Pinned Tweet
Let me try and articulate what is see is a genuine tension between Bayesian subjective credences across theories of the world and what constitutes a probability calculus within a possible world bundle.
1
0
4
968
Brian Gordon retweeted
We need to stop using generic terminology. AI is not a monolith. There are critical things we need to do with AI for economic prosperity, there are things we may have to do for reasons of national competition, there are things we will at some point need to do carefully for the sake of charting and taming the computational universe, and there are things we should totally avoid doing at this stage of history, and maybe ever, if at all possible.
If we can’t be precise about these issues because we’re saying vague things like “using coding agents to make AI progress should be illegal”, then we risk failing altogether. Each of the goals above depends on achieving the others.
If we can’t prosper we can’t lead the world into the future. Everyone needs to internalize this. Democrats cannot subjugate themselves to antigrowth. Republicans cannot subjugate themselves to technocapital. Both roads lead to despair.
If we can’t be precise about these issues because we’re saying vague things like “using coding agents to make AI progress should be illegal”, then we risk failing altogether. Each of the goals above depends on achieving the others.
If we can’t prosper we can’t lead the world into the future. Everyone needs to internalize this. Democrats cannot subjugate themselves to antigrowth. Republicans cannot subjugate themselves to technocapital. Both roads lead to despair.
18
54
385
38,507
Brian Gordon retweeted
Yes
The next decade won't reward tourists. It will reward people who combine audacity with discipline. Who understand that technology cycles come and go, but that building enduring companies requires clarity of purpose and consistency of execution.
Venture has always been about backing outliers. That hasn't changed. If anything, it matters more now.
1
8
117
17,191
Brian Gordon retweeted
The one word in @RichardSSutton's Bitter Lesson that misled silicon valley.
-----
A large portion of silicon valley misunderstood the Bitter Lesson essay to think that LLMs are bitter-lesson-pilled. I won't blame them. Some poor word choices in the essay give that impression.
In particular, changing a single word in the essay would have better conveyed the message Rich wanted.
Here are the last two sentences in Rich's essay. Very powerful sentences, but read carefully.
"We want AI agents that can discover like we can, not which contain what we have discovered. **Building** in our discoveries only makes it harder to see how the discovering process can be done."
Instead of writing "Building in", Rich should have written "Feeding in."
Here's the modified version.
"We want AI agents that can discover like we can, not which contain what we have discovered. Feeding in our discoveries only makes it harder to see how the discovering process can be done."
If I'm allowed to change two words, I'd make it "Training on".
--> "Training on our discoveries only makes it harder to see how the discovering process can be done".
LLMs are the ultimate example of where we train the model with our discoveries. We are not building it in, we are feeding it in. But that doesn't matter, it still short-circuits the discovery process and makes it harder to see how the discovering process can be done.
(It's easy to see why the original wording of 'building in' gives LLMs a pass. Although Rich tried to correct this in @dwarkesh_sp podcast I don't think he fully succeeded.)
"Feeding" or "Training on" also covers another class of methods that is currently ramping up -- behavior cloning. Many people in tech think that behavior cloning is bitter-lesson pilled. But I think Rich Sutton would disagree. Behavior cloning is an example of where we are training on our (micro-) discoveries instead of letting the model do the discovering.
This also goes to the crux of what general intelligence is and how to measure it. Intelligence is not about knowing! It is about knowing how to know! Measuring intelligence should be about measuring what the model knows about how to know! More on this soon!
-----
A large portion of silicon valley misunderstood the Bitter Lesson essay to think that LLMs are bitter-lesson-pilled. I won't blame them. Some poor word choices in the essay give that impression.
In particular, changing a single word in the essay would have better conveyed the message Rich wanted.
Here are the last two sentences in Rich's essay. Very powerful sentences, but read carefully.
"We want AI agents that can discover like we can, not which contain what we have discovered. **Building** in our discoveries only makes it harder to see how the discovering process can be done."
Instead of writing "Building in", Rich should have written "Feeding in."
Here's the modified version.
