AI’s GeoGuessr genius and the art of prompting well
David Peterson

A few days ago, Kelsey Piper, a writer for Vox, went quasi-viral with a photo of her kid flying a kite on a non-descript beach. Would OpenAI’s o3 be able to figure out what beach this picture was taken on?

Turns out, o3 was able to locate that spot!

Inspired by Kelsey’s result, Scott Alexander put o3 to the test with some even more obscure and intentionally difficult-to-locate images (all similarly stripped of metadata and reoriented to ensure the originals weren’t part of o3’s training data). The results were similarly stunning.

I have two reactions.

First, these sort of emergent skills continue to surprise/amaze/excite me. OpenAI did not train o3 to be good at GeoGuessr in particular. But it is! As it turns out, knowing a lot about everything makes you pretty good at a lot of things.

Second, o3 isn’t that good at locating images “out of the box.” It requires some creative prompting to make it work. Kelsey shared her prompt here. Go look at it. It’s a work of art, in and of itself. And when asked by somebody on that original thread how she came up with it, she had a great answer:

I love this answer. Because it’s a great example of something that I don’t believe is discussed enough: the art of prompting well.

When generative AI first burst onto the scene, "prompt engineer" became a buzzword and then, soon after, a punchline. But talk to anybody who is really steeped in using AI day in and day out, and they’ll be the first to tell you that there is genuine alpha - consistent, significant advantage - available to those who master AI interaction.

The reason, obvious as it may be, is because models are unlike traditional software. They're probabilistic rather than deterministic, responding to natural language instead of precise commands. This creates a dynamic where interaction quality dramatically affects outcomes in ways we've never seen with conventional tools.

Perhaps the most surprising aspect of this emerging landscape is who excels at it. Unlike previous technological revolutions that primarily rewarded technical expertise, exceptional AI interaction often comes from a different set of traits entirely. You don't need to be a programmer or have a technical background to be in the 99th percentile of AI users. What's required is something closer to curiosity, openness, taste, and the ability to interact with living systems in a responsive way. Look at what Kelsey did.

What makes this particularly fascinating is how AI transforms traditionally procedural knowledge work into something more akin to creative pursuits with a natural skill curve like a musical instrument (anybody can plunk out Chopsticks on a piano but just listen to some Chopin or Rachmaninoff for a taste of what expertise sounds like) or gaming (where button-mashing works for beginners but advanced techniques take you the next level) or cooking (where following recipes produces reliable meals but experimentation and improvisation yields amazing results).

These sorts of skill curves don’t normally exist in the arena of knowledge work. But, with AI, perhaps now they do?

You might imagine that the models will improve so much that the arbitrage available to great prompters will diminish. But I don’t think so. The mere fact that the models are probabilistic ensures some level of randomness to the output. As long as that exists, there will be alpha available to those who can prompt well.

Said another way, as base model capabilities improve, the ceiling for what's possible with expert use rises too. This is about comparative advantage, not absolute capability. Even as the floor rises for everyone, those who master AI interaction will continue to find ways to push the ceiling higher.

For the past 10+ years, we've talked about "10x engineers" with reverence. But what if AI is silently creating 10x marketers, 10x designers, 10x analysts…even 10x journalists or 10x academics or 10x CEOs?

Here’s a prediction (which, if you’ve spent time with an expert prompter, will not be hard to believe): AI will transform knowledge work into a creative pursuit with a genuine skill curve, just like piano or gaming or cooking. And, as a result, we may find 10x performers becoming much more common across all domains.

A few final reflections:

First, this has convinced me that AI isn’t just a tool but a creative medium worth mastering. Follow Ethan Mollick or Dan Shipper for a few weeks for some inspiration of what creative, next-level AI usage looks like.

Second, what’s compelling about this vision of the future is that it doesn’t diminish human expertise…it just transforms it. There will still be loads of dislocation and disruption. But I don’t foresee the wholesale replacement of humans across entire industries. Not yet, at least.

Third, this is truly the era of the principal (vs. the agent). Never before has so much power been placed at the fingertips of…well…anyone with an internet connection! If you can wield AI to your advantage, you will be able to do extraordinary things.

Engineering Success in a Technical Startup
For technical founders, extraordinary success is highly correlated with how obsessive a founding team is about developing an extraordinarily efficient GTM machine.

What if an AI Was Your Best Customer?
Gil explores how AI agents autonomously selecting and consuming tools can reshape the software industry.

