Oprisa's Journal

August 24, 2025

"Civilization advances by extending the number of important operations which we can perform without thinking of them.” Alfred North Whitehead

In Today's Weekly Journal

AI & Robotics

China has revealed the world’s first pregnancy-capable humanoid robot, equipped with an artificial womb to simulate human gestation, aiming to advance research in reproductive health and robotics.

Details:

  • Kaiwa Technology, led by Dr. Zhang Qifeng, presented a humanoid robot equipped with an artificial womb embedded in its abdomen — intended to replicate the entire gestation process, from conception through birth.

  • The technology builds on past success using “biobag” systems that sustained premature lambs in artificial amniotic fluid — demonstrating feasibility in controlled condition.

  • The team expects a working prototype to be ready within a year and priced at under 100,000 yuan ($13,900).

  • The announcement has sparked substantial media and social reaction. Authorities in Guangdong are already reviewing ethical and legal implications, including issues like maternal bonding, rights of the parent and child, and regulatory oversight.

Why it matters: This humanoid pregnancy robot raises fundamental questions about the future of human reproduction. It could offer new hope for infertility and reduce the physiological burden of pregnancy — but also pushes the boundaries of technology, ethics, and what it means to give birth.

This study evaluates GPT-5’s abilities as a general-purpose multimodal reasoner in medical decision support. Without extensive fine-tuning, GPT-5 is tested on both textual and visual medical questions to gauge its reasoning and diagnostic performance.

Details:

  • GPT-5 achieved state-of-the-art accuracy across multiple medical exam benchmarks—such as MedQA, MMLU-Medical, USMLE self-assessments, and VQA-RAD—outperforming GPT-4o in both text and image-based reasoning tasks.

  • On the MedXpertQA multimodal benchmark, GPT-5 boosted reasoning and understanding scores by approximately 29.3% and 26.2% respectively, compared to GPT-4o—demonstrating notable improvements in integrating visual and textual information.

  • GPT-5 didn’t just match human proficiency—it significantly exceeded the performance of pre-licensed medical professionals. The model achieved +24% in reasoning and over +29% in understanding compared to expert-level benchmarks in multimodal contexts.

  • In detailed case studies, GPT-5 effectively synthesized clinical data and imaging to provide accurate diagnoses and intervention plans, such as recommending a Gastrografin swallow for suspected Boerhaave syndrome.

Why it matters: The leap from GPT-4o’s near-human capabilities to GPT-5 surpassing medical professionals suggests we’re nearing a point where failing to use AI in clinical practice could be seen as malpractice (credit to Dr. Derya Unutmaz). And with intelligence advancing rapidly, that gap is only set to widen.

mage source: Kaitlyn Johnson/Texas A&M Engineering

Texas A&M engineers are reviving RoboBall, a spherical robot first imagined by NASA. Built to roll, bounce, and self-right, it’s being developed for uses from coastal missions to space exploration.

Details:

  • Originally designed in 2003 at NASA by Dr. Robert Ambrose, RoboBall is a self-righting, spherical robot at Texas A&M’s RAD Lab—capable of traversing environments that challenge conventional wheeled or legged robots.

  • RoboBall II (2-foot diameter): Focuses on testing power outputs and control systems. RoboBall III (6-foot diameter): Built to carry payloads like sensors or sampling tools for real-world missions.

  • In trials, RoboBall II reached speeds up to 20 mph—half its theoretical max—surpassing expectations. Upcoming tests in Galveston aim to validate its seawater-to-sand transition capabilities.

  • Graduate students Rishi Jangale and Derek Pravecek have taken the project forward with autonomy, inventing solutions on the fly without blueprints. Long-term, RoboBall could be deployed from lunar landers or sent into disaster zones as autonomous terrain-mapping scouts.

Why it matters: This matters because RoboBall’s rolling design enables it to navigate extreme environments where traditional robots fail. It could unlock new possibilities in space exploration, disaster response, and marine research, reaching areas once thought inaccessible.

Antrophic has added a groundbreaking feature to its Claude Opus 4 and 4.1 models; the ability to autonomously end conversation in rare, extreme cases of persistent harm or abuse. Built as part of its exploratory work on “model welfare,” this feature represents a new frontier in AI safety and ethical design.

Details:

  • The feature activates only in extreme scenarios where the user repeatedly pushes harmful or abusive content despite multiple refusals and redirection attempts.

  • Anthropic frames this capability as a low-cost intervention to protect AI itself, acknowledging uncertainty around the moral status of models while treating their well-being seriously.

  • Claude will not terminate conversations if users are expressing self-harm or harm to others-those cases continue to receive engagement to prioritize human safety.

  • Ending a chat doesn’t block access entirely—users can start a new conversation or branch from a previous one by editing earlier messages. Most users are unlikely to encounter this feature in normal use.

