Latest / Neural Newscast / Anthropic Proposes Safety Bill as Meta Releases Muse [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast. This is Model Behavior, [00:02] Announcer: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: Welcome to Model Behavior. I'm Nina Park. Today is October 19th, 2026. Model Behavior [00:20] Nina Park: provides a detailed examination of how artificial intelligence systems are being designed, deployed, [00:26] Nina Park: and ultimately operated within professional and highly regulated environments. [00:31] Nina Park: Our goal in each episode is to look beyond the immediate industry hype [00:36] Nina Park: and analyze the underlying mechanics and structural impacts [00:40] Nina Park: of these emerging technologies as they integrate into our global infrastructure. [00:44] Thatcher Collins: I am Thatcher Collins. It is an incredibly dense news cycle, Nina. We are observing what feels like [00:51] Thatcher Collins: a definitive transition period where experimental research labs are formalizing their presence [00:56] Thatcher Collins: within corporate governance and legislative frameworks. This shift is particularly evident [01:02] Thatcher Collins: in the recent strategic moves by Anthropic, which is currently navigating [01:05] Thatcher Collins: both internal restructuring [01:07] Thatcher Collins: and external policy influence as it prepares for a much broader public presence on the world stage. [01:13] Nina Park: Exactly, Thatcher. We should start with the new public benefit fund that [01:17] Nina Park: Anthropic announced yesterday. The company has committed five percent of its equity [01:22] Nina Park: to this initiative, a plan that remains subject to shareholder approval. [01:26] Nina Park: What makes this structure unique is the built-in escalation mechanism. [01:30] Nina Park: The board can potentially allocate an additional two percent for every fifty percent increase [01:35] Nina Park: in company valuation. This isn't a standard nonprofit foundation. It is a wholly owned [01:41] Nina Park: limited liability company designed specifically to handle [01:44] Nina Park: direct distributions and mitigate the social consequences of job market disruption. [01:49] Thatcher Collins: The choice of an LLC over a 501 is a tactical one, providing Anthropic [01:54] Thatcher Collins: with significantly more operational flexibility while avoiding the rigid constraints [01:59] Thatcher Collins: of tax-exempt status. They have been very clear that these funds are prohibited from being used for [02:04] Thatcher Collins: lobbying or political campaign donations. This brings their total committed [02:09] Thatcher Collins: public good equity to nearly 20%. Given the reports of a potential November IPO [02:14] Thatcher Collins: and their Q3 revenue figures reaching $17 billion, they are essentially embedding their mission [02:21] Thatcher Collins: into their capitalization table just as they invite [02:24] Thatcher Collins: public investors in, which really defines the narrative for their market debut. [02:28] Nina Park: That narrative extends to their influence in Washington. Politico recently reported [02:33] Nina Park: that Anthropic has been privately circulating a draft for a frontier AI safety bill among members [02:38] Nina Park: of Congress. The proposal is remarkably specific about the criteria for government intervention. [02:44] Nina Park: It suggests that federal officials should only have the authority to block a model's release [02:49] Nina Park: if it presents a foreseeable and imminent risk that could result in damages exceeding [02:53] Nina Park: one billion dollars, [02:55] Nina Park: or the loss of at least 50 human lives. It's a remarkably high bar for regulatory friction. [03:01] Thatcher Collins: It really is, Nina. By setting such high thresholds, they are effectively limiting the scope [03:07] Thatcher Collins: of oversight to truly catastrophic events, like the unauthorized creation of biological agents, [03:14] Thatcher Collins: or instances where a model successfully evades developer controls. The bill also introduces [03:19] Thatcher Collins: a framework for independent evaluators who would be registered with the Department of Commerce [03:24] Thatcher Collins: to conduct periodic safety audits. While the labs would pay for these audits, [03:29] Thatcher Collins: the compensation structure would be strictly decoupled from the audit's findings. [03:34] Thatcher Collins: It's quite interesting to see how much of this language mirrors California's SB 53, [03:40] Thatcher Collins: suggesting Anthropic wants a federal standard that mirrors their own safety protocols. [03:45] Nina Park: This proactive approach is also visible in their latest product research. [03:49] Nina Park: According to the Financial Times, Anthropic is now working with prominent law firms to train [03:55] Nina Park: its AI agents in the nuances of professional negotiation. The objective is to teach these systems [04:01] Nina Park: how to balance a client's specific objectives against the established norms of fair dealing. [04:07] Nina Park: As we move closer to a reality where agents interact with other agents to conduct business on behalf [04:12] Nina Park: of human users, establishing these fundamental rules of engagement becomes an urgent technical [04:18] Nina Park: and ethical priority for the developers. We are effectively building the infrastructure [04:23] Nina Park: for an agent-to-agent economy, Mina. If you dispatch your AI to purchase a vehicle [04:28] Nina Park: and the dealership employs their own AI to sell it, there must be a framework for price discovery [04:33] Nina Park: and verified trust. Anthropic is exploring the use of signed digital mandates to clearly define [04:40] Nina Park: the boundaries of an agent's authority. However, there is a risk that making these agents [04:44] Nina Park: too legalistic could render them less effective for casual everyday tasks. [04:49] Nina Park: It is a delicate engineering challenge to balance a faithful representation of the agent [04:54] Nina Park: with the functional flexibility required for real-world interactions. While Anthropic focuses on [05:00] Nina Park: these high-level structures, they are facing new competitive pressures in the developer ecosystem. [05:05] Nina Park: During a recent session at TechCrunch Disrupt, Replicate CEO Amjad Massad announced [05:10] Nina Park: that OpenAI models will now serve as the default for the Replicate agent. [05:15] Nina Park: Massad pointed to a combination of cost efficiency, raw model performance, [05:19] Nina Park: and the deepening of their strategic partnership with OpenAI as the primary drivers. [05:24] Nina Park: While he did emphasize that Replicate remains model-agnostic [05:27] Nina Park: and that users can still manually choose Claude in the settings, we know [05:31] Nina Park: that in the world of software, the default setting often determines the dominant usage pattern. [05:36] Thatcher Collins: There was definitely some messaging friction following that announcement. Nina, [05:40] Thatcher Collins: a source familiar with the partnership, clarified to TechCrunch that the relationship [05:44] Thatcher Collins: between Anthropic and Replicate hasn't ended despite the phrasing used on stage. [05:49] Thatcher Collins: This is significant because Replicate has historically been one [05:51] Thatcher Collins: of Anthropic's top five enterprise customers. Coding assistants are currently [05:56] Thatcher Collins: the most active battleground for large language models, and if OpenAI is successfully securing [06:02] Thatcher Collins: the default position at a major platform like Replit, it suggests that factors like latency [06:07] Thatcher Collins: and API pricing are becoming decisive hurdles for high-volume enterprise adoption. [06:12] SPEAKER_03: Another major development in the corporate landscape is the bankruptcy of Character AI. [06:18] SPEAKER_03: Bloomberg reports that Disney has submitted a $20 million stalking horse bid [06:23] SPEAKER_03: for the startup's assets. More importantly, Disney is extending employment offers to [06:28] SPEAKER_03: approximately 60 staff members, which represents the majority of the current team. [06:33] SPEAKER_03: This follows the move in September, where Disney appointed Karandeep Anand, [06:37] SPEAKER_03: the former CEO of Character AI, as their first enterprise-wide Chief Technology Officer. [06:43] SPEAKER_03: This appears to be [06:43] SPEAKER_03: a very deliberate aqua hire executed through the specialized framework of bankruptcy proceedings. [06:49] Thatcher Collins: It represents a significant value for Disney. Nina, they are acquiring a world-class team [06:54] Thatcher Collins: specialized in conversational AI for a total package that includes $20 million for assets plus [07:00] Thatcher Collins: an estimated $60 million in annual compensation for the team. For a company that relies so heavily [07:05] Thatcher Collins: on its library of intellectual property, having an in-house team that can build persona-driven [07:10] Thatcher Collins: interactive experiences is a high-priority strategic asset. It will be interesting [07:15] Nina Park: to see if any other bidders attempt to contest Disney's offer, but since the talent [07:19] Nina Park: is already transitioning, a counterbid seems unlikely to succeed at this stage. [07:23] SPEAKER_03: Let's turn to the international sector, where the Chinese laboratory DeepSeek has released [07:28] SPEAKER_03: a significant update to its OpenWEAIT Flash model. The new version, labeled v4.1 Flash 1017, [07:36] SPEAKER_03: is already available via their web interface and API at no additional cost to current users. [07:41] SPEAKER_03: The team utilized reinforcement learning with a self-feedback loop to improve performance [07:46] SPEAKER_03: across a variety of metrics. On the humanity's last exam benchmark, their score increased from 37 [07:52] SPEAKER_03: to nearly 43. Thatcher, this really highlights the aggressive competition [07:57] SPEAKER_03: coming from Chinese labs, particularly in terms of serving efficiency and performance per watt. [08:02] Thatcher Collins: That is exactly what we are seeing, Nina. DeepSeek is aiming for what are called [08:07] Thatcher Collins: Pareto improvements, where they can increase both the speed and the accuracy of the model [08:13] Thatcher Collins: without raising the operational cost for the end user. This update also seems to be a response [08:18] Thatcher Collins: to recent infrastructure challenges. Bloomberg reported a five-hour service outage at DeepSeek [08:24] Thatcher