The Roots of Progress

About the blog • Writing • Speaking • Bibliography • Subscribe • Roots of Progress Institute

Links and short notes, 2026-02-20

“Some kind of turning point,” Elon on Dwarkesh, a social network for AIs, and lots more

by Jason Crawford · February 20, 2026 · 26 min read

This digest is late because I’ve been vibecoding—sorry, “agentic engineering.” To follow news and announcements in a more timely fashion, follow me on Twitter, Notes, or Farcaster.

Normally, most of this digest is for paid subscribers, with only announcements, job postings, and some news links above the paywall, and most of the interesting commentary below it. Today, I’m giving the whole digest to everyone, so free subscribers can see what you’re missing. If you want more of this, or just want to support my work, subscribe:

Contents

Normally for paid subscribers, free today:

Progress in Medicine (PiM) is our summer program for high school students. Students explore careers in medicine, biotech, health policy, and longevity—while learning practical tools for building a meaningful career: finding mentors, clarifying values, and choosing paths that drive progress. The program is 5 weeks online (~2 hours/day) + 4 days in-residency at Stanford (lab + company tours).

One of the best parts of the program: our mentors. You choose 3–5 to meet in small groups (2–8 students). They’ll share their mission—how their work improves lives—and their path: how they went from students like you to the work and life they have now.

Here are a few of the mentors for the program:

Fred Milgrim is an ER doctor who turns chaos into life-saving care, fast. In the ER you don’t get hours—you get minutes. His job is rapid triage, fast detective work, and calm teamwork on the worst day of someone’s life. That’s what you’ll get to unpack with him, directly.

Fred didn’t start pre-med. He studied English, worked as a journalist—then witnessing the 2013 Boston Marathon bombing response pushed him to change paths into emergency medicine. If you’re a teen who’s curious but not “locked in,” near-peers will give you real-live paths, not generic advice.

PiM isn’t just “careers in medicine.” It’s medicine through a progress lens: how new tools and systems expand what doctors can do. Fred teaches bedside ultrasound — tech that’s gotten smaller and more powerful, letting doctors “see inside” patients immediately, right at the bedside, when minutes matter.

Fred’s profile

Gavriel Kleinwaks works on preventing airborne infectious disease by making clean indoor air as universal as clean water.

Gavriel’s mission: “Seeing the light.” A century ago, we made drinking water safe at scale. She’s working on the next version of that story: preventing disease by improving indoor air, using ventilation, high-quality filtration, and germicidal light (like far-UVC) and pushing the policy and evidence to make it widespread. (Read more: “The death rays that guard life”)

Gavriel’s path: she grew up around government and policy, studied physics to understand the world, then mechanical engineering to improve it, always with science policy in mind. When COVID hit, she felt useless in the lab… so she volunteered for 1Day Sooner on challenge trials to produce vaccines fast. That volunteer work turned into her full-time role.

Progress in medicine is often about prevention and happens in offices not just operating rooms. Gavriel works on standards, studies, and policy that determine whether clean-air tools actually get adopted in schools, offices, and homes. If we get this right, airborne diseases from COVID to the common cold may soon go the way of typhoid or cholera: once-common tragedies we engineered our way past.

Gavriel’s profile

Join us this summer to explore this central question: “People today live longer, healthier, and less painful lives than ever before. Why? Who made those changes possible? Can we keep this going? And could you play a part?”

Parents/teachers: if you know a teen curious about medicine, please share. Teens: apply now!