"We want AI agents that can discover like we can, not which contain what we have discovered. Feeding in our discoveries only makes it harder to see how the discovering process can be done."
If I'm allowed to change two words, I'd make it "Training on".
--> "Training on our discoveries only makes it harder to see how the discovering process can be done".
LLMs are the ultimate example of where we train the model with our discoveries. We are not building it in, we are feeding it in. But that doesn't matter, it still short-circuits the discovery process and makes it harder to see how the discovering process can be done.
(It's easy to see why the original wording of 'building in' gives LLMs a pass. Although Rich tried to correct this in @dwarkesh_sp podcast I don't think he fully succeeded.)
"Feeding" or "Training on" also covers another class of methods that is currently ramping up -- behavior cloning. Many people in tech think that behavior cloning is bitter-lesson pilled. But I think Rich Sutton would disagree. Behavior cloning is an example of where we are training on our (micro-) discoveries instead of letting the model do the discovering.
This also goes to the crux of what general intelligence is and how to measure it. Intelligence is not about knowing! It is about knowing how to know! Measuring intelligence should be about measuring what the model knows about how to know! More on this soon!
.@RichardSSutton’s Bitter Lesson essay is popular in Silicon Valley, for reasons Rich didn’t intend.
Why? One poor word choice in Rich’s essay!!
Had he used a different word in one place in the essay the world wouldn’t have thought LLMs are bitter-lesson pilled.
Want to make a guess which word it is? Or you can wait till tomorrow and I’ll tell you…..
16
22
150
25,064
Brian Gordon retweeted
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
408
709
6,172
1,361,023
For two generations, western firms moved to China for labor arbitrage advantages. Today, Chinese firms are making location decisions for market access.
As Prof Malcolm said: "No, I'm, I'm simply saying that life [profits, revenue], uh… finds a way."
As Prof Malcolm said: "No, I'm, I'm simply saying that life [profits, revenue], uh… finds a way."
Caixin: "Chinese companies are flocking to Africa, but with the emphasis on building factories rather than exporting goods, as the continent’s growing population and industrialization drive create new opportunities for manufacturers facing slowing demand and fierce competition at home."
caixinglobal.com/2026-10-30/in-…
0
0
2
115
Model-based engineering, accelerated with software
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
0
0
0
115
Brian Gordon retweeted
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
153
337
4,040
403,169
Investment banking is at least a bit unique among white collar industries in that an immense proportion of the labor hours revolves around 1) proof of work and 2) pitching. Both can and do expand to maximize returns on fixed assets (analyst seats).
Just spent some time catching up with an old coworker, who now works at the VP level at a boutique investment bank
Was asking him about analyst use of AI and thought his insight was quite fascinating
More people are using AI now vs just few months ago, and it absolutely cuts down a lot of time from mundane tasks
Just tedious things like arranging logos on a deck or managing a contact log for investors would require analyst resources behind it.
Now an Associate or VP with Claude can just do it by themselves with minimal effort
However, the work hours have not gone down at all for the analysts. Since senior folks know about the capabilities of AI, they are now staffing analysts across more deals at any given time
Not only that, MDs are chasing after new mandates more often than before.
Directors are coming up with new tasks for their analysts, like maintaining tear sheets or comp sets on niche verticals. VPs are asking for more pages and trying to show more analysis to clients than before
In summary, the duration of any single task has gone down because of AI.
However, this has been offset with new work that the firm and group simply never had capacity to do before
Instead of reducing hours, it has opened up room for his group to chase after more revenue and provide better/faster work to existing clients.
Suspect this is also what is happening across many industries and companies, not just in banking or finance
Walked away with the reinforced belief that the entire job loss narrative of AI is just doomerism
0
0
2
270
Strategy, to a much greater extent than is sometimes understood, is position and context-dependent. Founder-friendly notions of governance sometimes elide the differences between genuine idiosyncratic vision and strategic thinking and choice, for example.
absolutely remarkable that Uber, one of the most disruptive innovators / technologies of modern times, has now embraced the exact political playbook it had to overcome and is attempting to use it to stop the next disruptive transportation revolution. Shameful.
