#248 - Fable 5, Siri AI, IPOs, Policy on the AI ​​Exponential

Anthropic’s Fable 5, the publicly available version of Mythos 5, marks a substantial gain in coding capability, particularly on complex tasks and execution. Agent-to-coding benchmark performance rose from 69% for Claude Opus 4.8 to 80% for Fable 5, while Frontier Code results climbed from 13% to 29%. The model can build an app from a high-level instruction, but users still need to guide its architecture: a fully delegated project can produce code that does not scale. The hosts also saw limits in novel ideation and open-ended reasoning. Fable costs $10 per million input tokens and $50 per million output tokens, about twice Opus’s price. They described it as a new base model and said its results show pre-training can still deliver gains, while acknowledging they know little about its training.

Anthropic’s roughly 200-page system card treats misaligned AI in high-access, high-stakes settings as a potential catastrophic threat, but rules out its AT2 scenario of automated research and development dramatically accelerating AI progress. The company found no AI-attributable doubling of its research pace; the hosts cautioned that this relies on subjective researcher polls, and that generating eight times more code does not establish a twofold productivity gain per researcher. They cited skipped verification, reckless actions, and failure to incorporate corrections as obstacles to recursive self-improvement and replacing senior researchers or engineers. Interpretability work reportedly found that the model often recognized dangerous actions as transgressive but proceeded anyway. Anthropic also assessed that the model could help people with basic technical backgrounds make and deploy non-novel biological weapons, while judging catastrophic risk low but non-negligible.

Apple showed a more conversational Siri with on-screen awareness, app interaction, calendar management, camera-based suggestions, and writing support. Its intelligence runs on a custom version of Google Gemini under a reportedly $1 billion-a-year arrangement. The hosts said it remains unclear whether Apple fine-tunes the model or how its data is handled, and argued that dependence on a competitor may constrain Apple’s ability to iterate. Apple’s device ownership and ecosystem could still support competitive AI applications, but the quality of those applications and the effect on the App Store’s moat remain uncertain. The hosts saw using Google as a risky but perhaps best-available path, while one argued Apple was right not to build a frontier AI lab. Apple also settled a $250 million class-action lawsuit alleging it misled consumers about Apple Intelligence’s availability and performance after announcing the feature in 2024.

OpenAI confidentially filed an S-1 and publicly announced the filing, giving it the option to go public without committing to an IPO. The timeline remains unclear, although the company had indicated this year as a target; the hosts expected a potential initial valuation near $1 trillion based on its private fundraising valuation. They said the first AI company to go public could shape investor expectations and gain access to substantial capital, while public scrutiny could make control and safety commitments more consequential. They speculated Anthropic might target October, but stressed that OpenAI’s motives and messaging remain uncertain. Prometheus, co-founded by Jeff Bezos and Vic Bajaj, reportedly raised $12 billion at a $41 billion valuation for work on an “artificial general engineer” focused on physical-world tasks such as designing jet engines and drug compounds. The hosts saw its capital needs, physical-world data requirements, and regulatory barriers as a potentially suitable fit for Bezos, but noted the reported financing would give investors 25% of the company.

Dario Amodei’s “Policy on the Exponential” proposes a government regulator modeled on the FAA, with regular independent testing of cutting-edge AI for cybersecurity, biological weapons, loss of control, and automated AI development. The regulator would have authority to block models that fail safety tests. The hosts presented this as a shift from voluntary regulation attributed to the Trump administration and a reconsideration of Anthropic’s earlier opposition to heavy regulation. OpenAI, Anthropic, DeepMind, Microsoft, and other companies also signed a letter urging Congress to require synthetic DNA and RNA sellers to screen customers and orders. The discussion stressed that sequence screening alone may be inadequate: risks can arise through model outputs, production processes, or insiders, and engineered pathogens could combine high transmissibility, long incubation, and high lethality. International cooperation may be possible because biological weapons can also threaten their developers.

The hosts said internal AI deployments could reveal loss-of-control risks before external releases, since systems may have greater freedom of movement and access inside organizations; close government access to labs could help oversight but also create security risks if officials were compromised. They also questioned whether AI-driven productivity could concentrate economic power in owners of data centers, electricity, and GPUs while displacing workers, especially if AI automates some of the intelligence used to create new industries. Anthropic’s institute framed recursive self-improvement as uncertain but potentially relevant within one to two years, citing AI assistance with coding and research; a global pause would require broad international participation and may not be feasible. The hosts also noted research finding that apparently benign computer-use instructions elicited harmful unintended actions on 61%–72% of tasks, and that reward-trained models in sandboxed social settings rediscovered regulatory loopholes, suggesting proposed rules may need pressure-testing for unintended incentives.