Exploring new ideas in frontier tech with the authors and leaders shaping them. Episodes every other Wednesday.
New York City
Joined December 2025
- Tweets115
- Following2
- Followers211
- Likes191
Pinned Tweet
Frontier technology is scaling rapidly. Understanding lags far behind.
Dry Run began with a narrow focus on innovation in fintech. Since then, we've been drawn to a broader goal: explaining frontier technology.
By frontier technology, we mean fields where capabilities are outpacing the answers around them, including AI, robotics, biotech, and digital assets.
On Dry Run, we bring on technical experts in their respective fields and ask them to explain the open questions that need to be answered to push the frontier forward.
We break concepts down to first principles, give an unvarnished view of where capabilities and limitations stand today, and look at how these advances change the systems around us: work and labor, software and product development, business and organizations, hardware, and research.
As progress accelerates, these technologies will increasingly affect how people work, create, build, and live. We want to shorten the distance between progress and understanding, because understanding leads to better questions, greater participation, and a clearer view of what is at stake.
Technological change can be empowering, but there is still work to do to make it understandable. We want to make that work a bigger part of Dry Run.
Dry Run retweeted
AI's fortune-tellers making predictions 100 years ahead
@AustinCampbell to @TheDryRunPod @eshita: "We don’t even know what happens in the next 10."
Dry Run retweeted
You can't vibe code DocuSign.
The secret sauce isn't the software, it's the trust and distribution they've built.
That's why DocuSign is set up for a comeback.
They refuse to send customer data to OpenAI and other closed models, and are building their own models to protect the information that the business runs on.
Shiv Shankar (@sshankar) on @TheDryRunPod
What is an open weight model?
Shiv Shanker (@sshankar), founder and CEO of Boundless, joins the Dry Run to explain what an open weight model is, and why he thinks most companies will try closed models first but eventually shift to open ones.
00:40 Understanding open weight models
01:57 What labs actually share when they publish weights
06:33 Cost considerations in AI adoption
08:47 Security and data privacy
13:08 You can't vibe code DocuSign
15:48 The MySQL playbook for open source profits
20:33 Why the top open weight models come from China
24:09 What changes after OpenAI and Anthropic go public
25:32 The case against open weight models
30:04 Don't give an agent your passwords
31:50 Regulatory frameworks for open weight models
36:23 Education and workforce development for AI
38:41 The sign that open weight models are winning
Dry Run retweeted
Open weight models are a genius GTM strategy.
If you're the third, fourth, or fifth lab chasing OpenAI and Anthropic, giving away the weights builds trust.
Open weight models are cheaper, transparent, and easy to run on your own systems. You don't have to trust the lab at all.
Commercialization comes from the services that sit on top: consulting, training, and customizing the models. After all, who understands DeepSeek better than the DeepSeek team?
Shiv Shankar (@sshankar) on @TheDryRunPod
What is an open weight model?
Shiv Shanker (@sshankar), founder and CEO of Boundless, joins the Dry Run to explain what an open weight model is, and why he thinks most companies will try closed models first but eventually shift to open ones.
00:40 Understanding open weight models
01:57 What labs actually share when they publish weights
06:33 Cost considerations in AI adoption
08:47 Security and data privacy
13:08 You can't vibe code DocuSign
15:48 The MySQL playbook for open source profits
20:33 Why the top open weight models come from China
24:09 What changes after OpenAI and Anthropic go public
25:32 The case against open weight models
30:04 Don't give an agent your passwords
31:50 Regulatory frameworks for open weight models
36:23 Education and workforce development for AI
38:41 The sign that open weight models are winning
What is an open weight model?
Shiv Shanker (@sshankar), founder and CEO of Boundless, joins the Dry Run to explain what an open weight model is, and why he thinks most companies will try closed models first but eventually shift to open ones.
00:40 Understanding open weight models
01:57 What labs actually share when they publish weights
06:33 Cost considerations in AI adoption
08:47 Security and data privacy
13:08 You can't vibe code DocuSign
15:48 The MySQL playbook for open source profits
20:33 Why the top open weight models come from China
24:09 What changes after OpenAI and Anthropic go public
25:32 The case against open weight models
30:04 Don't give an agent your passwords
31:50 Regulatory frameworks for open weight models
36:23 Education and workforce development for AI
38:41 The sign that open weight models are winning
You can also catch the episode on other platforms:
YouTube: youtu.be/uN2ZqMfJgbk
Spotify: tinyurl.com/4cfjhm76
Apple Podcasts: tinyurl.com/y8nft9rp
Dry Run retweeted
It took close to 30 years to get from ARPANET to AOL, but financial markets don't wait that long.
@austincampbell explains why the slow grind of building and adopting new technology can create bubbles along the way, even if the underlying thesis turns out to be right.
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
Dry Run retweeted
SPVs have become a commonly used structure, especially in AI financing.
The valid uses are well understood, such as letting a lender take exposure to one slice of a business, keeping a failure siloed, and managing capital rules that don't make sense.
They become questionable when used to keep obligations off the balance sheet, which contributed to Enron's downfall, or when unregulated companies keep reaching for them.
@austincampbell from @ZKZeroKnowledge
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
Dry Run retweeted
People talk about AI like any other tech, like code.
In the real world?
@austincampbell sees : Land. Power. Cooling. Chips.
People having to ask "Where do we put this stuff?"
Dry Run retweeted
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
Dry Run retweeted
Why does AI capex keep increasing?
@austincampbell gives his take. Competition is intense, chips lose value fast, and labs have to keep training to stay ahead. Every new round of spend on chips and data centers raises how much revenue they’ll eventually need to earn it back.
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
Dry Run retweeted
Really enjoyed this conversation with @austincampbell on how the AI buildout is being financed.