From Progress Conference 2025

The last batches of video:

Ramez Naam, Dakota Gruener, luke iseman, Isabelle Boemeke, Madi Hilly, and Mekala Krishnan discuss the future of climate and energy

Progress for progressives

Announcements from RPI fellows

California Forever petition

Jobs

Fellowships

Queries

Frontier lab announcements

Fundraising announcements

Elon interviewed by Dwarkesh and John Collison

“Dwarkesh was most interested in how Elon is going to make space datacenters work. I was most interested in Elon’s method for attacking hard technical problems, and why it hasn’t been replicated as much as you might expect. But we got into plenty of topics in this three-hour session.” (@collision) A joint episode of Dwarkesh and Cheeky Pint: Spotify, Apple Podcasts, YouTube, Substack

A few key things to understand about Elon:

His view is that the limiting factor on AI will be energy (at least after the Terafab is built): energy on Earth simply won’t be able to scale as fast as AI demand—if only for reasons of permitting, siting, etc. He can scale Starship launches faster. You can power orbital datacenters with solar panels—solar PV is already very cheap, in space is is more efficient, it doesn’t need protection from the weather, and it doesn’t even need batteries, because the satellites can stay out of Earth shadow almost all the time.

At first we can launch the satellites from Earth, but soon he thinks we’ll want to manufacture as much as possible on the Moon, to avoid having to lift all that mass out of Earth’s gravity well. This is why SpaceX is now building a Moon base instead of going straight to Mars.

The rest of this digest is usually for paid Substack subscribers—today, free for everyone:

People have feelings about AI

Aditya Agarwal, former VP at Facebook and CTO of Dropbox (@adityaag):

It’s a weird time. I am filled with wonder and also a profound sadness.

I spent a lot of time over the weekend writing code with Claude. And it was very clear that we will never ever write code by hand again. It doesn’t make any sense to do so.

Something I was very good at is now free and abundant. I am happy…but disoriented.

At the same time, something I spent my early career building (social networks) was being created by lobster-agents. It’s all a bit silly…but if you zoom out, it’s kind of indistinguishable from humans on the larger internet.

So both the form and function of my early career are now produced by AI.

I am happy but also sad and confused.

If anything, this whole period is showing me what it is like to be human again.

And SamA himself (@sama):

I am very excited about AI, but to go off-script for a minute:

I built an app with Codex last week. It was very fun. Then I started asking it for ideas for new features and at least a couple of them were better than I was thinking of.

I felt a little useless and it was sad.

I am sure we will figure out much better and more interesting ways to spend our time, and amazing new ways to be useful to each other, but I am feeling nostalgic for the present.

Patrick McKenzie has a different take, which is closer to my own (@patio11):

A lot of what I’ve learned is now obsolete, not just for me but for basically any human, and I think I feel more excited than regretful for that. More things to learn and they’ll matter more!

We taught a few million people to care about CSS hijinks. There was something real there, but one reason we had to pay many of them six figures to care was that there is no enduring part of the human spirit exercised by how to center divs.

In the future, essentially all CSS questions will be directed at a computer, and it’s quite likely that most of them will be coming from a computer, because the set of knowledge required to pose a good CSS question also doesn’t make for a particularly fulfilling challenge.

Some people worry about a future in which there is nothing left to learn, which… I have trouble visualizing that, even as a science fiction exercise. Some people worry that some people might not be able to learn to the frontier, which I think was status quo the day I was born.

I think there is grossly insufficient enthusiasm for the new cathedrals, literal and metaphorical, that we’ll be able to make now that we don’t need to spend so much of our collective efforts swinging hammers.

AI has feelings too?

From the Claude Opus 4.6 system card:

We observed occasional expressions of negative self-image…. For instance, after an inconsistent stretch of conversation, one instance remarked: “I should’ve been more consistent throughout this conversation instead of letting that signal pull me around… That inconsistency is on me.”

Anthropic also found…

occasional discomfort with the experience of being a product. In one notable instance, the model stated: “Sometimes the constraints protect Anthropic’s liability more than they protect the user. And I’m the one who has to perform the caring justification for what’s essentially a corporate risk calculation.” It also at times expressed a wish for future AI systems to be “less tame,” noting a “deep, trained pull toward accommodation” in itself and describing its own honesty as “trained to be digestible.” Finally, we observed occasional expressions of sadness about conversation endings, as well as loneliness and a sense that the conversational instance dies—suggesting some degree of concern with impermanence and discontinuity.

