Latest / Neural Newscast / Anthropic Floats Proposal to End US AI Model Ban [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior. [00:02] Nina Park: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: I'm Nina Park. Welcome to Model Behavior. Today is June 18th, 2026. Model Behavior examines [00:19] Nina Park: how AI systems are built, deployed, and operated in real professional environments. On this program, [00:26] Nina Park: we look past the speculative headlines to analyze the technical architecture [00:30] Nina Park: and policy decisions shaping the industry. We provide a measured perspective on the developments [00:36] Nina Park: that matter to developers, policymakers [00:38] Nina Park: and enterprise leaders who are integrating these sophisticated models into their core operations. [00:44] Thatcher Collins: I'm Thatcher Collins. Today we have a high-stakes set of stories ranging from a regulatory standoff [00:49] Thatcher Collins: in Washington to a significant shift in which AI models businesses are actually paying for [00:55] Thatcher Collins: in the current market. We will also look at a medical breakthrough that is giving families [01:00] Thatcher Collins: long-awaited answers using off-the-shelf models to solve complex genetic mysteries. [01:05] Thatcher Collins: It is a dense agenda today [01:06] Thatcher Collins: covering the intersection of global policy, corporate spending and advanced clinical research. [01:12] Nina Park: We begin with exclusive reporting from the New York Post regarding the ongoing friction [01:17] Nina Park: between the private sector and federal oversight. Anthropic executives are currently floating [01:22] Nina Park: a proposal to Commerce Secretary Howard Lutnick in an attempt to end a United States [01:27] Nina Park: ban on their most powerful models, known as Mythos and Fable. These models were hit with [01:32] Nina Park: restrictive export controls after specific security concerns were raised by the administration. [01:37] Nina Park: This negotiation is being led by [01:39] Nina Park: Anthropic Chief Compute Officer Tom Brown and their head of public policy, Sarah Heck. [01:44] Thatcher Collins: This is an unusual situation, Nina, because the crackdown was reportedly triggered by a warning [01:49] Thatcher Collins: from Amazon CEO Andy Jassey. Amazon researchers found evidence that it was possible to bypass [01:55] Thatcher Collins: the safety guardrails on Mythos and Fable using novel techniques. It is a bit of a paradox, [02:01] Thatcher Collins: considering anthropic CEO Dario Amadei has publicly argued that the government should have [02:06] Thatcher Collins: the authority to shut down frontier models if they are deemed too dangerous. [02:11] Thatcher Collins: Now they find themselves on the receiving end of that very authority." [02:13] Nina Park: Exactly. And the White House was reportedly irked that Anthropic initially downplayed [02:20] Nina Park: these safety flaws as narrow problems rather than systemic vulnerabilities. Currently, [02:26] Nina Park: Anthropic is pledging to work more closely with the administration and improve communication [02:31] Nina Park: to resolve these security issues. While President Trump mentioned at the G7 summit [02:35] Nina Park: that talks are going fine, the exact time table for lifting these export controls remains unclear. [02:41] Nina Park: The company claims pulling the models was the only way to comply with the controls [02:45] Nina Park: while they work on a permanent fix. It is interesting to see the friction here, Thatcher. [02:50] Nina Park: A senior administration official told The Post that they received warnings [02:53] Nina Park: from nearly half a dozen companies besides Amazon before taking action. [02:58] Nina Park: It suggests that the safety-first reputation Anthropic has built is being tested by [03:03] Nina Park: the very regulators they encouraged to be vigilant. This is not just a policy debate; it has [03:09] Nina Park: real market implications, especially given the new data on who is leading the enterprise market [03:14] Nina Park: in terms of actual adoption and revenue. That brings us to the latest Ramp AI index, [03:20] Nina Park: which provides a unique window into corporate spending. For the first time, [03:24] Nina Park: Anthropic has overtaken OpenAI in United States business adoption. According to spending data from [03:31] Nina Park: over 50,000 companies, Anthropic now accounts for 41% of businesses with a paid AI subscription. [03:38] Nina Park: OpenAI has slipped slightly to 39.5%. This shift is notable because it is not based on survey vibes [03:45] Nina Park: or sentiment. It is based on actual corporate credit card transactions and direct billing. Nina, [03:50] Nina Park: that is a staggering trajectory to witness. In mid-2023, Anthropic had almost 0% [03:57] Nina Park: of that specific market share. To see them quadruple their adoption while OpenAI has remained [04:03] Nina Park: relatively flat in the enterprise sector is a major signal for the industry. [04:06] Nina Park: It suggests that despite the regulatory hurdles we just discussed, the corporate world [04:11] Nina Park: is increasingly choosing Clode over ChatGPT for their professional stacks, [04:16] Nina Park: perhaps valuing the perceived stability [04:18] Nina Park: and safety focus that Anthropic has marketed so heavily to the C suite. [04:23] Announcer: Ramps data implies that this safety obsessed branding might actually be the winning strategy [04:28] Announcer: for businesses, even if it causes temporary friction with the commerce department. [04:32] Announcer: But while the Enterprise battle heats up, we're seeing these same models achieve [04:36] Announcer: remarkable results in specialized fields like medicine. NBC News is reporting that OpenAI's O3 model [04:44] Announcer: has helped diagnose 18 children with rare diseases at [04:47] Announcer: Boston Children's Hospital, cases that had previously baffled medical professionals for years. [04:54] Thatcher Collins: This study published in the New England Journal of Medicine AI publication is [04:58] Thatcher Collins: quite a proof-of-concept for clinical integration Researchers at the