Seven Signals
An evolving VC investment decision framework.

Five Emerging VC Swimlanes
The venture market is evolving into five distinct swim lanes, each with its own rules, characteristics, and key players.

ENTERPRISE/TECH NEWS

Is AI normal? Despite all the speculation and fear surrounding AI, two Princeton AI researchers, Arvind Narayanan and Sayash Kapoor, wrote a 40-page paper urging everyone to calm down and treat AI like a normal technology. They argue that comparing AI to nuclear weapons or “superintelligent entities” does more harm than good. “We view AI as a tool that we can and should remain in control of, and we argue that this goal does not require drastic policy interventions or technical breakthroughs. We do not think that viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts, nor is it likely to be in our vision of the future.” Instead, they suggest we think of AI more like the internet or electricity: a general-purpose technology whose societal impact will be messy, slow, and tied to how we adopt it.

EU’s €200B AI investment plan. The European Union has unveiled a new €200B “AI Continent Action Plan” to help position itself as a global leader in artificial intelligence. Lagging behind the US and China, Europe has had far fewer AI achievements and has lower enterprise AI adoption. The plan hopes to close the gap and better position the EU as an AI leader. “This ambitious plan will involve investing €200 billion towards increasing the EU’s AI capabilities, aiming to transform Europe's robust talent pool and prowess in traditional industries into engines that will accelerate AI innovation in the EU.” The plan includes five key pillars: 1. Building a large-scale AI data and computing infrastructure, 2. Increasing access to data, 3. Fostering AI adoption in strategic sectors, 4. Strengthening AI skills and talents, and 5. Fostering compliance with the EU AI Act.

2025 exit outlook dims. While many were hopeful 2025 would be a record year for tech, underpinned by hopes of the exit market picking back up, Trump’s economic policies and new tariffs have reset expectations. The economic instability has had profound effects on the startup ecosystem. According to Avi Eyal, partner of Entrée Capital, even if the tariffs are reversed, “the IPO market probably won’t reopen before the end of 2025, and M&A activity is also slowing. In Israel, most startups are built for acquisition, so many will find themselves in distress.” That’s why VCs are getting anxious, Eyal says. “It’s not clear who will buy the startups in their portfolios or how they’ll exit their investments.”

HOW TO STARTUP

The new pricing playbook. Kyle Poyar, the pricing expert, shared his initial findings from a survey of 240 software companies. The results show that the old pricing playbook is dead. Only 15% of those surveyed still use seat-based pricing as their core pricing model, with just 3% seeing it as the future. Hybrid pricing (subscription + usage/success fees) is gaining momentum, with 40% using it as their core pricing model, up from 25% only a year ago.

Startup funding landscape. Head of Insights at Carta, Peter Walker, shares interesting data on the startup funding landscape. According to Carta data, a paradox is emerging: fewer rounds, but higher valuations. Regarding early-stage funding, both Seed and Series A deals are at a high for median valuations, but a low for round volume.

HOW TO VENTURE

Angular AGM excerpt. Gil recently shared this 12-minute excerpt from Angular’s closed-door Annual General Meeting with our Limited Partners back in February. The segment offers our candid, data-driven perspective on the significant shifts currently defining the venture capital market. Watch the excerpt and read the key takeaways here.

The evolution of VC. In the last few years, many of the largest VC funds, like Sequoia, Andreessen Horowitz, and General Catalyst, have shifted away from the traditional VC playbook to invest across a broader range of assets. Last week, news broke that Lightspeed, which manages $31B in assets, is also eschewing the traditional VC model and officially registered as a Registered Investment Advisor, enabling the fund to invest in assets beyond direct startup equity. While some are saying this signals that ‘venture capital is dead’ or that ‘the venture model is broken at scale’, our take is it’s more accurate to say: "these firms have opted out of the venture market by scaling to unprecedented levels that were never demonstrably successful in venture."

Reco secures Microsoft 365 Copilot by spotting risky prompts, protecting data, managing user access, and identifying threats.

FalkorDB was voted the Startup of the Year 2024 for Tel Aviv by HackerNoon.

Aquant
Senior Enterprise Account Executive (New York)

CruxOCM
Senior Software Engineer - Platform (Remote)

Reco
Detection Engineer (Tel Aviv)

Tensorleap
Head of Customer Success (Tel Aviv)