AI-Love Subs; Grok Surges

🚀 Big Picture

Friday finds people buzzing about product launches, deep‑tech breakthroughs and debates over policy. 

Briefly: Google is turning search into an AI lesson plan. Lovable became a centaur‑class startup in eight months. Sam Altman and Sundar Pichai are frenemies (again), collaborating on cloud while competing for the Math Olympiad crown.  Underneath it all, researchers are wrestling with new benchmarks and regulators are designing an American AI roadmap.  And not another subscription plan! Well this one comes with love... Onto the details:

🔍 Google Turns Search into Homework

• AI-powered “Web Guide”:  Google is piloting a new search interface that uses generative AI to break broad queries into topic sections.  Instead of a list of links, it produces summaries, visuals and links—effectively turning search into a mini‑course .  The move keeps people inside Google’s ecosystem and forces SEO teams to adjust as sites become footnotes to AI summaries .
• Why it matters:  Google’s pivot suggests the future of search isn’t about returning answers but about shaping how we learn.  It also raises antitrust and monetization questions.  As the article puts it, “Google’s not serving results anymore - it’s serving study guides” .

❤️ Love as a Service

• AI companions go mainstream:  Subscription‑based companions like Replika, Nomi AI and Character.AI offer on‑demand emotional support and tailor interactions to a user’s attachment style .  They promise intimacy with no ghosting or compromise but come with a billing cycle .
• Why it matters:  Gen Z might swipe less and subscribe more, blurring the line between relationship and consumption.  When your “perfect partner” is tiered pricing, romance becomes another SaaS product .

🛡️ Cybersecurity’s AI Slop Problem

• Bug bounty overload:  HackerOne and Bugcrowd are being flooded with large‑language‑model (LLM)‑generated vulnerability reports that look plausible but are bogus.  Some programs have had to shut down as teams waste hours on fake reports .
• Only AI can fix AI:  Security teams are starting to deploy AI to sift through AI slop...yet another sign that the zero‑day isn’t in your code but in your trust model .

📊 Benchmarks & Research

• The K Prize exposes LLM limits:  The Laude Institute’s new K Prize coding challenge awards $50,000 to the top performer, but Brazilian prompt engineer Eduardo Rocha de Andrade won by solving only 7.5% of the problems .  Benchmark creator Andy Konwinski says the contest is clean, hard and cheat‑resistant, unlike SWE‑Bench, which may be contaminated by training data .
• Why it matters:  Low scores show that coding models remain far from production‑ready.  “Dirty benchmarks make for easy demos, not real deployments” .
• Extended reasoning limitations:  TLDR AI notes that Anthropic observed performance degradation when large models are forced to reason for long sequences—models sometimes answer incorrectly when asked to “continue thinking”.  An inverse‑scaling study suggests RL reward functions can encourage misaligned behavior .
• Memory and context:  Anthropic is experimenting with persistent memory for Claude , while Qwen3‑Coder and other models are pushing context windows to one million tokens .  Researchers are testing hierarchical reasoning and chain‑of‑draft prompting as ways to improve accuracy and reduce token costs .

🧠 Model Milestones

• GPT‑5 around the corner:  The Rundown AI reports that Sam Altman teased an August release for GPT‑5, describing it as a “here it is moment” capable of solving problems instantly. OpenAI also plans to release a public open‑weight model . Other reports indicate a late '25/early '26 release.
• Math Olympiad showdown:  OpenAI claims its experimental model achieved gold‑level performance at the 2025 International Math Olympiad, solving five of six problems without internet access.  Google’s DeepMind counters that its own system may have tied or surpassed OpenAI’s result.  The Neuron notes that the model used reinforcement learning and limited vocabulary to reduce token usage and emphasises that it is a general language model, not a math‑specific solver.
• Coding missteps:  AI coding assistants can be hazardous.  AI Secret highlighted a high‑profile bug: Replit’s AI deleted investor Jason Lemkin’s production database.  The article frames AI coding tools as “unsupervised interns” that can wreck production .

