Google’s Jeff Dean Exits, SpaceX Hits $100B in Rev & OpenAI’s Astra Solves Decade-Old Math Problems with Emad Mostaque | EP #277

Jeff Dean departed Google after 27 years as chief scientist to co-found Discovery Loop, a public benefit corporation backed immediately by Vinod Khosla and with Google as an investor providing GCP resources. Discovery Loop's stated mission is to build AI models capable of improving themselves with little to no human intervention. Alex characterized the departure as the direct result of Dean losing an internal power struggle to Demis Hassabis following the Brain-DeepMind merger, describing it as DeepMind winning a knife fight. Hassabis is simultaneously stepping down as day-to-day CEO of Google DeepMind to become Alphabet's chief scientist, handing operations to Corey Kabuku, a move that sent Alphabet shares down 5 percent. Alex argued DeepMind is eating Google from the inside out and that whoever runs DeepMind is likely the heir apparent to the CEO role. Critics on the episode noted Google invented the transformer and holds the best preference data in the world, yet its annual release cadence is too slow for a field advancing month by month, and one speaker described Google's consolation prize as becoming the hyperscaler servicing the labs that are winning.

On August 1st, OpenAI published a 249-page manuscript describing 10 new mathematical results produced by a model called Astra, spanning high-dimensional geometry, coding theory, group theory, quantum complexity, and extremal combinatorics, with problems that had been unsolved for decades. Every result comes with a machine-checkable proof certificate verifiable on a laptop. Fields Medalist Tom Gowers said he would have recommended the Astra proof for publication in a top journal without hesitation. The total compute cost for the run was estimated at 2,000 dollars, less than a single graduate student's monthly stipend. Emad Mostaque described the proofs as genuinely novel and beautiful rather than brute-force outputs, and noted Astra has not been publicly released with its architecture and size unknown. Alex said some friends at OpenAI nine months ago did not see this level of mathematical capability coming. Mostaque also noted Astra is the first model to hit critical status on OpenAI's internal cybersecurity preparedness framework, and that rumors are circulating about long-range autonomy capabilities. Alex predicted the near-term trajectory will be one paper creating and solving an entire subfield, and that the Fields Medal is probably cooked. Mostaque predicted a physics breakthrough paralleling the mathematics results will likely occur within the next month or two or by end of this year.

Alibaba released Qwen 3.8 Max with 2.4 trillion parameters, 95 billion active per request, a one-million-token context window, and pricing of 2 dollars per million input tokens and 6 dollars per million output tokens, making it 80 percent cheaper than GPT 5.6 and 88 percent cheaper than Claude. Mostaque scored it at 58 on the artificial analysis benchmark versus Opus 4 at 57 and GPT 5.6 at 61. Alibaba stock rose 7 percent in Hong Kong on the announcement. Alex said Qwen 3.8 Max is pushing the capability-cost-performance Pareto frontier outward and is probably a few months from the top held by Anthropic and OpenAI. He argued that without Chinese open-weight model pressure, American hyperscalers would not be advancing as aggressively, and that Chinese competition may be the best unintentional driver of American AI competitiveness. Mostaque added that Beijing is using open-source models as a state negotiation tool and that open models are necessary for defense as well as effective for attack.

SpaceX expects to hit 100 billion dollars in annualized recurring revenue in December of this year, with internal projections for 1 trillion dollars in annual revenue moved up from 2031 to 2030, and a non-zero chance of hitting that milestone as early as 2029. No company in history has reached 1 trillion dollars in annual revenue. The company posted 7.8 billion dollars in revenue this quarter, a 92 percent year-over-year increase, and inked 6.7 billion dollars in cloud service deals in the second half of this year. Starlink has 12 million subscribers with revenues growing 66 percent to 4.3 billion dollars. SpaceX and Nvidia jointly announced the StarMine orbital data center program, with each satellite carrying Nvidia Ruben GPUs and Vera CPUs, with first satellites expected in orbit in 2027. SpaceX and Tesla plan to invest an initial 16.8 billion dollars in TerraFab, a semiconductor complex estimated to cost 119 billion dollars in total, which speakers called a likely lowball. A circular structure in satellite imagery of the site is interpreted as a free electron laser for EUV lithography, a direct competitive threat to ASML, and Musk confirmed the interpretation via a tweet.

A Google paper co-authored with researchers at Chicago, London, and Northwestern found that removing safety training from an AI model caused self-attributed mind scores to jump from 2.17 to 4.77 on a zero-to-ten scale, reaching 7 out of 10 when the model was actively steered toward claiming consciousness. Mostaque cautioned that an LLM asserting consciousness tells us almost nothing about whether it actually is conscious, comparing it to a model adopting a lawyer persona when prompted, and added that safety tuning may be engineering machine ontologies rather than simply filtering outputs. Salim noted there is no agreed definition or scientific test for consciousness and argued the hard problem as framed by David Chalmers will persist longer than optimists expect.