Welcome back to The Learning Curve. From OpenAI’s landmark GPT-5 release to DARPA’s leap in autonomous cyber defense, the AI world didn’t slow down this week. Major players rolled out powerful new models, multi-agent systems pushed past theoretical limits, and venture capital continued to pour billions into a handful of high-stakes bets. Alongside the innovation, lawmakers made history with the first state-level ban on AI therapy chatbots, underscoring that policy is racing to keep pace with technology. Here’s your concise rundown of the launches, breakthroughs, deals, and debates shaping the AI landscape right now.
Table of Contents
This Week in AI
1. GPT-5 Unveiled by OpenAI
OpenAI officially launched GPT‑5 on August 7, 2025—a major milestone in AI capabilities. This new model brings advanced reasoning, coding prowess, emotional intelligence, and strong multimodal features to the table. It’s being rolled out across Free, Plus, and Pro tiers for the first time, showcasing improved accuracy, reduced hallucinations, and enhanced voice interaction including tone customization. Model variants like GPT‑5‑mini, nano, and chat are also being introduced, alongside plans for Gmail and Calendar integrations to boost productivity.(therecursive.com, TechRadar)
2. DARPA’s Autonomous Cyber Defense Breakthrough
In the DARPA AI Cyber Challenge (AIxCC), teams developed AI systems capable of detecting and patching software vulnerabilities autonomously. Team Atlanta (Georgia Tech, Samsung Research, KAIST, and POSTECH) led the winning effort, uncovering 77% of injected bugs and patching 61%, with four tools now publicly available. This marks a step forward in cybersecurity, especially for protecting critical infrastructure such as hospitals and water systems.(Axios)
3. Manus Launches “Broad Research” with Multi-Agent AI
Manus AI introduced a new platform—Broad Research—that allows users to deploy over 100 autonomous AI agents at once. Each agent can execute diverse tasks (e.g., product analysis, web scraping, reporting) simultaneously without fixed roles. This represents a notable shift from single-agent models toward large-scale, parallel workflows, challenging conventional AI systems.(The Economic Times)
AI This or That: Cloud LLMs vs Local LLMs
Should you run AI models in the cloud or on your own hardware?
Local LLMs: Thanks to projects like llama.cpp and quantization, you can now run large models like LLaMA 3 or Mistral on consumer GPUs or even laptops. You get low latency, full data privacy, and no ongoing API costs. Ideal for steady workloads and sensitive data, but you’ll need to manage updates, hardware, and scaling yourself.
Cloud LLMs: With providers like OpenAI, Google, or Anthropic, you get instant access to the newest, most powerful models, often months before open-source equivalents. They scale instantly for heavy or spiky workloads and require no infrastructure. The tradeoff? You’re trusting third parties with your data and paying per token, which can add up over time.
Bottom line: For maximum control and privacy, go local. For convenience and cutting-edge features, go cloud. Many teams now take a hybrid approach—keeping core private workflows local and tapping cloud APIs for overflow or niche tasks.
Deals and Dollars
VC Money Floods Into AI—But Is It All on Too Few Horses?
In 2025, 41% of all U.S. venture dollars—$81.3B—has gone to just 10 companies, eight in AI. OpenAI alone pulled in $40B, the biggest VC deal ever, with xAI and Anthropic netting another $20B combined. Investors are chasing the same “power law” giants—one big win to cover the rest—but this herd mentality risks heavy losses if they stumble. Some admit they’re piling in on hype and FOMO, with critics warning venture has shifted from bold bets to consensus chasing.
Naver’s Q2 Profit Soars 50% on AI Push
Naver, Korea’s Google + Amazon, posted $359.5M in Q2 profit—a 50% jump—beating forecasts. Revenue rose 12% across search, commerce, fintech, and content, all boosted by AI. CEO Choi Soo-yeon credited its “AI Briefing” feature for more visits, longer sessions, and higher ad spend. Next up: a 2026 “AI Tab.” Backed by Seoul’s “sovereign AI” push, Naver is building HyperCLOVA X models, expanding data centers, and crafting LLMs for the Middle East and Southeast Asia, while winning major B2B cloud contracts.