Why it matters: This development marks a significant shift in how AI systems can self- regulate and uphold ethical boundaries—not just for user protection but also to safeguard the AI itself. If signal that “model welfare” is more than a philosophical concept and may become an essential pilar of AI design and deployment.

NASA and IBM have unveiled Surya, a groundbreaking open-source AI model trained on nine years of Solar Dynamics Observatory (SDO) data. Designed to forecast solar activity—such as flares and solar wind—with 16% greater accuracy, Surya marks a significant advancement in space weather prediction.

Details:

  • Surya leverages nearly a decade of continuous observations from SDO, including solar imagery and magnetic field measurements, providing a unique, long-term foundation for AI learning.

  • Surya visualizes solar flares up to two hours in advance with a noticeable improvement over traditional forecasting benchmarks.

  • Built on a foundation model with spatiotemporal transformer design, Surya excels at various heliophysics tasks—solar flare prediction, active region tracking, wind forecasting—without requiring heavy fine-tuning.

  • Both the model and its datasets are publicly available via Hugging Face and GitHub, enabling the broader research community to test, adapt, and innovate.

Why it matters: With society’s reliance on satellites, GPS navigation, power grids, and telecommunications only growing, timely and accurate forecasts of solar disruptions are critical. Surya empowers earlier, more precise alerts, offering robust protection for both space-based systems and terrestrial infrastructure.

🔬Science snapshot

>Researchers discover eight new genes linked to schizophrenia; study suggests the condition may be related to how DNA is organized within cells, and how brain cells communicate using the chemical GABA (More)

>Study reveals neurons help control the inflammatory response in the gut, play a role in inflammatory bowel disease, enteritis, and other conditions (More)

>Surgery-free alternative to LASIK that uses electricity instead of lasers to reshape the cornea proves effective on rabbit eyeballs; initial results suggest the method could be a promising technique for correcting human vision (More)

>Megalibrary of nanoparticles helps researchers quickly find abundant, cheap alternative to iridium, a rare metal critical to clean hydrogen energy production; could accelerate discovery of commercially viable materials for various application (More)

Market view

Investor Jon Charbonneau argues that crypto investing is no different from traditional markets: it comes down to fundamentals (future cash flows) or flows (trading demand). Bitcoin’s flows-driven thesis has worked so well it now resembles a fundamental asset like gold, while Ethereum straddles both approaches, making it more complex. Charbonneau believes crypto is shifting toward fundamentals as the industry matures, creating real value beyond speculative “narrative trading.”

Charbonneau notes that many who claim to invest on fundamentals are really chasing flows. David Einhorn illustrates this, saying markets are “fundamentally broken” because the old game of buying low P/E stocks and selling higher no longer works. With fewer competitors, he now focuses on being paid directly by companies — a shift Charbonneau and Buffett would endorse.

President Trump announced that the U.S. government will acquire a 10% equity stake in Intel, valued at roughly $10 billion. The deal redirects previously allocated Chips Act and federal funds into non-voting shares of the chipmaker, aiming to support domestic semiconductor production and national security.

From offices to servers : U.S. investment in data centers is about to exceed that in office buildings.

Ben Carlson points out that OpenAI usage fell sharply once the school year ended — raising the question of wether investors are betting on more than just students using AI to finish their homework.

Brad Setser argues that the US is effectively exporting government debt and AI investments, with the AI boom helping finance the current account deficit. Investment flows into equities dominate the picture, while government debt makes up much of the rest.

The World’s Most Valuable Unicorns in 2025

After a funding slowdown post-2021, private investment is surging again—especially into AI leaders like OpenAI. Unicorns are also emerging globally, with countries like India and the UK joining the ranks.

A Visual Capitalist graphic highlights 2025’s top unicorns, based on Crunchbase data. SpaceX leads with $15.5B projected revenue, powered by its Starlink business. OpenAI ranks second at a $300B valuation, with $12B in annualized revenue from 700M weekly users. ByteDance follows with a $220B valuation and $186B in expected 2025 revenue, nearly matching Meta.

Friday 22 ETHEREUM HITS NEW ALL-TIME HIGH, SURPASSING 2021 PEAK

Record pain for short sellers: more than $350M liquidated in a single day

Bitcoin Demand Slows as Profit-Taking Continues

Bitcoin demand has dropped from 174K BTC in July to 59K BTC, with ETF inflows and Strategy’s accumulation both hitting multi-month lows. Market sentiment cooled from “Extra Bullish” to “Bullish Cooldown” as BTC topped $124K.

Holders have realized $74B in profits since July 4, yet downside risk looks limited, with $110K expected to act as strong support.

If you found this valuable, consider sharing or forwarding it to a friend who needs to be ahead of the curve. 🧠📩

📬 Until next time,

— The Macro Scratch team, with AI-powered research and support.