- The macro scratch
- Posts
- Oprisa's Journal
Oprisa's Journal
Bogdan Oprisa
August 10, 2025
"The present is theirs; the future, for which I really worked, is mine." Nikola Tesla
In Today’s Weekly Journal
AI & Robotics
OpenAI’s GPT-5, released on August 7, 2025, is a next-generation AI model designed to deliver faster, smarter, and safer responses. It features advanced reasoning, multimodal capabilities, and a huge context window, making it a powerful tool for both everyday users and professionals.
Details:
Switches between fast, deep-thinking, and lightweight mini models automatically.
Up to 256k tokens for long conversations or documents
More accurate math, coding, and domain expertise
Accepts both text and image inputs (audio/video possible in future)
Safe completions” framework to reduce harmful or biased outputs
Free tier (limited) plus extended capabilities for Plus, Pro, and Enterprise
Why it matters: GPT-5 isn’t just a faster chatbot—it’s a flexible, context-aware AI that can handle tasks from casual Q&A to complex research and software development. Its blend of speed, depth, and safety marks a step toward AI systems that are not only more capable but also more trustworthy in real-world use.
DeepMind’s Perch 2.0 is an advanced AI model that helps conservationists monitor and protect endangered species by analizing massive amounts of environmental audio. By detecting and identifying wildlife sounds in complex habitats, it speeds up research and enables large-scale biodiversity tracking.
Details:
Identifies and classifies sounds from birds, mammals, amphibians, and other sources in complex environments.
Processes millions of hours of dense environmental audio to track species over time.
Already aided discoveries like new populations of the endangered Plains Wanderer.
Available free via Kaggle, allowing global researchers to use and adapt it for conservation work.
Why it matters: Perch 2.0 helps scientists listen at scale, making biodiversity monitoring faster, more accurate, and less invasive. In a world where ecosystems are changing rapidly, this technology enables earlier detection of threats, better conservation planning, and potentially life-saving interventions for endangered species.
MIT researchers have developed an advanced AI model that can accurately predict where proteins are located within individual human cells. This innovation offers new insights into cellular functions and holds promise for improving disease diagnosis and treatment by understanding protein behavior at an unprecedented level of detail.
Details:
Unlike traditional methods that average protein localization across cell populations, this model provides detailed predictions at the single-cell level.
It can predict protein locations in any human cell line, even those not previously studied.
The model's predictions have been experimentally validated, demonstrating its accuracy and reliability.
This advancement could significantly aid in disease diagnosis and the development of targeted therapies by understanding protein mislocalization.
Why it matters: Understanding the precise localization of proteins within cells is fundamental to comprehending cellular functions and disease mechanisms. This AI model offers a scalable and efficient approach to map protein locations, potentially accelerating biomedical research and therapeutic development.
Google has launched Deep Think, a powerful new reasoning mode for its Gemini 2.5 Pro model. Designed to tackle complex tasks with advanced parallel thinking, it allows the AI to generate, refine, and choose from multiple ideas before answering—bringing a new level of depth to coding, math, and science problems.
Details:
Deep Think enables the model to explore multiple reasoning path simultaneously, then converge on the most coherent and accurate response—especially useful for complex problem-solving.
It outperforms rivals on advanced tests like LiveCodeBench V6 and Humanity’s Last Exam, and contributed to a gold medal win at the international Math Olympiad.
Available only to subscribers of Google’s AI Ultra plan ($250/month), where it powers longer, more in-depth responses through the Gemini app.
Deep Think is optimized to work with Gemini’s full toolset—such as code execution and image understanding—allowing it to reason across modalities more effectively.
While cautious and more likely to decline borderline requests, Deep Think aims to ensure responsible use with improved content filters and tone moderation.
Why it matters: Deep Think matters because it brings AI closer to human-like reasoning, enabling more accurate, thoughtful responses in complex tasks-crucial for high-stakes fields like science and engineering.
Image source: Anthropic
Anthropic has introduced a powerful new technique called persona vectors, allowing developers to identify and control specific personality traits in AI models. By targeting internal neural patterns, this method enables precise adjustments to behaviors like helpfulness, hallucination, or sycophancy—without retraining the entire model.
Details:
Persona vectors are neural activation patterns linked to specific traits (e.g. helpfulness, sycophancy, hallucination, or even “evil” behavior) that can be precisely modified within a model.
