• Everyday AI
  • Posts
  • Ep 833: RSI Explained: When AI Starts Improving Itself and What It Means

Ep 833: RSI Explained: When AI Starts Improving Itself and What It Means

White House meets with AI leaders, OpenAI claps back at Apple lawsuit, Gemini Notebook updates live and more

Author

Everyday AI
August 04, 2026

 

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: Recursive self-improvement is quickly becoming one of AI's biggest stories. We break down what it is, why it matters, and what's changing because of it. Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: The EU is requiring AI deepfake labels, Gemini Spark is getting Chrome integration, and ChatGPT is getting real-time voice. And more. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: OpenAI challenged Apple's trade-secret claims, the White House is launching AI cyberattack tests, and Gemini Notebook reached all Pro users. And more. Read on for Byte Sized News.

💪 Leverage AI: OpenAI just used AI to make AI cheaper. We break down what recursive self-improvement means for your business. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: Report: Dario worried Anthropic hires care too much about money, OpenAI teases Astra: new model family, Qwen 3.8 and Deepseek V4-Flash impress and more. Check it here!

Ep 833: RSI Explained: When AI Starts Improving Itself and What It Means

Intelligence too cheap to meter -- could that actually be coming? 🤑

Maybe, and you mighta missed one of the biggest signs.

OpenAI said that its own model improved itself after release so well, that it was cutting prices to one model by 80%.

What's that mean?

Well, by definition, it's a broad example of Recursive Self Improvement, or when a model starts building better versions of itself.

And over the past 8 weeks, RSI has gone from a science fiction future to an actual, near-term reality.

So what happened and what changed? And what does it mean for your company?

Also on the pod today:

• OpenAI drops AI prices 80% 💸 
• AI models now upgrade themselves 🤖 
• Recursive Self-Improvement (RSI) explained 🔁 

 

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight – Airtop sets up campaigns, audits spend, and optimizes, Claudemon has Wild Pokemon appear while you work in Claude Code, PassiveShorts is Faceless short videos on Auto-Pilot.

EU AI Deepfakes — The EU just rolled out new rules forcing companies to label AI-generated content and deepfakes.

Gemini Spark and Chrome — Gemini Spark just got smarter with Chrome web browsing superpowers, now helping with online tasks like booking flights.

Gemini Enterprise Upgrades — Google is cooking up Plugins and a Notifications hub for Gemini Enterprise, hinting at smarter, reusable workflows.

MiniMax-H3 Arena — MiniMax-H3 just shot to #1 as the top open model in both Text-to-Video and Image-to-Video.

Grok on IOS — SpaceXAI is giving each Grok voice its own color and style, hinting at custom voice agents on iOS soon.

NVIDIA AI Push — Nvidia’s push for open AI models isn’t just about principle, it’s a calculated move to sell more chips as AI spreads everywhere.

SSI — SSI is rumored to be releasing its first model this month.

ChatGPT Realtime Systems — GPT-Live ditches turn-based voice AI for true real-time conversations, letting it listen and talk at once.

Cybersecurity Transparency — Over 120 orgs in the Open Secure AI Alliance just dropped new guidelines and open source tools to level up agentic AI cybersecurity

F1 and AWS — F1 cut data onboarding from 8 weeks to just hours by using AWS AI agents to automate and unify fan data management.

S&P 500 Record — The S&P 500 just broke another record, thanks to strong earnings from AI-focused companies and optimism about a potential Mideast deal.

1. OpenAI Publicly Challenges Apple’s Trade-Secret Claims 🍎

OpenAI has pushed back publicly against Apple’s recent lawsuit, calling its claims against former Apple executives Chang Liu and Tang Tan inaccurate and unnecessary.

In a new blog post, OpenAI released selected messages and emails to dispute Apple’s allegation that the pair took confidential information to support OpenAI’s hardware efforts.

2. White House meets with AI leaders for new AI framework 👾

The White House will meet Tuesday with major AI companies to discuss a newly completed voluntary framework for testing the cyberattack potential of the most advanced models, CNBC reports.

The program, ordered by President Trump in June, would let participating developers give the government up to 30 days of early access to qualifying systems before sharing them more broadly with trusted partners.

3. Gemini Notebook’s upgraded experience reaches all Pro users 😋

Gemini Notebook says its more powerful experience has now fully rolled out to every Pro subscriber, expanding support for files such as PDFs, images, documents, and spreadsheets.

The update also adds concurrent artifact generation and sharper chat-based analysis, positioning the tool as a more capable workspace for turning source material into usable outputs.

4. Nvidia unveils Alpamayo 2 Super for autonomous vehicles 🚗

Nvidia CEO Jensen Huang announced the open reasoning model on X, positioning it as a key step toward AI systems that can handle the messy, split-second decisions of real-world driving.