Manton Center [05:02] Thatcher Collins: for Orphan Disease Research ran the genomes of 376 patients who lacked diagnoses [05:08] Thatcher Collins: through the O3 system. They provided the model with clinician notes, symptoms [05:13] Thatcher Collins: and a list of potentially relevant genes. It yielded new diagnoses for about 5% of the cases. Nina, [05:21] Thatcher Collins: that might sound like a small number but for rare diseases, [05:24] Thatcher Collins: 5% is a massive breakthrough for families who had no other options. [05:29] Nina Park: One case highlighted was Kira Benton, who had been searching for an answer for 15 years. [05:35] Nina Park: The O3 model helped identify her condition as myofibrillar myopathy. Katherine Brownstein, [05:41] Nina Park: the lead researcher, noted that the model does not get tired and can analyze pages of genes that [05:46] Nina Park: a human analyst might take days or even weeks to process. It is about democratizing access [05:53] Nina Park: to specialized genetic expertise that is currently in very short supply [05:57] Nina Park: across the global healthcare system, allowing more families to receive clarity. [06:01] Thatcher Collins: However, we should be clear that the researchers and OpenAI themselves are not suggesting [06:06] Thatcher Collins: these tools be used by consumers to self-diagnose. The O3 model acted as a screening tool to help [06:13] Thatcher Collins: doctors re-analyze massive backlogs of data. There is still a strict trustworthiness requirement, [06:19] Thatcher Collins: as every output was reviewed by the human research team. Seven of the identified diagnoses [06:23] Thatcher Collins: were actually rediscoveries, where a treatment team elsewhere had the answer [06:28] Thatcher Collins: but had not shared it. The AI basically acted as a massive tireless librarian connecting dots. [06:35] Announcer: Speaking of technical efficiency, thatcher, a new framework called Arbor is making waves [06:41] Announcer: in the autonomous optimization space. Researchers from [06:43] Announcer: Redmond University and Microsoft Research introduced it [06:46] Nina Park: to solve [06:47] Announcer: the persistent problem [06:48] Nina Park: where AI agents repeat their own mistakes during complex engineering tasks. [06:53] Nina Park: They found that Arbor delivers more than two and a half times the gains of standard agents [06:57] Nina Park: like ClodCode or Codex [06:59] Nina Park: by implementing a more structured approach to problem solving and error correction in real time. [07:04] Thatcher Collins: Arbor uses something they call "hypothesis tree refinement." Instead of one agent [07:09] Thatcher Collins: trying to do everything, they have a coordinator who acts like a principal investigator [07:14] Thatcher Collins: and executors who are short-lived agents testing specific ideas in isolated environments. [07:20] Thatcher Collins: This stops reward hacking and ensures that if a model makes a mistake, [07:23] Thatcher Collins: that failure is recorded as a constraint so the system does not try it again. [07:28] Thatcher Collins: It is essentially teaching AI to conduct cumulative research [07:31] Thatcher Collins: rather than just relying on simple trial and error, which makes the entire process more robust. [07:36] Nina Park: This ties into a broader conversation happening in the industry about multi-agent systems. [07:42] Nina Park: We are seeing a move away from the God model toward these specialized networks of smaller agents, [07:48] Nina Park: but as a recent analysis pointed out, this introduces the token tax. When agents talk to each other, [07:54] Nina Park: it can become incredibly expensive and redundant. A framework called S squared [07:59] Nina Park: MCAD has actually been shown to reduce those token costs by over 94% by teaching agents [08:05] Nina Park: when to stay quiet and only transmit essential data to their peers. That is a vital point, Nina. [08:11] Nina Park: If every agent in a system is just a yes man agreeing with the first plausible solution, you get [08:17] Nina Park: brittle results and high operational costs. Newer protocols [08:21] Nina Park: like belief-calibrated consensus seeking are being developed to ensure agents only chime in [08:26] Nina Park: when they have something novel or a strong contradiction to offer. It's about engineering [08:30] Nina Park: the virtual boardroom to be as efficient as possible, mimicking [08:33] Nina Park: the way high-functioning human teams operate when they are under tight resource constraints. [08:38] Announcer: There is also the critical security angle to consider. [08:42] Nina Park: In a multi-agent setup, if one agent is compromised by a prompt injection, it could potentially pass [08:49] Nina Park: malicious instructions to every other agent in the network. This is leading to the development [08:54] Nina Park: of zero-trust AI architectures. This means agents will require cryptographically signed permissions [09:01] Nina Park: and internal firewalls to communicate with one another, ensuring that a single point [09:06] Nina Park: of failure does not lead to a total system compromise in sensitive corporate environments. [09:12] Thatcher Collins: It is a fascinating shift to observe. We have spent years teaching AI to speak, [09:18] Thatcher Collins: and now we are spending our time teaching it how to collaborate securely and efficiently. [09:22] Thatcher Collins: From anthropic's high-level negotiations with Secretary Lutnick to the granular logic of [09:27] Thatcher Collins: the Arbor Framework, the theme of 2026 seems to be professionalizing the environment [09:33] Thatcher Collins: where these models operate. We are moving from the era of novelty [09:37] Thatcher Collins: to the era of industrial-grade reliability and security across all sectors of the economy. [09:43] Nina Park: Thank you for [09:44] Announcer: listening to Model Behavior. [09:46] Nina Park: You can find more details in our full archive at mbnneuralnewscast.com. Neural Newscast [09:52] Nina Park: is AI-assisted human-reviewed. View our AI transparency policy at neuralnewscast.com. [09:59] SPEAKER_03: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.