Sam Altman: Singularity Slow-Down, Emad Runs 18 Grokbots, Waymo Slashes Hardware 83% | EP #283

Sam Altman said he was wrong to expect AI’s impact to arrive as quickly as he predicted after GPT-4 in 2023. He now expects a rising-tide transition, with economic inertia, institutional lag and people’s difficulty adapting slowing the effects of improving technology. Selim described the singularity as the point when institutions can no longer adapt to machine capability and said that breaking point is near. Alex argued that familiar interfaces can mask radical change, while faster progress depends on vertical integration; he noted OpenAI dropped its plan to own Stargate data centers in favor of leasing them.

Emad said models have only recently become capable enough for practical institutional use, citing improved code generation and competent intelligence becoming available through familiar tools such as iMessage. Grokbot launched in early beta on August 11, giving agents access to a cloud computer, browser and terminal, with specialist agents coordinated by a chief-of-staff agent. Emad said he was running 18 Grokbots across computers and subscriptions to operate systems, install and test models, and optimize them; he cited a 76% performance improvement for Alibaba’s 27B model on an NVIDIA 5090 at 64K context. Dave said he had approved 100,000 swarm agents, and a speaker predicted people could coordinate 50,000 agents in parallel within three months.

The speakers said large agent fleets will need ways to report activity and measure productivity, but disagreed about the right management model: Alex argued that messaging interfaces and human-style org charts will not scale, and that agents should manage other agents. More broadly, agents are shifting work from execution to judgment. A founder who once managed five tasks may now review 50 agent outputs, creating more need for direction, checking and decisions. Selim warned that unclear approval rules could leave people supervising agents around the clock. Emad said human attention is limited, and described turning off his research agents’ active workflow so he could review their output weekly while focusing on launches and fundraising.

Chinese open-weight models are putting pressure on US competitors’ pricing, though the speakers cautioned that rankings vary by task, cost, speed and context use. One comparison put Fable 5 at $15 per million tokens and a Chinese model at $0.14, with the latter said to deliver about 80% of Fable 5’s capability. Another placed Fable at 60 and a new GLM model at 57 on the Artificial Analysis benchmark, with GLM costing 100 times less. The speakers also discussed Moonshot AI’s Kimi Linear, which it said cuts context-memory use by 75%, enables six-times-faster decoding and supports a one-million-token context window. They described such architectures as attempts to reduce the memory burden of standard attention.

Poolside’s Emad said the company moved from coding models toward Laguna, an open-source model factory, and built a strong model for its size. After an earlier attempt to raise $2 billion for a Blackwell cluster failed, Nvidia committed $6 billion toward open-weight models and inference infrastructure with Poolside. Alex said Nvidia had previously invested about $1 billion and is non-exclusively licensing key Poolside intellectual property. Emad described the arrangement as an early step toward a broader open-source stack and said Poolside Infrastructure Company is building a 1.2-gigawatt data center that may need GPUs. The speakers saw Nvidia’s move as a way to expand across the AI stack and build demand for its GPUs; Alex questioned whether the part of Poolside not joining Nvidia would remain viable.

Waymo said its sixth-generation autonomous-driving hardware reduces cost from $115,000 to $20,000, an approximately 83% cut. The system uses a custom 5-nanometer chip rated at one quadrillion operations per second, 13 cameras, four LiDARs and six radars, with 42% fewer sensors than its predecessor; Waymo said it performs better and integrates heaters, wipers and sprayers into sensor pods. The purpose-built Zeekr robotaxi was put at $75,000, compared with an estimated $300,000 for a Gen 5 Jaguar equipped for autonomy. Speakers characterized Waymo’s use of Chinese-made vehicle hardware as vertical integration, but one preferred Western companies to use a Western hardware stack. Alex predicted Waymo might eventually make or white-label more hardware of its own while concentrating on software and hosted autonomy.

The speakers also questioned whether Anthropic can sustain its growth without secure, long-term compute access. Dave said its former data-retention policy sent enterprise data to Anthropic for 30 days even when Fable 5 ran in a secure Amazon Bedrock environment, prompting concern about proprietary information and pushing some customers toward Chinese models or other secure options. Anthropic has since changed its policy to let enterprise customers keep data on their own cloud infrastructure. The speakers differed on an IPO: Emad argued Anthropic should remain private, raise about $120 billion and pursue AGI, while Dave said delaying a public offering could send a damaging signal. They also cited a report that roughly 40% of Anthropic revenue comes from models hosted by third-party cloud providers.