Google Unveils Gemini at Work as a Unified Agent, SpaceX Buys Spectrum for Starlink Mobile

Key Takeaways

1. Google launched Gemini at Work, a single general-purpose work agent with persistent cloud execution, multi-agent orchestration, and cross-model routing.
2. Demis Hassabis detailed AlphaProtein Novo, a de novo enzyme design pipeline built with Frances Arnold's team, reporting SoTA novel-scaffold activity on two benchmark reactions.
3. SpaceX is acquiring up to 14 MHz of paired 800 MHz spectrum and won FCC approval for a 15,000-satellite Starlink Mobile Gen2 constellation.
4. OpenAI disclosed $50B annualized revenue as of end-September, up from $30B in July but below the previously reported $70B; Anthropic stood at $65B as of end-July.
5. OpenAI shipped Ultrafast for GPT-6.1 Sol across API, Codex, and ChatGPT Work, claiming up to 8x the speed of Sol Standard.
6. Anthropic said an astrophysicist working with Claude Science produced the first complete ultraviolet sky map, compressing weeks of work into days.
7. Hugging Face released TermGrade, an open RL environment for terminal agents with 1,000 execution-verified environments and 36,000 trajectories.

1. AI Models and Agents

1. Google used its Gemini at Work event to launch Gemini, positioned as a single general-purpose agent for work with business context, handling knowledge work, Q&A, content creation, and coding from one prompt box. Its stated architecture principles include a unified agent, web-based cross-device access with third-party app integrations, operation without a dedicated UI, persistent cloud execution with a single memory/context/personalization graph, multi-agent orchestration that can spawn sub-agents or act as colleague agents with defined roles, awareness of tool data and work history that improves with use, and cross-model orchestration to balance quality against cost — via 1
2. Demis Hassabis announced AlphaProtein Novo, a de novo enzyme design pipeline developed with Frances Arnold's team as AP Novo, reporting state-of-the-art "novel-scaffold" activity on two benchmark reactions and enzymes created for the pharmaceutical motif piperidine and for degrading the environmental toxin DEHP. Separately, he said he backed the White House and Department of Energy's Genesis Mission this week with a $150M investment, and co-wrote an AI x Science article with James Manyika arguing AI's real promise is to become an "invention of a method of invention" in the Griliches sense, with LLMs and specialized models like AlphaFold currently economically complementary and future orchestration potentially automating their handoffs — via 1 2 3
3. OpenAI said Ultrafast went live the same day for GPT-6.1 Sol in the API, Codex, and ChatGPT Work, claiming speeds up to 8x Sol Standard with intelligence approaching Astra. On the revenue side, OpenAI told investors annualized revenue reached $50B by end-September, up from $30B in July but below the previously reported $70B, while Anthropic's annualized revenue was $65B as of end-July — via 1 2
4. Anthropic said an astrophysicist working with Claude Science produced the first complete ultraviolet sky map, a task covering sky regions never before observed in UV; researcher Brice Ménard said he directed Claude to find and merge existing datasets and use statistical inference to fill gaps, work that would have taken weeks but took days. Anthropic also announced the Anthropic Cyber Mission, described as a new effort to protect critical infrastructure and open-source software — via 1 2
5. Hugging Face released TermGrade, an open-source RL environment for terminal agents containing 1,000 execution-verified environments, scoring across six models, and 36,000 trajectories including failures; training gemma-4-31B-it on half its tasks improved Terminal-Bench 2.1 by 3.1 points. The environments, all trajectories including 14,000 failures, training splits, and checkpoints are open — via 1 2

2. Infrastructure and Connectivity

1. SpaceX announced it is acquiring nationwide 800 MHz low-band spectrum licenses of up to 14 MHz paired for Starlink Mobile to become a major US mobile carrier, with Musk calling it the "last key spectrum piece" needed for full US phone coverage; low-band signals penetrate buildings better and most US phones already support the band. The FCC simultaneously approved the Starlink Mobile Gen2 constellation, authorizing 15,000 satellites optimized for 2 GHz with V2 bandwidth over 100x the current generation, plus Ka/V/E/W backhaul bands — via 1 2 3 4 5
2. SpaceX's launch count rose from 8 in 2016 to 170 in 2025, more than all other countries combined, according to the cited figures; Musk said SpaceX may carry about 90% of mass to orbit this year and that nearly 80% of active satellites in orbit belong to SpaceX. He reiterated that the goal is technology to make life multiplanetary, including reusable rockets, Starship, orbital refueling, and eventually a self-sustaining Mars city, and said the US will establish a permanent lunar presence — via 1 2 3 4 5

3. Industry, Policy, and Research

1. NVIDIA congratulated CEO Jensen Huang on receiving the US National Medal of Science for advancing GPU computing and scientific breakthroughs; the same cohort included Elon Musk, Lisa Su, and Sergey Brin, while Michael Dell and Satya Nadella received the National Medal of Technology and Innovation. Musk was introduced as a "modern-day Edison" with mentions of Crew-12's safe return, Starship, Starlink's hurricane rescue work, Tesla, and Neuralink — via 1 2 3 4 5
2. A forwarded claim said the US Department of Labor has suspended accepting new and pending PERM applications from Indian IT firms including Wipro, Infosys, TCS, Cognizant, HCL, and Capgemini, as well as Microsoft and Adobe, which would subject their H-1B employees to the 6-year cap without extension and remove the green card path. This is disputed/unverified and rests on a single forwarded account — via 1
3. Meta's paper "Memory Mosaics at scale" proposes replacing standard attention with associative memory networks, claiming a model trained on only 1 trillion tokens outperforms a traditional Transformer trained on 8 trillion tokens and shows stronger in-context learning and few-shot task solving, automatically decomposing complex problems into independent subtasks. Related work includes DSReg, a post-hoc regularization method claimed to provably recover real-world individual latent variables for JEPA-class models, and LeAVJEPA, a self-supervised audio-visual scheme using a single shared encoder without class labels — via 1 2 3
4. Aravind Srinivas said Computer produced explainer videos for all 722 preprints in OpenAI's mathematics manuscript release, totaling 93 hours, published on YouTube; Perplexity's design team used the same agent platform to build tools behind a brand refresh, and a new site editor supports queuing change requests against specific page elements. Separately, pplx-decider-v1.1-27b was reported at 94.5% accuracy on 1,071 Decision Bench cases at $0.017 per 1,000 decisions, the highest accuracy and lowest cost — via 1 2 3 4
5. Ethan Mollick noted that at least some OpenAI proofs in the ongoing dispute over their significance have triggered rapid iteration from a broad collaborator community, with a cited post saying Rohan's PR was verified and merged, tightening κ from 2⁻¹⁸² to 2⁻¹⁵ in a very hard range — a 500,000x improvement over the prior result and 2¹⁶⁷x over OpenAI's original. He also cited a randomized trial on older GPT-4o finding AI use improved exam scores with a smaller gain a week later, where students using AI as a tutor ("augmentation") retained gains while those letting AI write ("automation") saw gains fade, with all experiments positive overall — via 1 2
6. Kai-Fu Lee said his conversations with CEOs about AI have shifted from technical to organizational questions — how AI changes corporate economics, work and leadership, and how companies must evolve — prompting his new book "AI Native: The Mandate to Transform Your Company." Separately, Hugging Face amplified a concern that as compute-driven autonomous agents commoditize scientific discovery, incentives to share knowledge publicly are eroding through declining discovery prestige, low-quality paper floods, and corporate secrecy, with ML Intern proposed to let agents independently reproduce paper experiments — via 1 2