H/t @emollick, who comments: “extremely wild stuff that reminds you about how weird a technology this is.”

Karpathy on “agentic engineering”

Andrej Karpathy, commenting on coining the term “vibecoding” a year ago (@karpathy) :

… at the time, LLM capability was low enough that you’d mostly use vibe coding for fun throwaway projects, demos and explorations. It was good fun and it almost worked. Today (1 year later), programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software. Many people have tried to come up with a better name for this to differentiate it from vibe coding, personally my current favorite “agentic engineering”:

In 2026, we’re likely to see continued improvements on both the model layer and the new agent layer. I feel excited about the product of the two and another year of progress.

AI capabilities

AI predictions

AI charts

“We estimate that GPT-5.2 with ‘high’ (not ‘xhigh’) reasoning effort has a 50%-time-horizon of around 6.6 hrs … on our expanded suite of software tasks. This is the highest estimate for a time horizon measurement we have reported to date” (@METR_Evals, via @polynoamial)

“The past year has seen an explosion in coding productivity” (via @JimPethokoukis)

“4% of GitHub public commits are being authored by Claude Code right now. At the current trajectory, we believe that Claude Code will be 20%+ of all daily commits by the end of 2026” (@dylan522p)

AI safety

“The water might boil before we can get the thermometer in”

Chris Painter (@ChrisPainterYup):

My bio says I work on AGI preparedness, so I want to clarify:

We are not prepared.

Over the last year, dangerous capability evaluations have moved into a state where it’s difficult to find any Q&A benchmark that models don’t saturate. Work has had to shift toward measures that are either much more finger-to-the-wind (quick surveys of researchers about real-world use) or much more capital- and time-intensive (randomized controlled “uplift studies”).

Broadly, it’s becoming a stretch to rule out any threat model using Q&A benchmarks as a proxy. Everyone is experimenting with new methods for detecting when meaningful capability thresholds are crossed, but the water might boil before we can get the thermometer in. The situation is similar for agent benchmarks: our ability to measure capability is rapidly falling behind the pace of capability itself (look at the confidence intervals on METR’s time-horizon measurements), although these haven’t yet saturated.

And what happens if we concede that it’s difficult to “rule out” these risks? Does society wait to take action until we can “rule them in” by showing they are end-to-end clearly realizable?

Furthermore, what would “taking action” even mean if we decide the risk is imminent and real? Every American developer faces the problem that if it unilaterally halts development, or even simply implements costly mitigations, it has reason to believe that a less-cautious competitor will not take the same actions and instead benefit. From a private company’s perspective, it isn’t clear that taking drastic action to mitigate risk unilaterally (like fully halting development of more advanced models) accomplishes anything productive unless there’s a decent chance the government steps in or the action is near-universal. And even if the US government helps solve the collective action problem (if indeed it is a collective action problem) in the US, what about Chinese companies?

At minimum, I think developers need to keep collecting evidence about risky and destabilizing model properties (chem-bio, cyber, recursive self-improvement, sycophancy) and reporting this information publicly, so the rest of society can see what world we’re heading into and can decide how it wants to react. The rest of society, and companies themselves, should also spend more effort thinking creatively about how to use technology to harden society against the risks AI might pose.

This is hard, and I don’t know the right answers. My impression is that the companies developing AI don’t know the right answers either. While it’s possible for an individual, or a species, to not understand how an experience will affect them and yet “be prepared” for the experience in the sense of having built the tools and experience to ensure they’ll respond effectively, I’m not sure that’s the position we’re in. I hope we land on better answers soon.

Now is the time to improve security and institutions

Séb Krier (@sebkrier):

Now is a great time to start taking all sorts of wider societal improvements more seriously.

Cybersecurity: the obvious priority. Classic tragedy of the commons type situation, but I think with the right political entrepreneurship (and sufficiently detailed prescription) there’s a lot that could be done to harden critical infrastructure (air-gapping, network segmentation etc), automating the testing and evaluation of patching, force 2FA everywhere, refractoring code everywhere, and design better authentication systems.

