OpenAI’s Big Reset + A.I. in the Doctor’s Office + Talkie, a pre-1930s LLM

What It Covers

Hard Fork covers three topics: OpenAI's strategic restructuring including a renegotiated Microsoft deal, expanded Amazon partnership, and Elon Musk trial; AI adoption in medicine with physician Dr. Adam Rodman detailing tools like Open Evidence reaching 40% of US doctors; and Talkie, a research language model trained exclusively on pre-1931 text data.

Key Questions Answered

• •OpenAI's cloud strategy shift: OpenAI renegotiated its Microsoft exclusivity agreement, enabling it to deploy models through Amazon Web Services Bedrock and Google Cloud Platform. This matters because large corporate customers locked into AWS or GCP no longer need to switch to Azure to access OpenAI models. Amazon also committed a $50 billion investment, signaling OpenAI's push to displace Anthropic as Amazon's preferred frontier model provider.
• •AI subscription market bifurcation: OpenAI projects its $8/month ChatGPT Go tier will grow 36x to 112 million users in 2025, while its $20/month-plus subscriptions fall 80% to roughly 9 million. This mirrors Netflix's ad-supported tier strategy. Casual users doing basic queries won't pay $20/month, while professional power users will pay multiples of that for higher rate limits and latest model access.
• •AI scribe adoption in medicine: AI scribes — voice-to-text tools that listen to patient visits and draft clinical notes — have gone from experimental to commodity in under two years. Simultaneously, Open Evidence, a retrieval-augmented decision support tool with access to NEJM and JAMA, is used by approximately 40% of US physicians. In one 24-hour period, doctors consulted Open Evidence one million times.
• •Safe vs. unsafe AI health use: Dr. Rodman's green-light uses for patients include general health questions, preparing clinic visit questions, and reviewing wearable data summaries. Yellow-light uses include exploring new symptoms as preparation for a doctor visit. Red-light uses — never recommended — include asking AI to evaluate chemotherapy options or other nuanced management decisions, where model sycophancy can produce dangerously convincing wrong answers.
• •AI deskilling risk in medicine: A Polish study found doctors using AI polyp-detection technology for three months lost six percentage points of unaided polyp-detection accuracy. This raises concern that medical trainees who learn alongside AI may never develop baseline diagnostic skills. Harvard Medical School identifies this as its primary near-term AI concern, since medical education depends on making supervised mistakes to build clinical judgment.
• •Talkie forecasting methodology: Researchers built Talkie, a language model trained on roughly 240 billion tokens of pre-1931 text, to establish a decades-long forecasting track record before trusting machine predictions. The plan is to ask it to predict events five to ten years ahead of its knowledge cutoff, verify accuracy against history, then eventually apply the same validated scaffold to present-day forecasting questions with measurable confidence levels.