28:09 Sparse linear algebra is the bet
28:21 [Parzival] And what is the killer app for the wafer-scale engine in spa
28:25 Why Cerebras has an edge in space
29:53 Alex: fault tolerance in ionizing radiation
30:19 Production chips in space, a decade out
31:29 [Parzival] Which professions actually fall first to language models?
31:32 Lawyers and accountants vs language models
32:23 Alex teases Andrew on obscure knowledge
32:33 Why Intel and AMD failed at cellphone chips
33:57 Alex: massively transformative purpose
34:02 What made Intel say no to Apple
34:15 Intel chasing margins, sold its ARM division
34:23 Luck is not equally distributed
35:01 [Parzival] And why does the team with the most money lose, again and ag
35:04 Why the team with the most money doesn't win
36:24 Alex: Sam Altman as counterweight to Elon
37:12 You underestimate Sam at your cost
38:58 Alex: fabs in space don't need heavy launch
39:58 Building a fab on land is hard enough
40:25 [Parzival] What is China actually winning at, when it comes to AI infra
40:29 China's power infrastructure advantage
41:17 Alex: China siphoning American tokens
41:43 Being CEO is enough
41:49 Signoff
