Andrej Karpathy Joins Anthropic's Pretraining Team
Andrej Karpathy — one of OpenAI's founding members, former Tesla AI director, and the researcher widely credited with popularising "vibe coding" and the concept of Software 3.0 — has joined Anthropic's pretraining team. The move was announced on May 19, coinciding with Google I/O, and was confirmed through both Fortune and multiple AI community sources.
Karpathy will work with Anthropic's pretraining team led by Nick Joseph, focusing on leveraging Claude to accelerate large-scale training operations and foundational research. He is perhaps best known publicly for his educational work — the Neural Networks: Zero to Hero course, his advocacy for LLM-native development workflows — but his core expertise is in large-scale model pretraining and the infrastructure required to run it efficiently. His "Karpathy Loop," an automated research methodology that reduces training iteration time through self-directed optimisation, is reportedly one of the capabilities Anthropic wants to apply more broadly to its model development pipeline.
The move is significant for several reasons. Karpathy left OpenAI in 2023 and then Tesla's AI team before that; his decision to join Anthropic rather than return to OpenAI or start his own effort is a statement about where he sees the most interesting pretraining research happening. For the developer community, it suggests Anthropic is investing heavily in the foundations of future Claude models rather than purely in application-layer capabilities. Karpathy's public voice on what makes good pretraining data and what matters for model capability will likely carry more weight once he has direct visibility into Anthropic's methods.
Read more — Fortune
Google Shuts Down Project Mariner, Folds Technology into Gemini
Google has officially discontinued Project Mariner, its experimental AI agent that navigated the web by processing browser screenshots and interacting with UI elements in Chrome. The shutdown ended a 17-month project that was among the most ambitious attempts to build a general-purpose web automation agent using vision-based interaction rather than structured API access.
Project Mariner's approach — repeatedly capturing the browser viewport, reasoning about UI state, and generating cursor movements and clicks — was computationally expensive and slow relative to the emerging alternative of agents that operate directly at the code or API level. At I/O, Google confirmed that Mariner's core capabilities will be absorbed into Gemini Agent and Chrome's upcoming "auto-browse" features, reframed around a faster architecture that uses Gemini's understanding of web structures to interact more directly with page semantics rather than pixel-by-pixel vision processing.
The shutdown is a meaningful data point for the developer community building web automation tools. Vision-based browser automation — which several startups have been building commercial products around — faces a structural efficiency disadvantage versus approaches that combine a capable model with lightweight DOM access or API-level integration. Google's decision to pivot away from Mariner after 17 months signals that even with frontier model capabilities, the vision-only approach isn't the winning architecture for production web agents. Teams evaluating automation frameworks should note this shift: the momentum is clearly toward agents that understand web structure (HTML semantics, accessibility trees, structured APIs) rather than those that treat the browser as a camera feed.
Read more — AndroidHeadlines
DeepMind Co-Scientist: Multi-Agent AI for Scientific Research, Published in Nature
Google DeepMind has published Co-Scientist, a multi-agent AI system that autonomously generates, evaluates, and refines scientific hypotheses, in a Nature paper co-authored with researchers from Stanford and MIT. The system represents one of the most rigorous applications of multi-agent architecture to domain-specific research tasks and provides concrete evidence for how the pattern accelerates complex knowledge work.
Co-Scientist coordinates four specialised agents — Generation, Reflection, Ranking, and a Supervisor — through what the paper calls an "idea tournament" mechanism. The Generation agent proposes candidate hypotheses; Reflection evaluates them against existing literature and flags weaknesses; Ranking uses pairwise comparison to score competing hypotheses; and the Supervisor orchestrates the cycle, triggering additional rounds until convergence. Crucially, the system integrates directly with scientific databases — AlphaFold for protein structures, ChEMBL for bioactivity data, UniProt for protein sequences — allowing agents to verify hypotheses against real experimental data rather than relying solely on model-internal knowledge.
Real-world deployments produced actionable results: the system identified drug-repurposing candidates for liver fibrosis that were subsequently validated in wet-lab experiments, and proposed RNA-based therapeutic approaches for ALS that are now in early evaluation at Stanford. In both cases, the system compressed months of manual literature review and hypothesis generation into days. The paper includes a detailed description of safety mitigations — custom classifiers to prevent misuse for CBRN (chemical, biological, radiological, nuclear) research, scope constraints on hypothesis types the system will generate — which is a useful reference for developers building domain-specific research agents where misuse guardrails are required.
Read more — Google DeepMind
Musk vs Altman: Jury Dismisses All Claims, OpenAI Structure Preserved
A California jury has unanimously dismissed all claims in Elon Musk's lawsuit against OpenAI and Sam Altman, ruling that the statute of limitations had expired. Musk filed the suit in 2024 alleging that OpenAI had violated its original non-profit mission and that Altman had engaged in deceptive conduct during his tenure — claims the jury did not reach the merits of, finding instead that Musk had waited too long to bring them under California law.
The verdict preserves OpenAI's current corporate structure — a non-profit parent controlling a capped-profit commercial entity — without judicial determination of whether that structure is consistent with the original mission. Musk's team has indicated he will appeal, challenging both the statute of limitations ruling and seeking review of the governance questions that the jury never addressed. A successful appeal that reaches the merits would reopen fundamental questions about what duties OpenAI's board owes to its original non-profit purpose, and whether the commercial entity's relationship with Microsoft violates those duties.
For developers and organisations building on OpenAI's APIs, the immediate practical outcome is continuity: the commercial structure that enables OpenAI to raise capital and sign enterprise agreements is not under active legal threat. The longer-term governance question — how a non-profit lab with significant commercial revenues and a capped-profit structure is held accountable to its stated mission — remains unresolved and will likely resurface in regulatory contexts as AI companies grow in scale and influence. The trial produced a detailed public record of OpenAI's internal decision-making that researchers, regulators, and competitors will be studying for some time.
Read more — NPR
Links & Sources
You May Also Enjoy