0
0
4
199
The problem with teaching people about the ‘scientific method’ in elementary school is that some people seem to take this sketch super duper seriously and never gain a more nuanced understanding of science as a cognitive and social epistemological project.
i don't think people realize how quickly we're getting to the point where AI can do actual scientific research basically without humans
yesterday Nature published a paper about an AI-run laboratory in China that autonomously does DNA synthesis, amplification, sequencing, among than 11,000 experimental steps with over 20 different instruments
the humans are basically there to refill the reagents and maintain the equipment
at the same time Anthropic published research where Claude helped a team discover and determine the structure of a bacterial-virus vault particle, searching through 15 BILLION protein sequences and doing the computational analysis
just think about where this is going
AI comes up with an experiment, robots run it, instruments return the results, AI interprets them and proposes the next experiment. repeat
once that loop becomes reliable enough and generalizable, scientific progress is no longer bottlenecked by how much research humans can personally do
scientific discovery is becoming scalable
0
3
12
413
Brian Gordon retweeted
America’s "let it rip" approach to AI aligns with the technology's productivity potential and national security imperatives.
However, it has fallen short not only on safety and diffusion, but also on bringing society along.
The resulting backlash evident in the midterms was fueled by genuine concerns, information failures, and behavioral science dynamics.
Yet the alternative is not to slam on the brakes—it is to find common ground between these two extremes.
#economy #ai #tech
However, it has fallen short not only on safety and diffusion, but also on bringing society along.
The resulting backlash evident in the midterms was fueled by genuine concerns, information failures, and behavioral science dynamics.
Yet the alternative is not to slam on the brakes—it is to find common ground between these two extremes.
#economy #ai #tech
51
36
390
51,624
This is the central insight of the core capabilities literature in strategic management from the 1990s.
One of my favorite examples of Chinese tech spillovers:
China doesn’t just make solar panels. It makes solar manufacturing equipment too, including equipment for slicing silicon wafers.
This wafer-slicing equipment typically uses diamond-coated wires. China also has the world’s nd industry.
These capabilities also benefit China's silicon carbide (SiC) industry. SiC is one of the hardest industrial materials and requires specialized cutting equipment, often using diamond abrasives. Chinese companies like Gaoce have expanded from solar wafer-cutting equipment into SiC semiconductor manufacturing equipment.
SiC semiconductors are increasingly important in power electronics, particularly EVs, high-power industrial motor drives, and emerging solid-state transformers.
Solid-state transformers are a promising technology for managing complex power grids, especially those with lots of clean energy. As well as AI data centers.
(And now China is considering placing export controls on some of this solar manufacturing equipment)
0
2
2
192
Brian Gordon retweeted
DO THIS
A few ideas on trade and economic security for the new Democratic congress:
1. Legislate no new tariffs on North America without Congress's approval. Repeal 338 and prohibit new uses of 301 and 232 within North America. Trump should still renegotiate USMCA and bring a new deal to Congress to address rules of origin, etc. Given USMCA passed 89-10 in the Senate and 385-41 in the House, there should be bipartisan support for this. (AD/CVD tariffs would still apply).
2. Pass a bill excluding various price-sensitive consumer goods (coffee, etc.) from U.S. 301 tariffs.
3. Pass--but amend--the Remote Access Security Act to limit foreign access to U.S. compute, but carve in companies based in NATO allies so that the Act doesn't end up being used against close U.S. friends and partners.
4. Codify chip export controls on China and close loopholes, including the loophole that allows sales of AI chips to China in exchange for a rev-share with the U.S. government.
Over the longer term the U.S. needs to do a deep rethink of its trade and tariff statutes (e.g., make safeguards more flexible and useful and expand AD/CVD duties, while adding more oversight to 232 and 301). But that is probably not a Q1 2027 project, whereas I think the four ideas above could be....
