Builder of large-scale ML systems; adjunct prof @ucla; ex quant derivatives trader & startup founder. Tweets on AI, tech, science, & econ.
Berkeley, CA
- Tweets 17,742
- Following 3,685
- Followers 5,313
- Likes 62,538
Pinned Tweet
My favorite books and reading trace for 2025:
I had a five-way tie for the best books this year:
1) Mokyr’s A Culture of Growth on how societies build or block the conditions for sustained progress;
2) John Williams’ masterpiece Augustus, and the runner-up, Stoner, as intimate studies of character under the slow pressure of duty and power;
3) Dhami and Sunstein’s Bounded Rationality as the clearest bridge from “humans are cognitively weird” to “policy should be designed accordingly”;
4) Liu et al.’s Advances and Challenges in Foundation Agents as a big-map survey of how modern agents stitch together memory, planning, tools, and safety constraints into something that actually acts in the world, and;
5) von Neumann’s core computing writings (Compendium, EDVAC, Self-reproducing Automata, and The Computer and the Brain) arguing that computation is best understood as a class of automata with robustness coming from adaptation and distributed memory.
Links below.
3
4
31
12,510
Arun Rao retweeted
Eric Schmidt explains why a mutual pause of the AI race is impossible:
> US incentives will not stop, a local lab or China can cheat, and a UN check cannot read code that computers wrote or outrun US and Chinese labs.
> it cannot be verified. Inspectors would have to enter secret laboratories, including military laboratories, and review the full code base.
That code is written by computers, so no human team can read it line by line.
----
From "The Economist" YouTube channel, (link in comment)
27
30
113
13,906
OpenAI and Anthropic need to pace themselves and increase resources (compute and people) to safety efforts. If enough whistleblowers come out because they don’t think they can get leadership to invest in safe development of products, I can see IPOs getting cancelled and even labs being taken over by federal authorities and possibly shut down. I don’t want that (these are world class labs and I’m proud of them!), but it would be prudent for the Feds to intervene at some point and the labs need to treat safety as their first priority going forward.
More broadly, the rest of the industry does not have similar concerns and AI development will continue. Large labs like Google and Meta, and the 6-8 Chinese labs, are proceeding carefully and take safety very seriously and are expanding efforts and compute there. This is a story about 2 labs in SF that need to pace themselves and focus more on reallocating internal resources to safety and less on external policy efforts. They have some of the smartest and most talented engineers and scientists in the world and I think they can fix their problems. But if not, I can assure you the Feds will. The ball is in the court of @sama and @DarioAmodei.
New in The Atlantic:
@dgrobinson resigned this week. He was among the longest-tenured employees at OpenAI—and oversaw safety reports on 12 frontier launches.
He is very worried: “The time for trial and error is over.”
You can read his essay here:
theatlantic.com/technology/2…
4
5
1,417
Weekend reading: Ted Hughes’ interpretation of the Metamorphoses, “Tales from Ovid.” Brilliant and delightful.
“Before sea or land, before even sky Which contains all,
Nature wore only one mask -
Since called Chaos.
A huge agglomeration of upset.
A bolus of everything - but
As if aborted.
And the total arsenal of entropy
Already at war within it.”
3
241
Arun Rao retweeted
Xennials FTW!
The only generation snickering appropriately at AI.
26
18
258
27,347
Arun Rao retweeted
Jensen was done selling GPUs along time ago.
First came the GPU.. then CUDA.
then network
then generative ai
then sovereign
then agentic
then physical
more software, open source, financial lending mechanisms for LPS
Catch me if you can $NVDA
2
8
59
5,907
Arun Rao retweeted
We estimate AI infrastructure could soon support hundreds of millions or even billions of AI agents. At the high end, that’s enough to rival the working hours of the global human workforce.
