Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit with CEOs of Cerebras & Black Forest Labs
with Jason Calacanis, David Sacks, Chamath Palihapitiya, David Friedberg
10 Jul 20265 min read1h 10m
TL;DR
Cerebras CEO Andrew Feldman argues we have already hit AGI by any definition from 20 years ago, and that the next leap — super intelligence — will come from recursive 'loop maxing,' where models iteratively improve their own reasoning. Feldman reveals Cerebras has a $25 billion backlog, and explains that his chips' blisteringly fast inference is what makes extended reasoning runs (25–48 hours) tractable. The conversation also tackles open-source model sovereignty, the Anthropic government standoff over staged model releases, and why every nation from Kazakhstan to Armenia is now building out data centers.
Key Moments
Andrew Feldman
“what we're talking about now are data centers that are in the next several years going to use more power than the previous 50 years on Earth took.”
Feldman is describing the unprecedented physical scale of the current AI infrastructure buildout to Jason Calacanis.
“The irony is unlike many sort of exciting times in technology. They're trying to capture yesterday's demand, right? The demand is way outstripping our ability to build data centers and to fill them with hardware.”
Feldman explaining that AI compute customers like OpenAI and Anthropic are so desperate for capacity they order chips before they are finished being made.
“AGI I think I suspect you'll agree with me that we've hit it. we just haven't exactly deployed it fully. We have artificial general intelligence now.”
Feldman making the direct claim that AGI has already been achieved, framed against any definition that would have been used 20 years ago.
“powerful recursive gains are exponential, right? you get better, you do it again. And if you continue to get gain, the slope of that curve is so steep.”
Feldman explaining the mathematical logic behind 'loop maxing' and why recursive self-improvement is the path to super intelligence.
All-In is a podcast hosted by Silicon Valley veterans Jason Calacanis, David Sacks, Chamath Palihapitiya, and David Friedberg. The hosts debate the biggest stories in tech, venture capital, politics, and markets. This episode features Andrew Feldman, CEO of Cerebras Systems, discussing the AI infrastructure buildout and the state of reasoning models.
Takeaways
1
Fast inference chips make 48-hour reasoning runs viable Extended reasoning runs — where a model deliberates for 25 to 48 hours — produce qualitatively different and vastly better outputs, not just incrementally better ones. Cerebras's speed advantage (cited as ~15x faster) means what would take weeks of inference time on slower hardware becomes practical in a single day. This makes 'token maxing' and 'loop maxing' economically and operationally feasible for real workloads.
2
Open-source AI sovereignty is now a board-level decision Regulated industries — finance, healthcare — are increasingly choosing open-source models deployed on-premises to avoid data leakage and dependency on frontier model providers. Feldman notes that in practice the only meaningful open-source options today are OpenAI's OSS 12B or Chinese models (GLM, Kimi, Qwen), creating a gap that domestic US open-source efforts need to fill. The government standoff over Anthropic's staged model release accelerated this conversation, especially in Europe.
3
Prompt engineering is dying; intent understanding is here Models like OpenAI's Fable and o3 (referred to as '56') are increasingly inferring user intent without requiring precise prompt construction, marking a fundamental shift from earlier generations where small wording changes dramatically altered outputs. Feldman and Calacanis both observe that the AI now proactively suggests better framings, additional charts, or follow-up questions the user didn't think to ask. Adding explicit instructions like 'check your work and tell me what I haven't considered' dramatically improves output quality.
4
Cerebras sitting on $25B backlog Feldman disclosed that Cerebras currently has a $25 billion demand backlog, and stressed this is not unique — every major compute provider faces similarly insatiable demand that outstrips supply. Customers are placing chip orders before fabrication is complete, meaning the industry is chasing demand that already existed rather than speculative future use.
5
AI data centers now rival cities in power draw Individual AI data center buildings are consuming more power than midsize cities, and collectively the next few years of buildout will consume more energy than the previous 50 years of global compute combined. This physical scale is unlike anything in modern technological history. Nations as unexpected as Kazakhstan, Tajikistan, and Armenia are now racing to build large-scale data centers.
6
Staged model releases are reasonable, like pharma trials Feldman argues that asking AI labs to roll out sufficiently powerful models in stages — giving governments time to red-team and patch infrastructure — is no different from phased pharmaceutical trials. He specifically references Palo Alto Networks CEO Nikesh Arora finding previously unknown critical vulnerabilities in hours using a frontier model, underscoring that the threat surface is real. The problem is that political polarization is preventing a rational policy conversation about this.
7
Vertical chip integration is about sovereignty, not speed Hyperscalers like Amazon (Trainium), OpenAI (Jalapeno), and Google building their own chips is less about outperforming Nvidia and more about avoiding dangerous single-vendor dependency — a lesson learned from the x86/Intel era. Feldman argues you don't need the fastest chip; you just can't be entirely dependent on another company's silicon for a core strategic input. This same logic applies at the nation-state level with data center buildouts.
8
AGI is here by any prior definition, say hosts and Feldman Both Feldman and the host agree that current models have surpassed every benchmark — including the Turing Test — that researchers and science fiction authors would have called AGI 10–30 years ago. The meaningful frontier has shifted: the race is now explicitly toward super intelligence, driven by recursive self-improvement loops. The implication is that debates about 'when AGI arrives' are already obsolete.