Here's a roundup of AI news for November 4 (JST). The biggest story of the day is OpenAI's revelation that it designed hardware in under a week using GPT-6 Astra. Other highlights: AI agents are now running operations at SETI's telescopes, Palantir reported earnings, Gartner analyzed the cost advantage of open models, and Apple moved to regulate AI companions.
GPT-6 Astra Designs Hardware in Under a Week — OpenAI Accelerates Engineering
OpenAI's hardware team has revealed that it used GPT-6 Astra to design a device for Codex, its coding agent, in under a week. The model handled everything from the enclosure to the circuit board and firmware, and the prototype went through several design revisions before reaching production. The main driver of those revisions was reportedly manufacturing tolerances — the model didn't always fully account for the variability introduced during manufacturing.
The company also said that a newer internal model is scoring substantially higher on its own engineering benchmarks, though it hasn't disclosed details. In the accompanying official article, "Accelerating engineering for the physical world," OpenAI describes how ChatGPT and Codex are accelerating physical-world engineering work such as CAD, circuit board design, and simulation. LLMs that built their track record in code generation are now concretely expanding into the automation of design work.
AI Agents Begin Operations at SETI's Allen Telescope Array
The SETI Institute and Breakthrough Listen have started using AI agents to control telescopes and analyze radio signals at the Allen Telescope Array (ATA). Work that once took weeks now takes hours, and according to the team, this marked the first real-time detection of a pulsar in the Crab Nebula in history.
The search for extraterrestrial intelligence is an 'extremely low-hit-rate' endeavor: sifting through enormous volumes of observational data to surface only the promising signals. That agents are now taking on the search itself looks like a turning point in AI's role in scientific observation — shifting from assistance to operations.
Palantir's Q3 Earnings: a 'Blowout,' but Signs of Slowdown in Enterprise Revenue
Palantir's Q3 results were strong enough that the company's own press release declared it had 'crushed consensus estimates.' Yet the stock slipped modestly on Tuesday, and investor enthusiasm was far from unanimous. The reason: early signs of a slowdown in revenue from enterprise customers, a key growth pillar.
The 'Q4 guidance signaling decelerating growth' covered in our previous roundup has now been borne out by the actual results. While AI demand itself persists, this can also be read as a signal that expectations for individual stocks may have run too far ahead.
Open Models Handle 90% of Agent Workloads at 1/20th the Cost — Gartner Analysis
A Gartner analysis reportedly concludes that 'open models can handle 90% of agent workloads at 1/20th the cost of frontier models.' One real-world case drawing attention is UK fintech Revolut. On Reddit's LocalLLaMA community, a report that the company cut its AI coding costs to one-tenth by dropping Claude in favor of self-hosted open models has been making the rounds.
This isn't a story about frontier model performance becoming worthless. But the cost structure of the 'most powerful model for every task' approach is showing its limits, and matching models to tasks looks increasingly like the pragmatic choice for enterprises.
Apple Restricts Dating Apps and AI Companions to Users 18 and Older
Under new App Store rules, Apple will impose an 18+ age restriction on dating apps and AI companions that include sexual or romantic content. With the AI companion market expanding rapidly and concerns mounting over their impact on minors, the platform's line-drawing has become concrete. The change could also directly affect the revenue structure of 'AI lover' services.
Inference Startup Fal Doubles Valuation to $16 Billion
Fal, a startup that provides inference infrastructure for generative AI workloads such as image and video generation, has doubled its valuation to $16 billion after raising $1.3 billion in a round led by Bessemer. The company is just five years old. It's the latest sign that infrastructure demand continues not only for model training but on the inference side as well.