22 | Agent Infra & Systems @architectlabs | UIUC CS '23, Master's '25
Joined January 2024
- Tweets1.2K
- Following508
- Followers367
- Likes1.1K
The American healthcare system has a very ignoble reality: we are not achieving the optimal set of outcomes given the technological constraints and amount of resources we allocate. While the current technologies might not bring immortality, with better processes and resource allocation, people can live a lot longer and healthier lives. And the reason is because medicine is fundamentally a problem-solving discipline.
I am not saying the American healthcare system is terrible at all. It is actually probably the best in the world in treating a demographically very unhealthy population, driven by lifestyle factors outside of medicine (i.e. diet, sedentary movement, violence). The vast amount of money in the field and a lighter amount of governmental intervention in compared to other countries in healthcare also incentivizes continuous innovation. But, the healthcare system can absolutely prevent millions of premature deaths without that much effort, such as better prevention systems, and also by getting more data through better diagnostics for sick patients. And the reason we haven’t been doing that, is that the field doesn’t have enough good problem-solving people and attention spans.
The problem-solving nature of medicine originates from the 18-19th centuries when medicine becomes a formalized discipline. Our understanding of health and medicine, on the basis of the scientific method, is that it is a collection of biochemical and systematic cause-and-effect relationships.
Today, the American healthcare system is highly income-unequal: wealthier patients see better doctors and get much better outcomes. The wealthier breast cancer patients do not have access to super radically different therapies as patient on Medicare. Rather, the wealthier patient is paying for the doctor’s problem solving ability, having a better understanding of the standard of care, and that the doctor can reason through the patient’s case and find a viable path to treating the disease. And it’s not just the primary disease that requires extensive problem solving. Healthcare problem solving also is about managing secondary factors, such as comorbidities and side effects.
I very much hope that AI can better meet the need for healthcare problem-solving, especially for poor or rural patients around the country, and help close this income inequity in problem-solving.
Even as calculators have existed for a long time, and I myself use it to crunch numbers, I have always insisted on developing my mental math capabilities. This is very much because it is not every situation where I will have access to a calculator. It has become extremely useful: if someone pitches some ideas to me, and it smells like someone is full of BS, I can ask for some numbers, crunch them, and walk away.
The death of mental math is without a doubt already visible in its impacts. If Americans weren’t so terrible at mental math, Americans would be much more capable at managing their finances.
I am not looking forward on how machine intelligence erodes thinking. I am excited about the pursuit of knowledge it creates, that it can rigorously search through problem spaces, but I am not excited that it will create Swiss cheese holes in how humans understand their crafts. And this is a tool that precisely needs people to understand their crafts, because machine intelligence is not deterministic (and even if they are, poorly observable).
New essay. I am worried that we might entrust more and more of our thinking to machines, and gradually cease to sustain the habits and institutions through which we had once cultivated our minds.
writingruxandrabio.com/p/edu…
I am worried that we might entrust more and more of our thinking to machines, and gradually cease to sustain the habits and institutions through which we had once cultivated our minds.
Such a surrender might happen amid considerable prosperity and be celebrated, at every stage, as progress. Each concession would come with an excellent justification: an economy of effort, a saving of time or an answer superior to anything we could produce ourselves. Only slowly would it become apparent that, in relieving ourselves of the burden of understanding, we had also diminished our capacity to govern our affairs. And by then, a preference for machine rule might seem entirely sensible. It would materialise not through violent and forced takeover, but by our own volition, with our own diminished abilities the strongest argument for such an arrangement.
I think the scary part of LLMs here is less so about individual mathematicians being unable to compete with AI, it’s that individual mathematicians who can solve problems without the assistance of it will simply disappear. A diverse set of ways of thinking will just collapse. In this regard, math is becoming what has happened to visual arts, except even fewer has an emotional connection to it and care about it.
