Exponential View 🔮

Solar does not care about the Strait of Hormuz. Oil flows through chokepoints. Solar follows a learning curve. Great piece @azeem exponentialview.co/p/solar-superc…

We are living through the solar supercycle, the self-reinforcing loop in which every cost reduction of solar opens a new market and every new market funds the next cost reduction which opens the next market. exponentialview.co/p/solar-superc…

"Claude’s mobile app saw over 500,000 downloads on Saturday – its biggest day on record." @azeem @ExponentialView

"In Feb 2025, ChatGPT held 90% of US business AI subscriptions. A year later, Claude commands nearly 70%." @azeem @ExponentialView

Non-techie, casual: 1. @emollick's twitter account and his blog oneusefulthing.org 2. @random_walker's twitter account and the blog normaltech.ai 3. @cblatts 4. @krishnanrohit 5. @exponentialview's newsletter: exponentialview.co 🧵

Congratulations to @azeem and the team at @ExponentialView on a new milestone: 150k subs on Substack. 1400+ posts across 10 years. Remarkable catalogue of the rise of AI in the public sphere. Here's to 150 more!

"This looks like high volatility in vendor choice but low volatility in aggregate demand. A competitive market with stable spending but a stalling total addressable market for direct subscriptions." @azeem @ExponentialView

Not an AI researcher, but does feel like that the current RL training paradigm inherently implies a lot of wasted compute. Also, why the pay packages for top researchers are so high as they can save millions (maybe billions) of compute with more intelligent design & workflows as Dylan Patel has also highlighted. "These systems extract gains from post-training reinforcement learning (refining answers through trial-and-error) and extended inference-time reasoning. Compute is paid per query, not once during pre-training. Ord estimates that this burns 1,000 to 1,000,000 times more compute per insight than traditional training. Returns shrink faster to get to the next milestone." @ExponentialView @azeem

We have a chance to rebuild a better internet – one that’s faster, safer, and fairer. I discuss the future of the web and AI with @Cloudflare's co-founder & CEO Matthew Prince (@eastdakota), who's uniquely positioned to see what’s changing.

There are many thinkers whose writing has inspired me over the years. They fall into three clusters: - Essays that sharpen our intuition around exponential trends, - Essays that help us think in systems, - Essays that train us to recognize patterns across history.

Reversible computing. Interesting space to follow. @azeem wrote about this over a year ago exponentialview.co/p/reversible-c…

🧵🫧 1/ Is AI a boom or a bubble? I built a five-gauge dashboard to separate vibes from fundamentals. The full essay is free to read today.

For OpenAI to beat the Nasdaq... It would need a $1.5 trillion valuation by 2030. I think this is possible and I break down why here: youtube.com/watch?v=7XQvVt…

I discuss the recent research by @erikbryn, @econ_b and @Ruyuchen here: youtube.com/watch?v=mhT0WC…

AI and the future of media – with @TheAtlantic CEO, Nicholas Thompson youtube.com/watch?v=pKYQe5…

This conversation with @OpenAI CPO Kevin Weil (@kevinweil) was shot before GPT-5's release but somehow it's even more relevant now.

How can GPT-5 can be the most capable AI model yet still feel “modest”? Over the past week, I've been tracking the tensions in how people are reacting. They crystallise into five paradoxes: 1️⃣ The moving-goalposts paradox: As AI clears old benchmarks (like the Turing test), we decide they were never good measures of intelligence. The smarter the system, the less its wins feel like proof. 2️⃣ The reliability paradox: Fewer mistakes make the rare ones harder to predict AND more jarring. Accuracy rises, but trust doesn’t rise in lockstep. 3️⃣ The benevolent-control paradox: The more an AI anticipates our needs, the more it shapes our choices by default. Empowerment blurs into subtle steering. 4️⃣ The floor-ceiling paradox: Benchmarks show the biggest gains at the bleeding edge, but users notice them most in everyday tasks. High ceilings don’t matter if the floor still limits autonomy. 5️⃣ The negative-space paradox: Progress sharpens the outline of what’s missing. The closer we get to AGI, the more the gaps dominate perception. Progress can be real AND still feel strangely underwhelming.

Well said @azeem "Some of GPT‑5’s biggest gains are invisible. When it anticipates needs and makes decisions for you, you do not feel the friction it removes; you just experience the smoother path. But that invisibility makes it risky: the more the model guides you, the more your own curiosity and agency can atrophy... The trade‑off is subtle but important. Technologies that take over part of the thinking process can degrade the skills they replace"

@LuizaJarovsky The truth is, it's complicated. We wrote about how GPT-5 can be the most capable and the most underwhelming model yet exponentialview.co/p/the-paradox-…

Sam Altman: With GPT-5, you'll have a PhD-level expert in any area you need Me: Draw a map of North America, highlighting countries, states, and capitals GPT 5: *Sam Altman forgot to mention that the PhD-level expert used ChatGPT to cheat on all their geography classes...

@emollick Related: we unpack how GPT-5 can be the most capable AND most underwhelming model yet exponentialview.co/p/the-paradox-…

The issue with GPT-5 in a nutshell is that unless you pay for model switching & know to use GPT-5 Thinking or Pro, when you ask “GPT-5” you sometimes get the best available AI & sometimes get one of the worst AIs available and it might even switch within a single conversation.