Podcast Crossover: AIE, AGI, frontier lab strategy with ​ ⁨@matthew_berman⁩ and @swyxtv

Building AI Engineer from scratch

Swyx launched AI Engineer after watching front-end, cloud, and data engineering professionalize into their own fields with dedicated conferences. He bought the AI.engineer domain, wrote a blog post, and got a crucial endorsement from Andre Karpathy. The first conference required convincing people to take a leap of faith on an unproven event, but now the value proposition is clear: all the frontier labs competing on neutral ground, which benefits engineers and labs alike.

Etched and the inference chip wars

The discussion turned to Etched, the inference chip startup that just came out of stealth. Swyx framed them as next-generation Cerebras or Groq, optimized specifically for post-Transformer workloads that didn't exist when older chip companies started. The risk of custom chips is that model architectures could shift, but Swyx noted that GPT-style architectures have remained remarkably stable, making dedicated inference chips a reasonable bet.

Fable returns and the compute question

Claude's Fable model is back with an Easter egg hidden on the AIE website. Swyx hasn't experienced the rumored nerfing or model rerouting that users have complained about on X, though he acknowledged false refusals remain annoying. The bigger issue is speed: Fable is so slow that you only use it for the hardest problems, which Swyx suspects is exactly how Anthropic wants it. He pushed back on the narrative that Anthropic was compute-limited, arguing the delays were genuinely safety-related.

OpenAI's government equity offer

On the rumored OpenAI offer of 5% equity to the US government, Swyx said it tracks with OpenAI's relatively friendly stance toward the current administration. Drawing on his Singapore background, he noted that government stakes in strategically important companies are normal elsewhere and could help address the problem of a permanent underclass missing out on AI upside. Treating AI as a utility this early would be premature, he argued, comparing it to regulating electricity in Edison's era.

P(doom), timelines, and recursive self-improvement

Swyx put his p(doom) at around 90% on a 50,000-year timeline, but only 5% on a 50-year timeline. He argued that humanity has no inherent right to exist and that birthing a new life form carries real risks. On whether LLMs alone can achieve recursive self-improvement, he said no: current models mostly explore things humans have already explored, and true unknown-unknown discovery requires something beyond today's paradigm.

The efficiency problem and the limits of scaling

Data efficiency is the next frontier. Humans learn on millions of examples to become capable adults, while models require trillions of tokens. Swyx called this a 'sour lesson': every time we try to make human analogies to machines, we fail, but we still know the current approach is wildly inefficient. He expects Fable to mark the end of this LLM era if a 20-trillion-parameter model is the ceiling, because slowness and cost make it unusable for most tasks.

The Agent Lab playbook for founders

For founders, Swyx offered two words: Agent Lab. Don't pick a solution, pick a problem. Be the AI layer for dentists, lawyers, or finance professionals, and fold in whatever new capabilities emerge. The labs won't do the unglamorous work of custom integrations and customer support. On model routing versus going all-in on one provider, Swyx was skeptical: routing means you never fully exploit any model's capabilities, and the big wins come from going deep, not staying superficial across a hundred models.