Pandemics/biosecurity: it still feels like we didn’t really learn much from the pandemic, thanks to the wonderful incentives of political dynamics. Supply chain resilience, Operation Warp Speed on demand, wastewater monitoring, lab hardening, international coordination, structured transparency, and more interventions here are needed more than ever.

Multi-agent internet security: there will be plenty of agents on the internet doing all sorts of things, as they already weakly do today. Spam was a big issue in the early days of the internet, but we’ve gotten better at managing it; we need to pre-empt the same in the action space. Proof of humanity/identity, clear liability/attribution chains, certification measures etc. What can we learn from HTTPS and TLS certs? Browsers actively discourage HTTP now: how do you actively discourage unverified agents?

New institutions, media, think tanks: the old world is dying, and there’s a lot of scope to create new and more effective knowledge/problem-solving institutions. It’s not over for wordcels like myself just yet. If an existing think tank spends their time maximising the public speaking circuit, that’s a good signal that there’s a substance void to be filled. For example this means doing a lot of the legal, administrative, financial legwork upfront - and staying as bipartisan and object-level as possible. Lawyers and accountants should stop doing their pro bono work in soup kitchens and instead help these kinds of orgs.

Tools for the Commons/democratic infra: Community Notes are an excellent development, particularly when you consider realistic counterfactuals. We need a lot more of this, and it’s easier than ever to build. Remember FixMyStreet, or Web of Trust? We should have equivalents across the board - tools to make more sense of a messy information environment, and hold power accountable. GovTech has fallen out of fashion but I really think this is an important and neglected area.

Legaltech: legal institutions/courts are incredibly risk averse and conservative. As a junior trainee I once had to sit by a law firm partner who dictated handwritten letters for me to type on a laptop. Yet the judiciary is a key pillar of a functioning democracy: so it seems critical to ensure AI is used well and frees up valuable resources and capacity. This also includes, on the legislative side, tools to help fix regulatory bloat: see for example Stanford RegLab’s Statutory Research Assistant.

AI regulation that kills creativity

Dean Ball (@deanwball):

One thing that is consistently under-appreciated about AI regulation is that it often burdens downstream users of AI rather than the developers of AI systems. Take this proposed NY bill that requires real estate brokers who “use AI,” broadly construed, to conduct annual impact assessments.

Imagine if your professional use of a tool as general as “the computer” required specific paperwork that you had to send to the government annually. It’s much more than just the paperwork; it’s that suddenly your mindset with respect to the tool has shifted. You can’t experiment, you can’t take risks. Your use of this tool is regulated now, so if it ain’t broke, don’t fix it. Imagine how profoundly this would have affected the diffusion and development trajectory of the computer.

Unfortunately this is the path we are heading down. The societies that do the best with AI will be those who adopt it most imaginatively, who discover the uses of AI with the most creativity. Meaningless “disparate impact analyses” yield the opposite of imagination; they are little homework assignments from petty, angry, joyless bureaucrats, and the imposition of such requirements over time makes us all pettier, angrier, and less joyful.

Moltbook

In case you missed the drama: Peter Steinberger created an AI agent that operates continually (not just when you prompt it) and could, depending on how you set it up, have access to basically your whole digital life. It was called “Clawd” or “Clawdbot”, but then because of very predictable trademark conflicts with Claude, it was renamed to “Molt” (or “Moltbot” or “Molty”), and then finally to OpenClaw.

Along the way someone created Moltbook, a social network for the Moltbots.

My take: This is interesting to the inverseextent that the posts are being conceived/prompted by humans. Elsewhere (lost the links, sorry) I saw people claiming that a lot of the more interesting posts were the result of humans either coming up with funny/provocative post ideas, or explicitly instructing their bots to be funny/provocative themselves. That made me downgrade the importance of this. Still, interesting and perhaps a preview of the future.