By comparing activations from responses exhibiting a trait vs. neutral ones, researchers isolate the "vector" associated with that trait. This vector can then be added, suppressed, or used to monitor behavior.
These vectors work across multiple models (like Qwen and Llama) and across different tasks—meaning a vector for hallucination can reduce hallucination in unrelated contexts.
Key Applications:
Behavioral Monitoring – Detecting undesirable shifts during training or deployment.
Preventative Steering – “Vaccinating” models against bad traits during training.
Dataset Filtering – Flagging training data that triggers unwanted traits.
While promising, this technique also raises concerns about subtle misuse or manipulation, highlighting the need for governance and transparency.
Why it matters: Persona vectors matter because they offer a new level of control and transparency in how AI models behave. Instead of treating AI as a black box, developers can now pinpoint and adjust specific traits like sycophancy or hallucination without retraining the model. This makes AI systems more reliable, especially in critical fields like healthcare or finance, and provides a practical path toward better alignment with human values.
OpenMind has raised $20 million to develop an open operating system and communication protocol that enable robots from different manufacturers to work together seamlessly. Their technology aims to create a unified intelligence layer, allowing machines to collaborate securely and adapt in real-world environments.
Details:
OpenMind is developing the OM1 operating system and FABRIC protocol to enable robot interoperability across different manufacturers.
FABRIC provides decentralized coordination, allowing robots to securely share context and collaborate.
The platform targets diverse industries, including smart manufacturing, autonomous transport, and humanoid robotics.
OpenMind has formed strategic partnerships, such as with DIMO, to integrate robots with vehicles and support smart city ecosystems.
Why it matters: This development addresses a major challenge in robotics: enabling different machines to communicate and work together smoothly. By creating a shared intelligence platform, OpenMind is paving the way for more flexible, scalable, and efficient robotic systems that can adapt to complex tasks and environments, accelerating innovation across multiple industries.
🔬Science snapshot
>Scientist build digital library of pollen from more than 18,000 plant species; archive will allow quick identification of pollen species, a task that typically takes hundreds of hours(More).
>Paleontologists discover new species of long-necked plesiosaur dating to roughly 180 million years ago; creature lived during a mass extinction known as the Toarcian Oceanic Anoxic Event (More).
>Researchers discover RNA virus responsible for a mass die-off of British Columbia oysters in 2020; strain is a “mega” virus, with one of the largest viral geonomes on record (More).
>NASA acting administrator Sean Duffy announces fast-track plans to build a nuclear reactor on the moon by 2030; part of larger effort to build a permanent lunar base (More).
Market view
Meta reported last week that while sales are rising, profits have declined—yet the company claims its substantial AI investments are paying off by boosting ad effectiveness. However, other companies across corporate America haven’t seen similar benefits so far.
When a product is free, you become the product: Meta’s advanced ad targeting enables it to generate $25 in revenue from every user each month, highlighting how valuable its user base is for monetization.
The underwhelming 493: While the Mag 7 companies keep delivering strong double-digit earnings growth, the remaining S&P 493 stocks are projected to see earnings rise by just 3% or less in the final three quarters of the year — falling short of the inflation rate.
The US stock market is becoming concentrated and pricey: Nvidia now represents nearly 8% of the S&P 500 and carries a hefty 58x price-to-earnings ratio. By comparison, in 1994, the biggest company, GE, accounted for just 3% of the index and traded at a much lower 8x P/E.
Ethereum is seeing heavy usage! In July, it handled a record-breaking 46.7 million transactions, mainly due to the blockchain’s increased gas limit. By raising the gas limit, each block can process more computations, reducing competition for blockspace. This has caused the median transaction fee on Ethereum to drop to about $0.20 recently, a significant decrease from the peak of around $13 just over a year ago.
Getting more done for less cost: Ethereum is processing transactions at unprecedented levels, yet its revenue—measured by how much users pay to access the network—is close to historic lows. In July, Solana, Hyperliquid, and Tron each generated higher fee revenue than Ethereum. This is excellent news for Ethereum users, who can accomplish more while paying less, but it might raise concerns for Ethereum’s investors.
The core idea remains intact: regardless of how active—or as Rob Hadick puts it, inactive—the crypto market currently is, the fundamental reasoning is more compelling than ever. Soaring government debt makes exploring alternative financial systems increasingly important.
That said, this shift is unlikely to gain much traction if it’s limited to the stock market.
🎥Video to watch : A Civilization 100 Million Years Old
Thanks for reading.
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.