Sam Altman’s AI Safety Scenarios, Arda’s $1.5B Debut, and SpaceX’s AI Revenue
This digest covers Sam Altman's three AI safety endgames, the $1.5 billion launch of Bob McGrew's Arda, SpaceX's significant AI revenue, and Anthropic's proposed AI safety legislation.
Podcast В· 4 min
Sam Altman outlines three AI safety endgames
OpenAI CEO Sam Altman recently shared three potential scenarios for the future of AI safety with staff. These include the establishment of a global watchdog similar to the International Atomic Energy Agency, a 'Demis plan' (named after DeepMind's Demis Hassabis) that merges top labs into a single entity, or a 'terrible' outcome where one lab races ahead to dominate computing power. Altman emphasized that OpenAI must choose its path within the next year, a process he termed 'pacing 2.0,' which must include defined speed limits for self-improving AI.
Ex-OpenAI research chief Bob McGrew launches Arda at $1.5B valuation
Former OpenAI Chief Research Officer Bob McGrew has emerged from stealth with Arda, a startup focused on automating factory assembly line setups. The company launched with a $1.5 billion valuation following a $160 million Series A funding round. Arda aims to do for manufacturing engineers what coding assistants do for programmers, using AI to plan builds from design files. Defense firm Anduril, Arda's first customer, reported that assembly instructions now take 15 minutes compared to the previous 10 hours.
SpaceX reports $7.5B in AI-related revenue
SpaceX has revealed that its new 'Super Intelligence' division generated $7.5 billion in revenue during the third quarter, accounting for roughly half of the company's $14.6 billion total revenue. This division encompasses Grok, the AI coding tool Cursor, and a 'neocloud' business that provides data center compute power. The company reported capital expenditures of $19.6 billion this quarter, largely driven by investments in data centers and Starship rockets.
OpenAI acquires Orchestra Bio
OpenAI has acquired Orchestra Bio, a startup specializing in AI agents that track and monitor drug development studies. The technology will be integrated into OpenAI's Rosalind life-science tools to help flag delays and ripple effects in complex drug programs.
Otter.ai secures $320M in funding
Otter.ai has raised $320 million from General Catalyst's Customer Value Fund. The financing is structured around returns on marketing spend rather than traditional equity sales, supporting the continued growth of its meeting intelligence platform.
Anthropic proposes AI safety bill
Anthropic has drafted a federal AI safety bill that would empower the Department of Commerce to request a judicial block on model releases if they pose an imminent risk of causing 50 or more deaths or $1 billion in damages. This proposal reflects the company's ongoing efforts to establish regulatory frameworks for frontier AI models.
Anthropic's Claude designs scorpion antivenom proteins
Anthropic published research conducted with Citrini Research, in which Claude designed proteins to bind to and disarm scorpion venom toxins. Out of 318,484 candidate designs generated by the AI, 211 were tested in labs, with 17 successfully binding to their targets. While the project did not result in a universal antivenom, it demonstrates the potential for AI to accelerate drug discovery.
Meta's Muse app vulnerability exposed shared chats
Security firm Claroc discovered that shared chats from Meta's Muse app were findable via public web records without the original link. The issue stemmed from the app's URL structure, which included chat titles, allowing passive DNS services to index them. Meta has reportedly paused the sharing feature while investigating the vulnerability.
ChatGPT introduces PIN-locked chats
OpenAI has rolled out a new feature for ChatGPT allowing users to secure sensitive conversations behind a PIN or Face ID. These locked chats maintain a separate memory from standard interactions, enhancing privacy for specific workflows.
DeepSeek releases DeepInfer code
DeepSeek has released the code for DeepInfer, the engine powering its V4.1-Flash AI. The release includes support for Huawei chips, enabling firms to deploy the model without relying on Nvidia hardware.
Skyfall AI unveils SKY-001
Skyfall AI has introduced SKY-001, a model designed to draft 3D designs for parts and buildings. The company claims the system is significantly faster and cheaper than existing frontier models.
Zoom adds deepfake detection
Zoom announced a new deepfake risk detection feature scheduled for December. The tool analyzes live video and audio for signs of alteration and provides a confidence indicator to users.
Spotify removes 50M+ 'slop' tracks
Spotify has removed over 50 million tracks identified as 'slop' or spam from its recommendations this year. The platform is also restricting AI-persona artists from its recommendation algorithms by default to prioritize human artistry.
Radical Numerics reports FDA clearance for AI-designed virus
Radical Numerics claims a significant milestone in AI-designed medicine, stating that the FDA has cleared a virus designed by its Omnii-1 AI for a patient with a treatment-resistant infection.
Thore Graepel debuts Metis Reasoning
Former DeepMind researcher Thore Graepel has launched Metis Reasoning, an AI startup applying AlphaGo-style techniques to chatbots to enable them to verify their outputs against evidence.