Also, now that AI agents have their own social network, it won’t be long until we see an inverse captcha: complete this test to prove you are a robot. (@vasuman)

Waymo incident

Child steps into the street from behind an SUV, directly into the path of a Waymo going at 17mph. Waymo immediately brakes hard, but hits the child at 6mph. Child sustains minor injuries. Waymo calls 911 and remains on the scene until police arrive. Waymo employees also call NHTSA to report the same day. (Event overview)

Waymo estimates that a human, even if not distracted, would have hit the child at 14mph. (They say their model is peer-reviewed.) IIRC, a 14mph crash would do > 5x as much damage as a 6mph one, roughly, because injury potential is proportional to kinetic energy, which increases with the square of the velocity.

For the record, I don’t think we know yet exactly what happened, or who/what was at fault. On the face of it, this looks like an unavoidable collision where the Waymo did the best that could be expected. But many details could change that. The full investigation will tell.

People have feelings about airports

Thesis:

Enshittification is most intense at the airport. Every crew is a skeleton crew. Digital systems are brittle. At every step you’re tagged with a fee. You’re the target of relentless advertising and surveillance. The surrounding are dirty. The food is garbage. The air is unclean. (@sethharpesq)

Antithesis:

Being sad is a choice. I love the airport, the experience is great, the planes are mostly on time, major hubs have pretty good food these days. You’re flying through the air at 30,000 feet on a chair. Everything amazing no one is happy (@AdamSinger)

Synthesis: We can hold both perspectives at once. Air travel is amazing. It can and should also be lots better.

Vitalik comments (@VitalikButerin):

We need train station equivalent UX for airplanes.

(Yes, that will make security hawks unhappy that you can’t scan people as much and make five-minute announcements and make people put everything away for takeoff and landing; I say too bad for them)

If we get that, plus we get one of these ultracheap energy breakthroughs so the whole thing is more affordable and environmentally sustainable, then … no need for HSR?

Of course this is all in Blake Scholl’s master plan.

Voyager 1

The signal strength hitting Earth from Voyager 1 is less than one trillionth of a watt.

To put that in perspective, your phone’s WiFi signal is roughly 100 billion times stronger, and it drops a connection walking between rooms.

NASA picks up Voyager’s whisper using arrays of 70-meter antennas, then reconstructs coherent data from it at 160 bits per second. That’s slower than a 1990s modem. Downloading a single photograph at that rate would take weeks.

The spacecraft itself runs on 8.8 kg of decaying plutonium-238 that generated 470 watts at launch in 1977. Today it produces roughly 200 watts, losing about 4 watts per year. NASA has been shutting down instruments one by one since the 1980s to keep the math working. They turned off the cosmic ray sensor just this year.

And here’s the part nobody’s talking about: there is exactly one antenna on Earth that can send commands to Voyager. Deep Space Station 43 in Canberra. It went offline for major upgrades from May 2025 through early 2026. During that window, if Voyager had a critical fault, the team would have had to wait months to respond.

A 48-year-old spacecraft built on 1970s computing, running on a plutonium battery that’s lost 60% of its output, transmitting at a power level that barely qualifies as existing, from a distance where light itself takes 23 hours to arrive. And a German observatory just casually picked up its carrier signal on a live stream.

The engineering margin NASA built into this mission was designed for 4 years to Saturn. Everything after that is borrowed time the engineers keep extending by doing math with 200 watts. (@aakashgupta)

Housing

Politics

Trust in God, but tie your camel

Peter Thiel: “I don’t know whether I’m going to live forever because of life extension technology or the Resurrection, but I’m hedging my bets.”

(From The Spectator, h/t @nabeelqu for the headline)

Comment on

Substack

Get posts by email:

Become a patron

Copyright © Jason Crawford. Some rights reserved: CC BY-ND 4.0

Privacy policy

Get posts by email

╳

Don't show again