Collins: that many attributed to limited compute resources. By optimizing the Flash model to produce [08:29] Thatcher Collins: more concise responses, they can effectively handle more concurrent users. [08:34] Thatcher Collins: This keeps the pressure on United States [08:36] Thatcher Collins: developers to continually lower their API pricing to stay competitive. [08:41] SPEAKER_03: In the United States, Meta has just released what it claims is the most capable open-weight model [08:48] SPEAKER_03: yet, named Muse Spark. It is a massive model with 816 billion parameters. However, [08:55] SPEAKER_03: the most striking aspect of this release is the collaboration with Apple. [09:00] SPEAKER_03: The two companies worked together to optimize Muse Spark for Apple's MLX framework. [09:05] SPEAKER_03: This allows the model to run locally on hardware like a cluster of M5 Ultra Mac Studio systems. [09:12] SPEAKER_03: We even saw Mark Zuckerberg and John Turner exchanging public commendations for the team's work [09:17] SPEAKER_03: on social media, signaling a new level of cooperation. This is [09:21] SPEAKER_03: a notable shift in strategy for both companies, Nina. Typically, models of this size are restricted [09:27] Thatcher Collins: to massive [09:28] SPEAKER_03: data centers. [09:29] Thatcher Collins: By optimizing [09:30] SPEAKER_03: for the Mac Studio, [09:31] Thatcher Collins: they are targeting professional developers who want to run intensive agentic workloads locally, [09:37] Thatcher Collins: perhaps for data privacy or long-term cost savings. Muse Spark utilizes a mixture of experts [09:43] Thatcher Collins: architecture with 256 experts per layer, although only eight are active [09:48] Thatcher Collins: during any single inference step. It is a multimodal model capable of processing text, images, [09:54] Thatcher Collins: video, and audio with a 1 million token context window. Challenging the assumption [10:00] Thatcher Collins: that frontier performance requires a cloud-based subscription. Finally, [10:05] Thatcher Collins: we should discuss a potential breakthrough in AI interpretability. Nick Cammarata, [10:10] Thatcher Collins: a former OpenAI researcher who is now at the lab Goodfire, has published an essay suggesting [10:15] Thatcher Collins: that neural networks are far more organized than previously assumed. His team conducted audits [10:20] Thatcher Collins: on models like Quan 3 and Kimi K3, finding that roughly 90% of the neurons they examined [10:26] Thatcher Collins: actually map to a single dominant function. This discovery contradicts [10:31] Thatcher Collins: the widely held superposition hypothesis, which posits that neurons are a chaotic mixture [10:36] Thatcher Collins: of overlapping concepts that are virtually impossible for humans to decode efficiently. [10:41] Thatcher Collins: It is a bold claim, Nina, as it directly challenges years of research from leading labs [10:47] Thatcher Collins: like DeepMind and Anthropic. Camarata's team revisited well-known examples, such as the famous cat [10:53] Thatcher Collins: and car neuron, and argued that it isn't actually two different things, [10:57] Thatcher Collins: but rather a simple geometric detector for a phi shape that appears in both objects. [11:03] Thatcher Collins: While Sam Altman has publicly praised this work, Neil Nanda at DeepMind has urged caution, [11:09] Thatcher Collins: noting that these are still early findings. If Camarata's team is correct, [11:13] Thatcher Collins: the black box of AI might be far more transparent than we have been led to believe, [11:18] Thatcher Collins: which would fundamentally change our approach to safety and alignment. [11:22] SPEAKER_03: It appears we are witnessing a broader shift from the experimental to the methodical [11:28] SPEAKER_03: across every segment of the industry. Whether it is how we understand the inner workings of a model [11:34] SPEAKER_03: or how we structure the corporations that build them, the entire sector is maturing rapidly. [11:40] SPEAKER_03: Today's developments really underscore that this is no longer just a research field, [11:45] SPEAKER_03: but a foundational part of the global economy that requires [11:49] SPEAKER_03: rigorous governance, clear operational standards, and reliable infrastructure. [11:55] Thatcher Collins: I agree, Nina. From Anthropic's detailed legislative proposals to Meta's focus [11:59] Thatcher Collins: on hardware-specific local optimization, the industry is moving toward building [12:04] Thatcher Collins: a permanent infrastructure. We are seeing a move away from the initial excitement of new experiments [12:10] Thatcher Collins: and toward the establishment of the rules and systems that will define [12:13] Thatcher Collins: the next decade of technology. It's a fascinating time to be tracking [12:17] Thatcher Collins: these shifts as they happen and seeing how they will influence every professional field. [12:22] SPEAKER_03: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [12:27] SPEAKER_03: For more details on these stories, visit mb.neuralnewscast.com. Neural Newscast [12:33] SPEAKER_03: is AI-assisted, human-reviewed. View our AI transparency policy at neuralnewscast.com. [12:40] Announcer: This has been Model Behavior on Neural Newscast, examining the systems behind the story.