Today there is a Silicon Valley obsession with taste. Taste is the idea of getting people and machines to generate thoughtfully created and differentiated works instead of mass-produced, semi-deterministic set of works. If you want to acquire taste, you have to get really good at a craft, like the Japanese mochi makers. Only when you get really good at a craft, you could actually understand how to apply tools like LLMs effectively, by knowing how to plan your workflows better. Sadly, American universities have been, in my opinion, rather terrible at making students get really really good at a craft.
Let’s take writing for example. I am certainly not the best writer in the world. I was simply not trained on rhetoric nor literature. But I myself write a Substack, imperfectly, without the use of LLMs in the content. The one reason I still write on Substack is that it makes me better at reading. I not only get to express my ideas, I also get to know who doesn’t care about my time and energy by throwing me LLM slop to read.
By the way, this is not my idea that AI will damage the ability for people to think competently. This is an idea that I read authored by @RuxandraTeslo yesterday, and she could have only wrote this after years of rhetorical understanding and refinement.
The impacts will extend far into domains where human creativity will be critically necessary. Let’s take AI safety for example. You can’t delegate all of your thinking to AI if you want to competently evaluate threats to AI safety. If we look at AI safety as a branch of computer security, you need to hypothesize creatively about what failure modes AI can take, in aggregate and not just on small things that show up on evals, and reason around them.
nitter.cf/ruxandrateslo/status/2…
New essay. I am worried that we might entrust more and more of our thinking to machines, and gradually cease to sustain the habits and institutions through which we had once cultivated our minds.
writingruxandrabio.com/p/edu…
I am worried that we might entrust more and more of our thinking to machines, and gradually cease to sustain the habits and institutions through which we had once cultivated our minds.
Such a surrender might happen amid considerable prosperity and be celebrated, at every stage, as progress. Each concession would come with an excellent justification: an economy of effort, a saving of time or an answer superior to anything we could produce ourselves. Only slowly would it become apparent that, in relieving ourselves of the burden of understanding, we had also diminished our capacity to govern our affairs. And by then, a preference for machine rule might seem entirely sensible. It would materialise not through violent and forced takeover, but by our own volition, with our own diminished abilities the strongest argument for such an arrangement.
Since this quoted post had a nontrivial amount of traction, I am going to tell another story from this chapter of my life in college. While the post I quote tweeted can seem like a linear life path, I am open about the fact is not all sunshine and roses; there was plenty of struggle. And I empathize with people who have struggled.
My “liberal arts” development was probably one of the most unusual of American college students. When I was 18 years old, in 2022, I took what I thought at the time was a huge gamble: to go to Switzerland, on exchange at EPFL as a computer science student in my last year of undergrad, as a fresh young adult. I basically was forgoing recruiting and gunning for Big Tech internships because I wanted to explore the world.
I learned immensely during this time. I visited 13 countries, and was hopping between hostels, train stations, churches, museums, and town squares. I learned so much just by being in unfamiliar places, seeing people who lived very different lives as I did. While on exchange, I met students from at least 50 countries. Another more anti-establishment part of my heart was also developing, knowing how much American universities charge for liberal arts courses but not rigorously educating its students on style, rhetoric, or interpersonal skills.
I have never been to Switzerland before then. The only thing I knew about the country was that it spoke three languages, German, French and Italian, and that it is really, really expensive. My best way of putting it is, I was very glad I went, it was an extremely formative chapter of my life, however I definitely would have enjoyed it a lot more and been a lot happier if I went at 20, 21, or 22 on exchange.
I was intellectually and socially mature enough to do fine over there. But spiritually and emotionally I had some to catch up. For much of my time there, I had a language barrier. People in Lausanne spoke French. It was only towards the end of the exchange, after December, especially during the January exam period, that I finally learned and practiced enough French to be able to speak it conversationally compared to the meager public school French I studied.