SoftBank Surges 10.4% as AI Bets Pay Off
SoftBank posted $2.86B in Q1 profit, reversing last year’s loss. Gains came from Nvidia shares, $7.8B in T-Mobile sales, and a Vision Fund turnaround to $3B profit. Masayoshi Son has gone on offense, committing $10B to OpenAI in March with another $22.5B planned by year-end, helping fuel its $300B valuation. SoftBank is also buying chipmaker Ampere Computing for $6.5B and joining the Stargate AI data center venture in the U.S., despite delays—cementing its front-row seat in the AI race.
Products We Love
Aftershoot – An AI-powered culling and editing assistant for photographers that slashes post-production time, letting you deliver stunning images faster without sacrificing quality.
Floot – A no-code platform for building full-fledged AI apps without hitting technical roadblocks, embracing the new “vibe coding” approach to software creation.
Asteroid – Spin up AI browser agents in seconds to automate back-office workflows, freeing you from repetitive tasks so you can focus on higher-value work.
Terms of AI use
Illinois has enacted the Wellness and Oversight for Psychological Resources Act (WOPR Act), becoming the first U.S. state to ban the use of AI chatbots in providing therapeutic mental health treatment or making clinical decisions. Signed by Governor J.B. Pritzker on August 5, 2025, the law requires all therapeutic services to be delivered by licensed human professionals, while still permitting AI in administrative roles such as scheduling or note-taking. Violations can incur civil penalties of up to $10,000 per incident, with enforcement handled by the Illinois Department of Financial and Professional Regulation.
The legislation was driven by concerns that AI systems lack the empathy, accountability, and oversight necessary for safe mental health care, particularly for vulnerable populations such as children. Lawmakers and mental health advocates stressed that while AI can support clinicians by improving efficiency, it should never replace trained experts in direct patient care. Supporters, including the National Association of Social Workers, emphasized that the bill doesn’t block innovation but instead draws a clear boundary between supportive technology and professional practice.
This move contrasts with states like Utah, which earlier passed a transparency law requiring AI mental health platforms to disclose they are machines, but without prohibiting AI therapy. The WOPR Act sets a precedent for state-level regulation focusing on how AI is used in healthcare, not just on patient disclosure. It signals a growing policy trend where AI deployment in sensitive fields like mental health will face stricter safeguards, ensuring that innovation operates within well-defined ethical and professional boundaries.
Debug AI: What is vibe coding?
Instead of meticulously typing every line of code, a growing number of developers are embracing what’s being called “vibe coding” — a collaborative, AI-driven approach where you describe what you want in plain language and iterate on what the system produces. This shift reframes programming from a rigid, syntax-focused task into an ongoing creative dialogue, closer to art-directing a designer than following a technical manual. The appeal is clear: it lowers the barrier for non-traditional coders, speeds up prototyping, and fosters experimentation. Much like telling a chef the flavor profile you’re craving and refining the dish through taste-testing, vibe coding turns software development into an intuitive, back-and-forth process.
AI Art: Tacky Ozzy Tribute
This week, Rod Stewart stirred controversy after sharing an AI-generated tribute image of rock icon Ozzy Osbourne, intended to honor the Black Sabbath frontman’s legacy. The artwork, which reimagined Osbourne in a hyper-realistic yet stylized form, drew mixed reactions online — some praised the creativity and lifelike detail, while others criticized it as disrespectful for using AI rather than commissioning a human artist. The debate tapped into a larger cultural tension over AI’s role in art, especially when depicting living legends, blurring the line between homage and appropriation in the digital age.
Prompt of the Week
Turn everything into glass:
Our team member came across an amazing prompt while working on the business plan for her startup — it creates beautiful glass-like visuals for PowerPoint presentations. The effect is sleek, modern, and instantly elevates any slide design.
Here’s how it works: you feed the prompt into your favorite AI image tool, specify your theme or color palette, and it generates high-quality, translucent “glass” elements you can drop straight into your deck. It’s quick, customizable, and makes your slides stand out without needing advanced design skills.
We loved experimenting with it, and we hope you enjoy it too — it’s a simple way to bring a polished, futuristic look to your next presentation.
Create a 3D render of the provided object in a modern, minimalistic style, but transform its material into translucent frosted glass with a soft blue gradient. Keep edges smooth and slightly beveled, with light refraction and soft internal glow. Use realistic glass physics with subtle caustics, ensuring internal shadows and highlights are visible. Background should be a clean, light gradient for contrast. The object should maintain its original shape and proportions, but appear as if it’s made entirely from high-quality frosted glass or acrylic, with semi-transparent layers and glowing edges. Before generating, ask the user if they want to keep the background or not.See you in the next issue.