OpenAI Partners With APA on ChatGPT Teen Safety
ALSO: Astra’s math proofs, Google AI 91%, Valar’s $1B bet
Welcome to Guru99 AI Report!
Top Story: Welcome back! From 1 in 8 teens leaning on AI for emotional support to disputed math proofs and Google’s 91% accuracy problem, today’s stories reveal where AI is quietly reshaping trust. Ready to explore what’s really happening?
🧠 OpenAI Teams With Psychologists on Teen AI Safety
Brief Buzz:
OpenAI is partnering with the American Psychological Association to make its AI safer for teens. The goal is practical, evidence-based guidance for young users. It comes as more teens quietly turn to chatbots for emotional support.
- Together they’ll create resources for parents, clinicians and school psychologists, and teens.
- Priorities include supporting teens in distress and flagging unhealthy AI overreliance.
- It builds on existing OpenAI safeguards like parental controls and an age-prediction model.
- 12% of US teens already use AI for emotional support or advice, per Pew Research.
- The effort follows a wrongful death lawsuit alleging ChatGPT contributed to a 16-year-old’s suicide, which OpenAI denies.
💡 Why Should You Care?
If a teen you know leans on chatbots for emotional support, this shows AI makers are starting to build guardrails – but human connection still matters most.
🧮 OpenAI’s Astra Claims 10 Math Proofs, Draws Misconduct Charges
Brief Buzz:
OpenAI says an unreleased model called Astra produced ten new proofs for decades-old math problems. Each comes with a machine-checkable certificate. Then mathematicians read the full 249-page paper – and the reaction turned sharply critical.
- Astra is OpenAI’s unannounced “next major model”; the ten results came from an internal version.
- Every proof ships with a Lean 4 certificate on GitHub – machine-checkable, so the logic can be verified independently.
- The headline result constructs a non-sofic group, a question open in group theory since 1999.
- OpenAI puts the total compute at roughly $2,000 – but that counts only the runs that succeeded.
- Named mathematicians say two proofs reuse prior work without credit; OpenAI is now revising its paper.
💡 Why Should You Care?
The real step up: proofs you can check by machine, not just trust. But “AI solved it” oversells – the model stays private, and humans built the borrowed ideas.
🔍 Google’s AI Overviews Are 91% Accurate – At Massive Scale
Brief Buzz:
A New York Times–commissioned analysis by AI startup Oumi found Google’s AI Overviews are factually accurate about 91% of the time. Impressive – until you spread that error rate across the roughly five trillion searches Google handles each year.
- Oumi ran 4,326 queries through the SimpleQA benchmark: 85% accurate on Gemini 2, 91% on Gemini 3.
- At Google’s scale, the Times calculates tens of millions of wrong answers hourly – if overviews appear on ~half of searches.
- Over half of correct answers were “ungrounded” – citing sources that didn’t support them – worsening from 37% to 56%.
- Google disputes the method, arguing SimpleQA contains flawed answer keys and that grading one AI with another adds errors.
💡 Why Should You Care?
Ninety-percent accuracy feels safe, but across billions of searches the misses add up. Double-check AI answers on anything that matters – especially health.
⚡ AI Money Is Moving Into the Physical World
Brief Buzz:
AI’s next investment boom may be moving below the model layer. Fresh mega-rounds for nuclear power and critical-minerals startups suggest investors are betting that AI’s biggest bottlenecks won’t just be chips. They’ll also be electricity, cooling, and raw materials.
- Valar Atomics raised $1 billion led by Sequoia. Its Nvidia collaboration targets a 30-megawatt, water-conserving AI factory.
- Mariana Minerals raised $310 million led by Khosla Ventures, with continued backing from a16z and Breakthrough Energy Ventures.
- Mariana says MarianaOS can roughly halve 5-to-10-year project execution timelines. That is a company claim, not an independent benchmark.
- Valar’s $6 billion valuation was reported by Bloomberg and Axios; the company did not disclose one itself.
- The IEA projects global data-center electricity use rising from 485 TWh in 2025 to about 950 TWh by 2030.
💡 Why Should You Care?
AI is becoming a physical-infrastructure story. That creates opportunities beyond model makers, but also brings nuclear, mining, permitting, supply-chain, and execution risks.
🤖 AI Won’t Replace You, But It’s Raising the Bar
Brief Buzz:
Job hunting in the AI era is brutal – LinkedIn data shows hiring is still well below pre-pandemic levels on both sides of the Atlantic. But at VivaTech in Paris, business leaders argued that workers who pair AI fluency with hard-to-automate human skills still hold a real edge.
- LinkedIn’s Economic Graph shows hiring stuck well below pre-pandemic levels – roughly 25% lower in Europe and 24% in the U.S.
- An estimated 70% of the skills most jobs require will shift by 2030, with AI as the main catalyst.
- Employers now want a blend of hard, soft, and AI skills – including knowing when to use AI and how to judge its output.
- In software engineering, tools like Codex and Claude Code are pushing hiring beyond raw coding toward collaboration and change management.
- Panelists flagged adaptability and curiosity as essential to getting hired.
💡 Why Should You Care?
The bar for getting hired keeps rising. Specialized expertise alone no longer cuts it – pairing technical chops with communication, emotional intelligence, and a real willingness to keep learning is now the differentiator.
💸 New “Tokenomics” Group Wants to Measure AI’s ROI
Brief Buzz:
How much does AI actually cost – and is it worth it? The Linux Foundation just launched the Tokenomics Foundation to answer exactly that. Thirty companies, including IBM, Oracle, and JPMorganChase, are backing shared standards for measuring AI spend and returns.
- The Linux Foundation launched the Tokenomics Foundation on August 4 to standardize how businesses measure AI costs and returns.
- 30 founding members – including Accenture, IBM, Oracle, SAP, ServiceNow and JPMorganChase – back the vendor-neutral effort.
- Token use could jump 24-fold by 2030, Goldman Sachs projects, widening the gap between AI spend and measurable value.
- The group will build cost frameworks, benchmarks, and a “cost per call” metric via the open FOCUS billing spec.
- It runs alongside the FinOps Foundation, echoing how cloud spending once got its own accounting playbook.
💡 Why Should You Care?
If your company uses AI, you’ll soon get clearer answers on whether it’s paying off. Shared standards could turn confusing token bills into comparable numbers.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
