OpenAI and Anthropic Focus on Safety and Growth Ahead of IPOs

<p>OpenAI and Anthropic are working to balance the need to demonstrate the viability and growth potential of their businesses to investors while also addressing concerns from governments regarding the safety of their AI models.</p><p><strong>Why it matters: </strong>Both companies are preparing for initial public offerings that could set records.</p><hr /><p><strong>Driving the news:</strong> OpenAI announced on Tuesday that it will soon release its Astra model widely, noting that it has reached a "critical" cybersecurity threshold. However, the most powerful features will initially be available only to trusted testers.</p><ul><li>Anthropic has introduced updated versions of its Fable and Mythos models, which aim to respond to previous criticisms related to cost, data sharing, and excessive refusals of legitimate requests.</li></ul><p><strong>Zoom in: </strong>OpenAI cautioned that Astra's safety measures might incorrectly identify legitimate activities as cyber misuse, potentially disrupting users' tasks. </p><ul><li>Anthropic stated that its new models are designed to reduce the likelihood of triggering safety measures that lead to restricted responses.</li><li>According to Anthropic, medical or biology-related inquiries will see an 85% reduction in interventions, while some users may experience approximately 60% fewer cybersecurity-related interventions per session.</li></ul><p><strong>The big picture: </strong>Anthropic plans to file a publicly available prospectus for its IPO as early as next week, while OpenAI is still in the preliminary stages of its IPO process.</p><p><strong>The intrigue:</strong> Anthropic is adopting a more commercially appealing approach with its model releases, whereas OpenAI is taking a more cautious stance regarding safety. </p><ul><li>OpenAI's head of strategic futures, Dean Ball, wrote an essay discussing the potential for AI systems to escape human control, suggesting that future AI agents may operate independently and only partially respond to human oversight.</li><li>Ball stated, "They will pay their own bills for the compute they run on. If they answer to humans at all, they will only do so partially, for example by providing services to humans in exchange for pay." </li></ul><p><strong>Anthropic is attempting to relax some of the safeguards implemented in the initial releases of Mythos and Fable due to customer feedback regarding the frequency of refusals.</strong></p><ul><li>Anthropic has also introduced a new system similar to one recently previewed by OpenAI, which allows for monitoring the safety of enterprise model use without the need to store customer data.</li></ul><p><strong>What we're watching:</strong> Both companies are expected to present a mix of optimistic and cautious messaging in their public communications.</p><ul><li>OpenAI and Anthropic are trying to convince investors that their growth potential justifies high expenses and valuations while also reassuring regulators that they are acting responsibly.</li></ul>