0
3
17
4,509
This is the theory behind the design of PhD programs actually
Does trauma make artists more original? 🎨
Using data on 39,000 paintings this paper shows that when Nazi Germany invaded their country, artists painted ~⅓ less, but their work became measurably more novel and also stayed that way after the war.
ericjeong.com
0
0
1
96
Brian Gordon retweeted
What if AI makes us individually more productive while quietly weakening the networks that make collective discovery possible? That could reshape science.
Our new - very early - paper, "The Missing Colleague," with Jake Prokopets and Raymond Kim.
Comments welcome: zenodo.org/records/239397…
Our new - very early - paper, "The Missing Colleague," with Jake Prokopets and Raymond Kim.
Comments welcome: zenodo.org/records/239397…
1
1
11
326
Brian Gordon retweeted
Policy implication: we should invest in sectors that generate a continuous stream of hard problems, like for example science. I reckon that Vannevar Bush didn't think about this use case for endlessness when he wrote about the endless frontier!
0
1
4
138
Brian Gordon retweeted
AI hyperbole has consequences. "AI will take 50% of jobs" & "AI will cure cancer" are both overstatements. The positive overstatement promotes investor excitement, & the negative fear in citizens.
Combination is dangerous concoction driving a simultaneous bubble & revolution.
Combination is dangerous concoction driving a simultaneous bubble & revolution.
‘“That’s complete BS,” said Jeff Leiden, executive chairman of Vertex Pharmaceuticals. Vertex has used roughly 50 AI companies to comb through its data set of 10,000 drug compounds, Leiden said. Not one managed to generate a working drug.’ 👇🏼
wsj.com/tech/ai/silico…
72
52
377
52,213
Brian Gordon retweeted
I also think this about science more broadly. We’ve built and become accustomed to a way of doing science that we take satisfaction in and enjoy. When the goalposts (and what’s within reach) move, we’ll do it again.
In fact, we’ve done it before but the progress was slower and less noticeable. We know the feeling of our advisors/predecessors getting their PhDs for pushing the frontier of science through activities now possible via click of a button on the internet. We spent years developing skills that are now completely obsolete and replaced by more advanced methods/technologies. We are used to that level progress over generations and decades but not within years (or months, or days). It’s disorienting to be sure.
The new challenge is actually grappling with the idea that machines are getting good enough that they are meeting and exceeding human capacity in a number of dimensions. Here, it’s worth being honest about what percentage of human scientific endeavors were particularly generative in the first place. While we’re going to get lapped in many ways, I don’t think we really know what it looks like if humans are freed up to spend less of our time on context gathering/pattern matching and more of our time towards generative activities. We don’t really know what recursive self-improvement looks like for *humans* when the bottlenecks shift this rapidly. But I’m pretty excited to find out
In fact, we’ve done it before but the progress was slower and less noticeable. We know the feeling of our advisors/predecessors getting their PhDs for pushing the frontier of science through activities now possible via click of a button on the internet. We spent years developing skills that are now completely obsolete and replaced by more advanced methods/technologies. We are used to that level progress over generations and decades but not within years (or months, or days). It’s disorienting to be sure.
The new challenge is actually grappling with the idea that machines are getting good enough that they are meeting and exceeding human capacity in a number of dimensions. Here, it’s worth being honest about what percentage of human scientific endeavors were particularly generative in the first place. While we’re going to get lapped in many ways, I don’t think we really know what it looks like if humans are freed up to spend less of our time on context gathering/pattern matching and more of our time towards generative activities. We don’t really know what recursive self-improvement looks like for *humans* when the bottlenecks shift this rapidly. But I’m pretty excited to find out
If I did not believe in the intrinsic and underrated usefulness of mathematics, I would be somewhat sad about what AI is doing to it.
I think Jevons paradox will strike again: much cheaper and deeper mathematics will become even more of the fundamental infrastructure of the future than it is today.
Math will no longer be centered on conjectures posed by humans. It will be continually postulated, solved, processed, and utilized by AI systems as they work on the hard scientific, engineering, and computing problems of the future.
0
4
19
2,340