The exact number of agents depends on the efficiency of the models they run on and whether soaring demand for AI continues. We analyzed different scenarios 🧵
ALT Chart: potential concurrent agents from 2025–27 memory shipments by model, from 30–60 million (Claude Fable 5) and 97–195 million (GPT-5.6 Sol) to 240 million (Kimi K3), 571 million (GLM-5.2) and 1.9 billion (DeepSeek V4 Pro).
66
132
720
113,720
The major political question in Western countries now isn’t left vs right but rich and old vs everyone else (esp workers and the young). The correct solution is eliminating payments and pensions to the 50% of the wealthiest retirees (or phasing it out by wealth levels sharply). Politicians don’t want to go there, but it’s the main issue as more than half of govt budgets go to retirees and the rich ones get most of it.
2
296
Arun Rao retweeted
'More striking is how fast AI took the lead. Just eighteen months ago, the best AI models fell short of the average accountant’s ~37% score. Today, models ace those same tasks.'
'These results are provocative. So much so that we considered not publishing them for fear of misinterpretation. But we think transparency about the findings matters as people and institutions prepare for rapidly advancing AI.'
“We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study.” Eighteen months ago they scored well below human accountants
Good discussion here: mercor.com/blog/human-baseli…
164
335
2,759
3,295,959
“It will be interesting to see if economic demand for intelligence in the form of AI tokens can saturate. I suspect not.”
I’d go further and say I think the space machine economy of SIs doing research and inventing stuff will be orders larger than the human economy at some point - right now it’s incipient, but humans are underestimating it.
Casey Handmer on why Starlink and Starmind are so valuable:
"If you run the math: What is the revenue per watt of transmitted microwave power for Starlink? It's a billion times higher than space-based solar power.
SpaceX is making money hand over fist with Starlink.
So you take that model and say, we make lots of money transmitting data from your content delivery service like Netflix, but instead let's transmit tokens of some hyper intelligent model.
Once you've got the hardware flying, what is the comparative value between your average Starlink bit versus your average Starmind bit? It's probably at least 1,000 times more valuable. It's not even in the same category.
It will be interesting to see if economic demand for intelligence in the form of AI tokens can saturate. I suspect not."
3
551
Arun Rao retweeted
Future population projections expect that fertility rates stop declining and roughly stabilize.
This new NBER paper asks what evidence we have to expect this. It finds there is none. "Instead, where and when fertility has fallen low, it has tended to keep falling."
nber.org/papers/w35824
11
77
305
33,941
Arun Rao retweeted
Dickinson's 4 guiding principles for governing emerging tech:
1) Default to Generality
2) Institutional Humility
3) Let Courts Do Their Work
4) Let Innovation Proceed by Default
"A system committed to both liberty and progress must place its trust not in anticipatory design, but in adaptive response to human innovation." 🙌
4
7
551
Casey and the Terraform team are building a generational company and one of the most fascinating and fun startups in the world. It brings me joy to follow these updates and I expect his company will be larger than the oil and gas majors one day, and it deserves much more reporting. Every time
I read about their milestones I hear “I'm going to have to science the shit out of this.”
CC @HenryTricks and @benmorenne.
The reactor team at @terraformindies has demonstrated successful production of high purity methanol at scale at our Muroc desert test site.
Let's review our progress in 2026. In January, we closed on our Muroc test site. By March, construction was underway. By May, we had 1.8 MW of solar installed. In July, we produced CO2 on site (x.lingyaoai.com/CJHandmer/status/20801…), and in August, hydrogen (x.lingyaoai.com/CJHandmer/status/20847…).
Terraform Industries has built an incredible team to revolutionize not just one or two key technical pieces, but the entire stack required to convert sunlight and air into fuel. Like plants, only 100 times higher productivity per acre and requiring no irrigation, fertilizer, or pesticide.
Two years ago, Terraform produced pipeline-grade natural gas for the first time. Since then, we've scaled up our methane reactor, increasing production by 250x and reducing contaminants by 10x, which allows us to increasing revenue per unit of product by 100x because it's now at chemically pure grade.
