Which diseases will you have in 20 years? This AI accurately predicts your risks

Key Takeaway: A modified GPT model, Delphi-2M, uses medical records and lifestyle data to estimate 20-year risk for 1,258 diseases, often matching or beating today’s single-disease predictors.

More Insights:

• Trained on 400,000 UK Biobank participants; inputs include history, age, sex, BMI, tobacco and alcohol use.

• Combines risks across many conditions, avoiding the need to run dozens of separate tools.

• Matched/exceeded single-disease models and outperformed a multi-disease biomarker algorithm.

• Works best for diseases with predictable progressions (e.g., some cancers), producing probability timelines.

• External test on 1.9M Danes showed only a slight accuracy drop; limits include recording only first disease events, with broader validation planned.

Why it matters: If reliable and generalizable, tools like Delphi-2M could shift healthcare from reactive treatment to targeted prevention—prioritizing screening and lifestyle interventions for those most at risk—while raising crucial questions about bias, privacy, and how predictions should (and shouldn’t) guide care.

Discover your next AI-powered teammate on the HubSpot Marketplace

Your perfect solution is waiting! Through intelligent recommendations, the HubSpot Marketplace makes it easy to find apps, services, all-new Breeze Agents & Assistants, and more. Search by outcome, discover trending tools, and accelerate your growth on HubSpot.

British AI startup beats humans in international forecasting competition

Key Takeaway: A British startup’s AI, ManticAI, placed 8th in the Metaculus Cup—outperforming many human forecasters—highlighting rapid gains in AI prediction while top humans still hold a narrow edge.

More Insights:

• ManticAI beat scores of enthusiasts and some professionals in a 60-question summer contest.

• Performance leap: last year’s best bot was ~rank 300; now a bot is top 10.

• Metaculus’s CEO says AI could match or surpass top humans by 2029, but humans currently lead on complex, interdependent questions.

• ManticAI decomposes problems and assigns tasks to different LLMs (e.g., OpenAI, Google, DeepSeek), often countering crowd averages—an antidote to groupthink.

• Research from Tetlock and views from Good Judgment suggest AI excels in data-rich, frequent metrics (e.g., inflation), while humans keep an edge where judgment and sparse data dominate—making human+AI teams strongest.

Why it matters: If AI can consistently rival top human forecasters, decision-making in finance, policy, and risk could shift from intuition-driven to rigorously ensemble-driven—where the winning edge comes from designing the best human–AI collaboration, governance, and verification loops rather than betting on either side alone.

OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites

Key Takeaway: OpenAI, Oracle, and SoftBank unveiled five new U.S. data center sites under Stargate, lifting planned capacity to nearly 7 GW and >$400B in investment—tracking to lock the full $500B, 10-GW target by end of 2025.

More Insights:

• Oracle-led sites: Shackelford County (TX), Doña Ana County (NM), and a yet-to-be-named Midwest location, plus a potential +600 MW expansion near Abilene, TX.

• SoftBank-led sites: Lordstown (OH) and Milam County (TX); together can scale to ~1.5 GW in ~18 months.

• Scale & jobs: Over 25,000 onsite roles expected, with tens of thousands more across supporting industries.

• Hardware & readiness: Abilene flagship is live on OCI; first NVIDIA GB200 racks arrived in June with early training/inference underway.

• Investment pace: Oracle–OpenAI agreement envisions up to 4.5 GW more capacity and a partnership exceeding $300B over five years; >300 site proposals reviewed since January.

Why it matters: This is AI industrial policy turning into concrete—literally. Massive, power-dense campuses signal a new competition axis: whoever can stand up cheap, reliable compute fastest will shape the next wave of AI capability, regional job growth, and the grid itself.

Other stuff

• Nvidia plans to invest up to $100B in OpenAI

• California issues historic fine over lawyer’s ChatGPT fabrications

• The Oakland Ballers let an AI manage the team. What could go wrong?

• ChatGPT joins human league, now solves CAPTCHAs for the right prompt

• The profile of Anthropic’s Dario Amodei

• OpenAI and Jony Ive poach Apple designers, target key suppliers for hardware push

• The AI Kids Take San Francisco

• New encryption lets LLMs answer encrypted prompts

• ChatGPT Is Blowing Up Marriages as Spouses Use AI to Attack Their Partners

Hope you enjoyed today's newsletter

⚡️ Join over 300,000 people using the Superpower ChatGPT extension on Chrome and Firefox.