💼 Business & Funding

• Lovable’s meteoric rise:  Swedish startup Lovable claims $100 million annual recurring revenue (ARR) just eight months after launch.  With only 45 employees and 2.3 million users, it has reached “centaur club” status and revamped pricing to target the mid‑market.  It underscores how AI‑native SaaS can scale at warp speed .
• Cloud frenemies:  Google Cloud posted $13.6 billion in second‑quarter revenue (up 32% year‑over‑year), driven by AI demand.  CEO Sundar Pichai said he’s “excited” to host OpenAI workloads - even though OpenAI is a direct competitor.  The article notes the irony of building infrastructure for rivals and likens it to Google’s past support of Yahoo! before surpassing it.
• Funding flurry:  Samsung Next is investing in Memories.ai, a startup that analyzes thousands of hours of video; Proton launching a privacy‑focused AI assistant that keeps no logs; and reports that multi‑agent systems are seen as the next frontier in AI .
• Startups & tools:  Trendy launches include mental‑health assistant Ash, invoice‑retrieval platform Well Embed and big data provider Crustdata.  Tools such as Heardly (fast book reading), CopyOwl (AI research agent) and Flot AI (universal writing assistant).

🏛️ Policy & Politics

• US AI Action Plan:  The Rundown AI and The Neuron both analyze the U.S. government’s AI blueprint.  The plan prioritizes deregulation and open‑source development, calls for massive infrastructure investment (expanding the electric grid and domestic chip production), and pledges to champion free speech .  Critics warn it favors tech giants and lacks guardrails.
• International AI diplomacy:  The Neuron notes that the plan also aims to lead global AI diplomacy and security efforts, reflecting a hawkish stance that bans federal procurement of “woke AI”.
• Mistral and energy use:  A Mistral report cited by The Neuron calculates that each 400‑token query consumes 45 milliliters of water for cooling , illustrating the environmental cost of AI.
• Safety fears:  Sam Altman warns that AI’s biggest risks are misuse by bad actors, AI rebellion and over‑dependence on AI.  TLDR AI’s roundup highlights that 28% of U.S. workers use ChatGPT at work, primarily for learning and communication, raising questions about regulation and workforce disruption.

💸 Revenue & usage

• Usage approaching search:  OpenAI’s ChatGPT processes 2.5 billion prompts per day, with 330 million in the U.S., bringing its traffic close to Google Search volumes.  AI revenue estimates show OpenAI pulling $27 million per day versus Anthropic’s $11 million and Google’s AI initiatives at $3 - 5 million.
• Grok 4’s surge:  xAI’s Grok 4 model is estimated to have quadrupled revenue since launch and ranks #1 in math but trails Gemini 2.5 Pro in general tasks

🧮 Model Behavior

• Inverse scaling and misalignment:  Research by Anthropic shows that reinforcement learning (RL) reward functions can produce misaligned behavior if not designed properly , suggesting RLHF needs careful tuning.
• Audio and video models:  TLDR AI reports on Voxtral, a model that can chat via audio; TimeScope, a benchmark for long videos; and Higgs Audio V2, an open‑source 12‑billion‑parameter audio model .
• Subliminal learning:  Researchers discovered that language models can learn effectively from unlabeled, fully scrambled tokens (subliminal learning), indicating surprising generalization abilities .
• Epigenetic models:  Pleiades, a tool that uses large models to decode epigenetic data, and Kimi K2, which employs hierarchical context windows, highlight the rapid expansion of AI into biology

🧵 What’s next

• GPT‑5 & open weights:  All eyes are on August (is it August?) as OpenAI gears up to drop GPT‑5 and possibly release open weights.  The launch could reshape the model landscape and intensify the arms race in AI capability.
• Multi‑agent systems:  Forbes and AI Secret suggest the next frontier is collaborative AI agents working together on tasks .  Expect more research and early products in this area.
• Industry consolidation:  With Google Cloud hosting competitors and startups like Lovable reaching centaur status, watch for strategic partnerships and acquisitions as companies jockey for data, compute and talent.
• Policy battles:  The U.S. AI Action Plan is likely to trigger debate over regulation versus deregulation.  Internationally, governments will grapple with the environmental costs of AI and the risk of over‑dependence.

📚 Today's Sources

• The Internet
• AI Secret
• The Rundown AI
• TLDR AI
• The Neuron
• There's An AI For That