But it was always clear to me that natural gas was only one piece of the puzzle. It is possible to convert natural gas into other hydrocarbon products like gasoline or jet fuel, but given our precursors are CO2 and hydrogen, we can do one better by directly synthesizing methanol as well.
Methanol is far easier to convert into longer chain hydrocarbon products, either via the MeOH->DME->ethylene pathway, MTG, or via variations on Fischer-Tropsch.
Methanol also has significantly higher revenue per molecule than pipeline natural gas, which forms an essential part of our cost-reduction-through-profitable-scale-up strategy, explained in more detail in our post: terraformindustries.wordpres…
TL;DR: If our learning rate is above 11%, our margins will strictly increase as we scale up into a market worth about $10t/year. Not including induced demand.
Amir joined the team earlier this year and speed ran the methanol reactor development program, successfully demonstrating recirculation and heat recycling before integrating these new technologies in a production scale reactor, built from the ground up at Terraform's facility in Burbank.
As soon as the hydrogen team was done, we shipped our latest reactor up to the desert, plugged it in, and began activating the catalyst. An extended test campaign through the heat of summer followed, punctuated by bunnies nibbling on hoses!
The hard work paid off, with Terraform's production of significant batches of high purity methanol, now packaged and ready to ship. A huge congratulation to the entire team, including Amir, Enric, Dominic, Hugo, and @lucie_nurdin.
With the completion of this test, Terraform has now demonstrated TRL 9 operation of all of its subcomponents, each developed from scratch, in house, to deliver the highest possible value to the end customer.
Terraform will never rest until we have broken the geological and geographical monopoly on oil, unlocking unconditional energy abundance for all of humanity. Join us!
5
646
Arun Rao retweeted
What if an agent could decide how to structure its own computation?
Agentic Meta-Reasoning: Let the model decide what to explore, what to build on, what to verify, and where to spend its next unit of compute: effectively constructing its own computational graph as it reasons. 🧵
Paras Dahal, @anton_bakhtin , @TacoCohen , Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, @syhw , @rsalakhu , @prfsanjeevarora , @jaseweston
8
32
226
29,506
We need data providers like @OrnnExchange to release detailed charts like this as a public good so markets just see how unpredictable these curves are (eg chips rising in price over time, vs depreciating), and to watch whether the values keep going up over time as new software updates and end use cases make the HW more valuable. Whoever releases it will be the gold standard of data quality and transparency for this emerging asset class (and ideally we see it across GPUs, TPUs, etc).
1
3
452
I expect to see the entire yield curve keep shifting up as AI factories and data centers with one year payback periods and 8-10 year lives keep sucking up all large amounts of global capital. It will get more intense as compute gets securitized and becomes fungible, with standard insurance contracts and rental markets.
One of the smartest CFOs I know said that 10%+ IG rates are possible in the next few years and we should be stress testing corporate and bank balance sheets for that and beyond. I would keep portfolios low on duration and seek high cash flow stocks and floating rate bonds.
A reminder (below) of where the US Treasury curve is:
With the exception of the 2-year, every maturity on this Bloomberg screen is now above 5%.
This poses a tricky question for a range of investors:
Step in to lock in the highest nominal and real yields in years,
or
stay on the sidelines as the market contends with both a structural supply/demand imbalance among longer-term participants and unsettling volatility as "fast money" increasingly fills the void?
#economy #bonds #markets #yields a
1
4
623
Arun Rao retweeted
I’m very excited to share @nvidia Kumo Tabular, a new family of foundation models for tabular data.
Kumo Tabular establishes the new Pareto frontier across the entire accuracy–inference-time tradeoff.
Just as importantly, we are releasing it openly: open weights, open-source software, and a permissive license for commercial use.
HuggingFace: huggingface.co/nvidia/Kumo-T…
GitHub: github.com/NVIDIA/structured…
32
156
1,096
126,072