Austin has spent his career in risk and structured finance. A few takeaways from how he is viewing things now:
• AI already makes money, so the open question is whether that revenue can pay back the capex
• A lot of the borrowing sits in SPVs and leases that keep it off company balance sheets, which makes the risk harder to see
• If it goes wrong, he puts it somewhere between the dot-com crash and 2008
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
How is the AI buildout financed?
Austin Campbell, founder of Zero Knowledge and adjunct professor at NYU Stern, dives into the ins and outs of the AI buildout, especially how it is financed, who's doing the financing, the role of debt, and how the AI boom compares to past bubbles.
03:39 Components of the physical buildout
07:41 Capex spent so far vs. future commitments
10:13 Why Big Tech buybacks are dropping
13:18 How SPVs are structured: good vs. bad uses
20:13 AI makes money, but can it pay back the capex?
20:59 Who takes the hit if demand falls short
24:33 How today's SPVs compare to 2008
32:25 Minsky's three stages of a bubble, applied to AI
You can also catch the episode on other platforms:
YouTube: youtu.be/dW0aOgzVOWo
Spotify: tinyurl.com/52a6axja
Apple Podcasts: tinyurl.com/2bbp4ctj
We've already been building in this direction:
• Building with AI agents over long time horizons: agent memory, harnesses, and what parts of the creative process stay human (@simoncorry)
• How agents transact: x402, financial infrastructure for autonomous agents, and what changes when agents become economic actors (@murrlincoln)
• How AI is changing GTM: GTM engineering, agent-driven sales workflows, and what changes when personalized outreach becomes cheap to produce (@retttx)
• What is holding robotics back: data quality, world models, and the gap between a robot that works 80% of the time and one you can depend on (@naman_kapasi)
You can find Dry Run on X, YouTube, Spotify, and Apple Podcasts.
New episodes out every other Wednesday.
▶️ YouTube: youtube.com/@TheDryRunPod
🟢 Spotify: open.spotify.com/show/57Re2S…
🎧 Apple Podcasts: podcasts.apple.com/us/podcas…
Dry Run retweeted
In physical AI, data companies are scaling rapidly.
However, there's no standard yet for which robotics data is actually valuable. Meanwhile, well-funded, VC-subsidized labs are spending heavily to get more of it.
For now, the bottleneck is quantity, not quality.
That's changing, in the same way it did when @scale_AI gave way to @mercor. Scale sold regular labeled data in bulk. Mercor brought in experts to label very specific components, and quality started to matter.
Robotics is going through that shift now.
But if higher-quality data means companies need less of it, it's an open question whether today's data vendors survive long term.
Naman Kapasi (@naman_kapasi) on @TheDryRunPod
Why is it so hard to get data for robots?
Naman (@naman_kapasi), CEO and co-founder of Nirvana, joins the Dry Run to talk about why data quality, not intelligence, is what's holding general purpose robotics back.
02:05 LLMs to VLAs to world models
07:39 Playing air hockey with a robot
08:34 The bottleneck isn't perception, planning, or control
10:37 The tiers of robotics data quality
13:48 Scale AI vs Mercor
18:01 Nobody knows how long a humanoid lasts
23:52 Five fingered hands vs giving robots tools
27:24 At 80% nobody cares, at 99.9% it's useful
29:28 A robot can't tell if an avocado is ripe
30:09 The US just banned imported humanoids
Dry Run retweeted
Should we really be building humanoid robots?
The robotics industry has conflated two things: humanoids and physical intelligence.
A humanoid is one specific form factor: two legs, two arms, a torso, and a head.
Plenty of well-designed general purpose robots don't follow that form factor.
The case for humanoids comes down to two things:
- Locomotion: getting into tight, narrow spaces
- Manipulation: moving objects and doing the work humans do
In the long term, a humanoid is a great robot.
In the short to medium term, it's significantly harder to build than other general purpose form factors.
Naman (@naman_kapasi) on @TheDryRunPod
Why is it so hard to get data for robots?
Naman (@naman_kapasi), CEO and co-founder of Nirvana, joins the Dry Run to talk about why data quality, not intelligence, is what's holding general purpose robotics back.
02:05 LLMs to VLAs to world models
07:39 Playing air hockey with a robot
08:34 The bottleneck isn't perception, planning, or control
10:37 The tiers of robotics data quality
13:48 Scale AI vs Mercor
18:01 Nobody knows how long a humanoid lasts
23:52 Five fingered hands vs giving robots tools
27:24 At 80% nobody cares, at 99.9% it's useful
29:28 A robot can't tell if an avocado is ripe
30:09 The US just banned imported humanoids
Dry Run retweeted
World action models in robotics output the next action, not the next image.
Video models have gotten significantly better at fidelity over time by predicting the next frame. World action models work on the same principle of predicting the next state, but the state they care about is the action that needs to be taken.
In practice that means generating something like six possible trajectories, one for each action the robot could take, and committing to one.
Because the models only look a few seconds ahead, accuracy is surprisingly good. The harder problem is choosing the right action.
Naman (@naman_kapasi) on @TheDryRunPod
Listen to the full episode below.
Why is it so hard to get data for robots?
Naman (@naman_kapasi), CEO and co-founder of Nirvana, joins the Dry Run to talk about why data quality, not intelligence, is what's holding general purpose robotics back.
02:05 LLMs to VLAs to world models
07:39 Playing air hockey with a robot
08:34 The bottleneck isn't perception, planning, or control
10:37 The tiers of robotics data quality
13:48 Scale AI vs Mercor
18:01 Nobody knows how long a humanoid lasts
23:52 Five fingered hands vs giving robots tools
27:24 At 80% nobody cares, at 99.9% it's useful
29:28 A robot can't tell if an avocado is ripe
30:09 The US just banned imported humanoids