Latest / Neural Newscast / Karpathy Joins Anthropic for Pretraining Research [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior, [00:03] Nina Park: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: I am Nina Park. Welcome to Model Behavior. [00:17] Nina Park: Today is May 21st, 2026, and we are examining how artificial intelligence [00:23] Announcer: systems are being built and operated within professional environments. [00:27] Announcer: Our focus today is on a major research hire at Anthropic, and several significant updates [00:33] Announcer: to the Claw developer ecosystem that are changing how engineers interact with code. [00:39] Thatcher Collins: It has been a particularly [00:41] Nina Park: dense week for the team at Anthropic, Nina. [00:44] Nina Park: The biggest headline involves the arrival of Andrei Karpathy. [00:48] Thatcher Collins: The OpenAI co-founder and former Tesla AI director has officially joined Anthropic to lead [00:53] Thatcher Collins: a specialized team focused on pre-training research. This represents [00:57] Thatcher Collins: a significant shift in the competitive landscape for foundational model development. [01:02] Announcer: That is right, Thatcher. Reports indicate that Carpathia, who recently founded [01:07] Announcer: the AI education startup Eureka Labs, will now return to foundational large language model research. [01:13] Announcer: His extensive background at both Tesla and OpenAI makes him one of the most recognizable [01:19] Announcer: and respected figures in the field. His decision to move to a direct rival is [01:23] Announcer: a major talent acquisition. It makes me wonder how this fits with his stated mission at Eureka Labs. [01:29] Announcer: He was very vocal about moving into AI education full time last year. Joining Anthropic [01:35] Announcer: to lead pre-training suggests that the pull of frontier research is still stronger [01:39] Announcer: than the independent education path. Or perhaps Anthropic offered a unique research environment [01:44] Announcer: that allows him to pursue both? Beyond the personnel moves, we also saw the Code [01:49] Announcer: With Claude developer event in London this week. MIT Technology Review [01:53] Announcer: noted a striking statistic from the gathering. Nearly half the developers in attendance admitted to [01:59] Announcer: shipping pull requests written by Claude 4.7 without reading the code first. [02:04] Announcer: Anthropic is leaning into this level of trust with a new feature they call dreaming. [02:08] Announcer: Let us dig into that dreaming feature, Nina. According to technical documentation from Anthropic, [02:14] Announcer: it allows Claude code agents to essentially write notes to themselves [02:17] Announcer: during the development process. When a subsequent agent works on that same codebase later, [02:23] Announcer: it reads those notes to get up to speed. It's a way for the AI to consolidate patterns [02:27] Announcer: and record errors autonomously." "Exactly. The ultimate goal here is to move beyond [02:33] Announcer: the traditional model where humans are constantly prompting the AI. Instead, [02:38] Announcer: Anthropic wants Claude to prompt itself, testing and tweaking code until it runs correctly [02:43] Announcer: without human intervention. They're colloquially calling this 'letting Claude cook, [02:48] Announcer: ' with the intent [02:48] Announcer: that developers shouldn't even see the intermediate error messages during the process. [02:53] Thatcher Collins: That is exactly where I see some significant friction, Nina. If developers are shipping code [02:59] Thatcher Collins: they have not actually read and the AI is managing its own error logs internally, [03:05] Thatcher Collins: we are creating a massive transparency gap. Security researchers have already warned [03:10] Thatcher Collins: that AI generated code can introduce subtle vulnerabilities [03:14] Thatcher Collins: that humans might easily miss if they are not performing rigorous manual reviews. [03:18] Announcer: Interestingly, we are seeing AI being used heavily on the defensive side of that equation as well. [03:24] Announcer: Recent reports show that Claude Opus 4.7 is being integrated directly [03:29] Announcer: into professional bug bounty workflows. Using tools like the Kiro IDE, [03:34] Announcer: ethical hackers are automating reconnaissance and payload generation [03:38] Announcer: on platforms like HackerOne to find those exact vulnerabilities and security flaws much faster. [03:44] Thatcher Collins: It really is becoming an arms race within the same model family. To support [03:49] Thatcher Collins: this rapidly growing developer ecosystem, Anthropic also announced the acquisition of a startup [03:55] Thatcher Collins: called Stainless. Stainless specializes in technology that converts API specifications [04:00] Thatcher Collins: into production-ready software development kits across various programming languages, [04:05] Thatcher Collins: including Python, Go, and TypeScript. This move streamlines the entire integration process [04:11] Announcer: for users. That acquisition is clearly a strategic move to secure the integration layer [04:16] Announcer: of the stack. As Anthropic shifts the focus of Stalins toward the cloud platform, [04:21] Announcer: they are making it significantly easier for businesses to connect autonomous agents [04:26] Announcer: to their existing internal APIs. It's about removing the technical friction [04:30] Announcer: of manual tool building and making the deployment of these agents almost instantaneous. [04:36] Thatcher Collins: It is clear that Anthropic is moving from being a simple model provider to becoming [04:41] Thatcher Collins: a full stack infrastructure company for AI agents. Between hiring Carpathia for foundation research [04:46] Thatcher Collins: and acquiring Stainless for ecosystem integration, they are tightening their grip [04:50] Thatcher Collins: on the entire developer lifecycle [04:52] Thatcher Collins: from initial training to final deployment and maintenance in a production environment. [04:57] Announcer: It certainly remains to be seen if the industry standards security and maintenance best practices [05:02] Announcer: can keep pace with this unprecedented level of automation. As these systems become more autonomous, [05:08] Announcer: the role of the human supervisor is changing fundamentally. Thatcher, thank you for sharing [05:13] Announcer: your perspective on these developments today. It is a fascinating moment for the industry. [05:18] Thatcher Collins: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [05:24] Thatcher Collins: For more in-depth analysis of the AI landscape, visit mbnneuralnewscast.com. Neural Newscast [05:32] Thatcher Collins: is AI-assisted, human-reviewed. View our AI transparency policy at neuralnewscast.com. [05:39] Thatcher Collins: Neural Newscast is AI-assisted, human-reviewed. [05:43] Thatcher Collins: View our AI transparency policy at neuralnewscast.com. [05:48] Nina Park: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.