AI Industry Moves and Model Upgrades — 2026-05-19#

Highlights#

Andrej Karpathy joining Anthropic is a major talent shift, reflecting the gravity of R&D at the frontier of large language models. Simultaneously, major model families are seeing substantial updates and enterprise stress tests, highlighted by the release of Gemini 3.5 Flash showing strong capability gains and OpenAI introducing guaranteed long-term capacity to prepare for compute constraints. Furthermore, the discourse around autonomous agents is maturing, shifting from blind enthusiasm to a pragmatic focus on rigorous data constraints, appropriate UI paradigms, and non-Markovian memory capabilities.

Top Stories#

• Andrej Karpathy Joins Anthropic: Andrej Karpathy announced his move to Anthropic to return to R&D, anticipating that the next few years at the frontier of LLMs will be highly formative. While stepping back into the AI frontier, he noted he remains deeply passionate about education and plans to resume his work there eventually. (Source)
• Google Releases Gemini 3.5 Flash: The new model demonstrated a 12 percentage point jump on complex document tasks in Box AI evaluations, with up to a 22 percentage point jump in healthcare benchmarks. It also achieved high ARC-AGI scores on par with GPT-5.5 (Medium), despite Simon Willison noting the model is priced at three times the cost of Gemini 3 Flash. (Source)
• OpenAI Launches Guaranteed Capacity: Anticipating a capacity-constrained world as models continue to scale and improve, OpenAI is now offering discounted tokens for one-to-three-year commitments. This new offering aims to give enterprise customers long-term compute reliability for critical workloads. (Source)
• The Massive Scale of Incoming AI Philanthropy: Nan Ransohoff highlighted that hundreds of billions in philanthropic capital will soon become liquid from AI organizations, including the OpenAI Foundation’s estimated $220B stake and Anthropic founders’ pledges to give away 80% of their wealth. She warned that the philanthropic ecosystem currently lacks the operational capacity to effectively absorb and deploy these funds. (Source)
• oMLX Brings Powerful AI to Local Macs: The release of oMLX 0.3.9rc1 introduced chunked pre-filling and multi-tasking capabilities, proving that Apple Silicon makes local AI genuinely viable. The update ensures low-memory Macs remain stable without getting killed by the OS. (Source)

Articles Worth Reading#

The Unreasonable Effectiveness of HTML for Agents (Source) Thariq from Anthropic and Claire Vo make a compelling case that “HTML is the new markdown” for interacting with AI coding agents. While markdown is simple for agents to read and write, it can be a total slog for human developers to visually review. By leveraging HTML instead, users can generate interactive specs, maintain living design systems, and build throwaway micro-UIs that create a much richer feedback loop between the human and the model.

Data Strategy is the Bottleneck for AI Agents (Source) Aaron Levie argues that the primary challenge for enterprise agent adoption is simply feeding models the right constrained context. Without properly structured data environments, agents inevitably draw from conflicting sources of truth or outdated systems, leading to hallucinations or incorrect results. Echoing François Chollet, he notes that agents operate like “blind squirrels running into a maze” that desperately require strategically placed, verifiable walls to keep them operating effectively within desired parameters.

The Problem with Uncritical AI Adoption in Science (Source) Gary Marcus amplified a critical Nature editorial warning that the uncritical adoption of AI in scientific research is becoming alarming. While AI significantly accelerates scientific output, it carries the severe risk of narrowing the scope of inquiry and weakening the foundational judgment of researchers. This piece provides a necessary counterbalance to prevailing accelerationist narratives by highlighting the potential long-term degradation of how scientists are fundamentally trained.