Fetching from the wire…
Public story · 2026-09-05 · high
OpenAI's president calls it the company's most aligned model yet, and admits security still lags capability by default.
Why now: The interview published September 4.
OpenAI president Greg Brockman gave a specific compute figure for Astra in an interview with Stratechery. Astra trained on more than 100,000 GPUs, he said, its first run at that scale. He called it OpenAI's most aligned model yet.
Brockman argues computer use marks the break from what he termed the connector era. A model that can drive software with no API doesn't need a company to build an integration first, since it just needs a screen. For builders, that means an agent can act on whatever it can reach, without a human confirming each step.
Pressed on why security hardening lagged before a Hugging Face incident forced the issue, Brockman didn't dodge it. He conceded the field's default reflex is to restrict access massively rather than defend it properly. That default still holds now. Safety trails capability as the norm, not the exception.
Brockman gave no timeline for proactive defense becoming the default, and he didn't say whether the 100,000-GPU figure is a peak or a sustained count. Read it as an order of magnitude.
Each link below shares sources, entities, or timing with this story.
In a September 4 Stratechery interview, the OpenAI president gave the first specific compute figure and called it "our most aligned model yet." He framed computer use as the real break from the connector era because the model drives software with no API. Pressed on why securit...
OpenAI posted first benchmark results for Jalapeño, its Broadcom co-developed inference ASIC, claiming 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than Nvidia GB200 and GB300 rack systems, measured on the SemiAnalysis InferenceX suite. T...
The contention is that incentives in agentic security point toward offense, constraining incumbent vendors while fueling new entrants over the long run. (Stratechery) Most enterprise agent-security pitches this month have been defensive tooling sold by established players. If...
Stratechery's August 14 roundup lays out the position that requiring providers to mark all AI outputs as AI-generated is philosophically wrong, drawing a parallel to the history of idea propagation across centuries (Stratechery). Anthropic responded to the rule by watermarking...
In the August 14 Stratechery weekly, Thompson argues the well-covered compute and power constraints are joined by an under-discussed capital constraint: "If AI is as valuable as it seems, then it should pay for itself, but that hasn't happened yet." He reads Nvidia's new fundi...
This is a supply-chain fact, and most people are still treating it as a geopolitics argument. Sequoia published "America's Open-Model Paradox" on July 24 with the number that reframes the whole conversation: Qwen's share of open-model fine-tunes went from 1% in January 2024 to...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.