AI news4
Broadcom seeks over $60 billion in debt to bankroll AI chips for Anthropic and OpenAI
Broadcom is in talks with lenders, including Apollo Global Management and Blackstone, to raise more than $60 billion in debt, potentially structured as a $60-70 billion senior secured tranche plus a $30 billion junior tranche, bringing the total package to as much as $100 billion. The financing builds on a partnership Broadcom, Apollo and Blackstone struck in June to fund a $35 billion expansion of Anthropic's computing capacity using Broadcom's custom chips, part of a plan to enable more than 20 gigawatts of AI compute by 2028.
bloomberg.com ↗Why it matters
chipmakers are now raising debt at the scale of sovereign wealth funds just to keep pace with frontier-lab compute demand, a financing pattern that puts real leverage risk behind the AI buildout rather than equity alone.
A new "blind" benchmark finds frontier AI models recover the core idea of a scientific paper from its bibliography alone just 3-15% of the time
Researchers posted "Reconstruction" to arXiv on 17 August, testing seven frontier models against 643 papers across six scientific domains: given only the reference list a paper cited before publication, models had to guess the paper's actual idea. A multi-agent pipeline that ran competing hypotheses through a Swiss-tournament-style bracket did better, reaching 23-42%, roughly 2.4 times the single-model baseline.
arxiv.org ↗Why it matters
a bibliography gestures at a problem space, but models still can't reliably make the creative leap from existing literature to a genuinely new hypothesis, a concrete data point against claims that frontier models are already doing novel scientific reasoning.
Google launches a "Preferred Sources" button as publisher search traffic keeps falling to AI-generated answers
Google introduced an embeddable button letting readers designate a publication as a preferred source across Search, Discover and Google News, saying people are twice as likely to click through to a preferred source when one is available and that over 345,000 unique sources have been selected since a May rollout of the underlying feature. The move follows data showing publisher referral traffic from Google fell roughly 34% over the past year, with small publishers losing about 60% of referrals as AI Overviews increasingly answer queries without a click.
techcrunch.com ↗Why it matters
this is Google's most concrete concession yet that its own AI answers are hollowing out the traffic model publishers depend on, and a button people have to actively click is a modest fix for a structural problem.
Moderna and Merck's AI-designed personalized cancer vaccine clears its first Phase 3 trial
The companies announced 19 August that their individualized mRNA melanoma vaccine, intismeran autogene, combined with Merck's Keytruda, met its main goal in the 1,137-patient INTerpath-001 trial, significantly improving recurrence-free and distant-metastasis-free survival versus Keytruda alone. Moderna's AI reviews sequencing data from a patient's tumor and blood to predict up to 34 patient-specific neoantigens most likely to trigger an immune response, turning per-patient vaccine design into a manufacturing process rather than a bespoke research project; Moderna shares rose more than 120% on the news, the largest single-session gain in the company's history.
statnews.com ↗Why it matters
this is the first individualized mRNA neoantigen therapy to clear a Phase 3 trial, a concrete proof point that AI-driven personalization in medicine can move past small pilot studies into large-scale, statistically significant results.
AI in the nonprofit sector2
Civic-tech nonprofit Code for India launches a 90-day agentic-AI hackathon for public-good projects across India
Code for India launched "Code for a Billion: Bharat Agentic-AI Hackathon 2026" on 15 August (India's 80th Independence Day), a fully virtual, open-source, 90-day program running through 15 November where individuals and teams worldwide build agentic-AI projects addressing education, health, climate, governance and financial inclusion for underserved populations. Projects must be released as open source and deployed, not just prototyped; founder Karl Mehta framed it as "90 days of actual impact, not just development," with winners announced 5 December.
tribuneindia.com ↗Why it matters
a 90-day deploy-or-nothing structure, rather than a weekend demo format, is a direct answer to the usual hackathon complaint that winning projects rarely ship; the program is still accepting participants as it runs.
Crisis Text Line joins Common Sense Media's Youth AI Safety Institute as a founding standards partner
The global mental-health crisis nonprofit announced 17 August it will help develop AI safety standards for products addressing the most severe youth risks, including self-harm, suicidality and emotional manipulation, contributing over a decade of frontline crisis-response data to a multi-month standards process expected to conclude in late 2026 or early 2027.
natlawreview.com ↗Why it matters
a nonprofit that has spent a decade watching how young people actually signal distress is now shaping the technical standards AI companies will be measured against, not just commenting on them after the fact.