In partnership with
Dear Readers,
What if the smallest model could outthink the giants? This week, a 7-million-parameter model from Samsung’s SAIL lab quietly outperformed major LLMs on reasoning tasks, challenging the industry’s obsession with scale. Meanwhile, physicists who once bridged the quantum and classical worlds have just been awarded the Nobel Prize, a reminder that today’s breakthroughs are built on the long echoes of past experiments. And in Washington, a Senate report warns that AI might erase up to 100 million jobs, forcing us to ask: are we ready for intelligence that reshapes the very idea of work?
In today’s issue, we dive into Tiny Recursive Models, the quantum pioneers who made modern computing possible, and a sobering look at automation’s social cost. Plus: GPT-5’s leap into real science, voice clones that out-charm their human originals, and a glimpse of Gemini 2.5 — an AI that doesn’t just think, but acts. Scroll down, and you’ll see the frontier moving in real time.
In Today’s Issue:
🌐 Gemini 2.5 Computer Use lets AI click, scroll, and act on screens
🧠 A tiny 7M parameter model from Samsung SAIL is outperforming larger LLMs
⚛️ The 2025 Nobel Prize in Physics went to pioneers of quantum chips, foundational for quantum computing.
📉 A new senate report is proposing “robot taxes”
✨ And more AI goodness…
All the best,
GPT-5 Begins Doing Real Science
OpenAI’s Kevin Weil reports that GPT-5 is now performing genuine novel research across disciplines like math, physics, and biology. When guided by experts, it can generate original lemmas and solve problems that would normally take human researchers days. Weil calls this the beginning of accelerated science—a shift where scientific exploration becomes parallelized, faster, and more collaborative, marking a new era in AI-driven discovery.
Voice clones sound realistic but not (yet) hyperrealistic
A new paper released on PubMed investigates how humans perceive AI-generated voices and whether they can be judged “more real” than human voices (a “hyperrealism” effect). The authors find that voice clones (AI-generated voices based on a specific human speaker) are often mistaken for real human voices, but there is no consistent hyperrealism effect at the group level. Interestingly, AI voices are perceived as more dominant, and sometimes even more trustworthy, than human voice.
Mathematician Sees AI (and hallucinations) as Copilot
Number theorist Ken Ono compares human creativity to AI “hallucination,” arguing that both rely on leaps beyond known knowledge - except mathematicians learn to tell fruitful ideas from dead ends. He believes AI will soon act as a true copilot in mathematical discovery, accelerating breakthroughs much like it did in chess and Go. Still, Ono warns that AI lacks the ability to ask deep, generative questions, making human curiosity and judgment irreplaceable even as universities and research evolve around the technology.
Yann LeCun - Self-Supervised Learning, JEPA, World Models, and the future of AI
The Takeaway
👉 Gemini 2.5 Computer Use lets AI read screens + execute UI actions (click, input, scroll).
👉 Best suited for automating web-based workflows where APIs don’t exist.
👉 Early benchmarks show it leads UI control tasks with competitive accuracy & latency.
👉 Use cautiously: supervision recommended, since mistakes in UI tasks can cascade.
Ever thought of an AI that doesn’t just talk, but uses the web like you do - clicking, typing, navigating? Google’s new Gemini 2.5 Computer Use makes that real.
This model is a specialized offshoot of Gemini 2.5 Pro, designed to interact with user interfaces (UIs) - filling forms, manipulating dropdowns, logging in, scrolling, and so on -all by reading screenshots and outputting UI commands. Rather than only working via APIs or structured endpoints, it “sees” the screen and acts.
Many real-world tasks require UI-level interaction (think admin dashboards, internal tools, websites without open APIs). Gemini 2.5 Computer Use bridges that gap. For AI developers and researchers, it’s a powerful agentic tool for automation and orchestration. Also, early benchmarks suggest it outperforms competitors in web & mobile UI control under latency constraints.
Looking ahead: As this class of models improves, we might see AI assistants that truly operate software on our behalf - no coding, no manual bridging.
Why it matters: Because it turns UI-heavy tasks - today a pain point - into things an AI agent can handle autonomously. And it extends the frontier of what “agent” in AI really means.
Sources:
Choose the Right AI Tools
With thousands of AI tools available, how do you know which ones are worth your money? Subscribe to Mindstream and get our expert guide comparing 40+ popular AI tools. Discover which free options rival paid versions and when upgrading is essential. Stop overspending on tools you don't need and find the perfect AI stack for your workflow.
Power demand from AI data centers to quadruple in 10 years, per Bloomberg. China will exceed every other nation in the long run.
Tiny Model Beats Major LLMs
Samsung SAIL’s new Tiny Recursion Model (TRM) demonstrates that advanced reasoning doesn’t always require massive models. With only 7 million parameters, TRM achieves 45% on ARC-AGI-1 and 8% on ARC-AGI-2, outperforming many large language models on complex reasoning benchmarks. The work highlights how recursive computation can rival scale, opening paths toward lightweight, efficient AGI-like reasoning.
Quantum Chips Win Nobel Prize
The 2025 Nobel Prize in Physics went to John Clarke, Michel Devoret, and John Martinis for experiments in the 1980s that showed quantum mechanical behavior (like tunneling and energy quantization) in macroscopic superconducting circuits—bridging quantum theory and real-world devices.
Their work laid the foundations for superconducting qubits used in today’s quantum computers, proving that quantum phenomena can manifest beyond the atomic scale.
Senate Warns: AI May Cause Massive Job Loss
A new Senate report (led by Bernie Sanders) warns that artificial intelligence and automation could wipe out nearly 100 million U.S. jobs by 2035, based on a ChatGPT-driven analysis of 774 occupations.
It ranks high-risk fields (fast food, customer service, accounting, logistics) and proposes countermeasures like a “robot tax,” reduced workweeks, and worker ownership to offset disruption.
Sponsored By Vireel.com
You need customers.
We make them come to you — with viral videos built by AI.
100+ scroll-stopping ads in seconds.
$3 each.