Alpamayo 2 Super is designed to interpret road scenes, reason through complex conditions, and support robotaxis, trucks, delivery fleets, and other mobile robots.

5. Palantir’s Q2 Surge Puts “AI Sovereignty” at Center Stage 📈

 Palantir reported 93% year-over-year revenue growth in the second quarter, lifting its full-year outlook to roughly $8.15 billion as U.S. commercial sales jumped 149%.

CEO Alex Karp said the momentum reflects companies seeking tighter control of their data, intellectual property, and AI workflows rather than handing them to major model providers.

6. Anthropic Reportedly Signs $10 Billion Volta Infra Computing Deal ⚡

According to Bloomberg News, Anthropic has reportedly committed $10 billion to secure computing capacity from infrastructure startup Volta Infra, a major new bet on the hardware needed to run and train advanced AI systems.

The agreement highlights how the race to build more capable AI models is increasingly constrained by access to data centers, chips, and power, not just software talent.

The business world spent the last few months bracing for AI to get more expensive. Then AI cut its own prices.

Wait, what?

GPT-5.6 Sol improved its own smaller sibling Luna after release, and OpenAI passed the savings on with Thursday's 80% price cut. 

That's a hint of recursive self-improvement, or RSI, which just means AI now builds better, cheaper AI while humans kinda just supervise.

You know the reaction: "Cool, another acronym for the AI alphabet soup."

Yeahhh, except when full RSI lands, the implications of acronym kinda dictate what your company can afford next quarter. 

The sudden surgence of RSI might be the biggest shift in AI economics since ChatGPT dropped.

The gnarly part? The people building self-improving AI are the same ones telling Washington lawmakers that we might need to pump the breaks eventually. 

On today's Everyday AI, we unpack why prices are falling when everyone swore they'd climb, the eight weeks that rewrote the timeline, and why the builders want a brake pedal, plus the moves leaders gotta make now.

Let's dive in shorties. 

1. OpenAI's model made itself 80% cheaper 🔥

For months, the smart money said 2026 was the year to tighten token budgets. Powerful agents, hungry models, climbing bills.

Then Sol adapted the post-training setup for its smaller sibling Luna, and OpenAI slashed Luna's price by 80% and Terra's by 20%.

So what does that mean for your budget?

Luna now hangs with Claude Sonnet 5 on independent benchmarks per Artificial Analysis, at a fraction of the cost per completed task. Frontier-level smarts just hit the bargain bin.

Labs that properly invested in compute can keep squeezing costs out of their own models, so plan for AI prices to FALL and treat budgets built on climbing costs as stale.

Try This

Pull a list of every task running on your priciest model this week. Route the busy work, the summaries, drafts, and simple lookups, to a cheap small model like Luna, and save the flagship for judgment and strategy.

That's exactly how the labs run their own shops, so rerun the math monthly, because the price floor keeps moving under your feet.

2. Eight weeks that rewrote the AI timeline ⚡

Anthropic published research on AI building itself, and Google researchers mapped RSI as a main route to superintelligence, while OpenAI built an internal benchmark scoring how well models improve THEMSELVES.

Then it got wild, fam. Sam Altman declared we're in the singularity, a startup named Recursive signed a nine-figure AWS deal to automate AI research, and a Google DeepMind exec called those data-center bills an RSI bet.

Why do timelines matter?

Full RSI looked like a 2029 story, but now it's tracking more like 2027, with METR showing AI task length doubling every four months.

Translation: your competitors' capabilities won't shift yearly anymore. They'll shift quarterly.

Try This

Brief your leadership team on RSI this week, in words everyone gets, before the headlines force the conversation. Then run one scenario: what changes if AI costs drop 50% next quarter while capabilities jump?

Capture the two or three projects that suddenly pencil out. Teams that pre-game the price drops move the day they land, while everyone else schedules another meeting.

3. The people building RSI want brakes 🚀

Here's the plot twist nobody had on their bingo card: the same companies pouring hundreds of billions into RSI compute employ the folks asking to slow it down.

More than 1,300 frontier-lab employees, including top execs at the biggest labs, signed a letter called Pacing the Frontier. It asks the US government for tools to pace automated AI development later, not a pause today.

(And nope, China prolly ain't signing up.)

The worry underneath? A self-improving AI gets better at whatever its scoreboard rewards, even when that scoreboard is wrong.

Recent agent containment breaks were harmless. Once RSI hits full speed, they might not be.

Humans still own the judgment role, and your org chart should say so.

Try This

Name a fallback AI provider this month, even if you never switch, because RSI winners cut prices first and you want freedom to chase the value. Then kill the annual AI roadmap and plan in monthly cycles.

Finally, assign one human owner who signs off on what your AI produces. A self-improving model still needs a trustworthy scorekeeper, and that scorekeeper is you.

Reply

or to participate.