At the time, I also applied to CS PhD programs at around a dozen of universities. I treated PhD applications like undergrad apps and I was NOT knowing what I was getting into. My statements of purposes (SoPs) were unfocused and aimless. And I eventually was rejected by all of them. To this day, I do not regret applying, because I found out it was not the right path in life for me.
The social environment was also very different. EPFL was much more academically “gloomy intense” than UIUC. My fellow exchange students were all at least 21-22, often 24-26, while I struggled with relating to the younger undergrad students speaking mostly in French. I was away from my happy, easygoing friends at UIUC. So, my least favorite part of going on exchange, easily, was the loneliness. And it says a lot, because I remembered being super broke at the time, I had to get creative to save money, but it never bothered me as much as being lonely.
However, it was not a horrible time, otherwise I could not have remembered it so fondly. I made friends with a few exchange students and Swiss classmates. And I also saw many places that left deep impressions on me. I learned what it meant to just trust people. People who were kind and honest made my time there a lot happier, and below I’ll show some pictures of places I visited where I really trusted people.
📍Slovakia
📍Kraków, Poland
📍Ticino, Switzerland
📍Zürich Christmas night market
When I was 17, I applied to 26 universities in the United States. I got rejected by 20, waitlisted by 3, and admitted by 3. My best offer, which to this day I am still grateful for, was UIUC.
I looked at the UIUC curriculum and found out I could speedrun the class and finish my undergrad in just four semesters. I used my AP/IB credits, overloaded some semesters, tested out of some, took summer classes, transferred general education classes from community college, and decided to go to EPFL in August 2022 to finish my last semester of undergrad there. I graduated from UIUC at 19 in May 2023.
I did this partially to save on the tuition my parents have to pay. But also because I don’t want to put off life experiences. Even if I enjoyed school, and student life, if not I wouldn’t have finished my master’s at 21 two years later, I didn’t want to spend immense amounts of time stuck in the same intro classes where I already know the knowledge. I think, with a little bit of planning and ingenuity, you don’t have to put a pause on valuable life experiences, such as solo travel while I was abroad in Switzerland at EPFL.
Here were some of my EPFL “integration week” photos around mid-September 2022, a week and a half before school started.
If you work at a startup, feeling slightly overwhelmed (though, not being burnt out) and being pulled in many different directions is not a bad thing.
It means there are enough necessary work for you to optimize your workflows and become more efficient.
I think the strongest argument against AGI right now is that current AI (especially LLMs) do not have a strong sense of time at all. Memory solves part of the time management problem, but not all of it.
And I have not yet seen how world models properly address this; they have massive mathematical and reasoning power, but it is hard to see yet how they can manage the most critical resource in the universe. If anything, we can make the argument here that self-driving systems are closer to AGI.
When I saw Terence Tao writing his blog post worrying about AI’s impact on math research, I feel sad that a lot of really smart people are losing their self-worth as AI gets better.
When I was in middle school, I thought I wanted to be a mathematician too! Then I did math competitions. I didn’t make Canadian Math Olympiad when I was in high school in Canada in 10th grade; and I barely missed the USAMO cutoff in both grades 11 and 12 after moving to Pittsburgh. Was I disappointed? Incredibly. I spent years on this craft. But I realized that the space of problems worth pushing the human frontier for is way bigger than just theoretical math.
So, I decided to major in computer science. I went to UIUC CS, took many of the most useful undergrad and grad classes, and now I am building software infra and systems for chip design with @architectlabs . It’s not an easy problem! But it is an interesting problem. There’s always more work to do, more engineering processes to rigorously build and evaluate. And wonderfully, I get to work with smart, interesting colleagues.
There’s plenty of interesting, meaningful problems to solve. For smart young people, you have both the neuroplasticity and also the lack of life + work experience commitments to choose any path you want to take. You can mold yourself down any path you find interesting. If it’s not theoretical math and its relatives (theoretical physics and theoretical computer science), there are so many problems in applied spaces, like hard tech, that desperately need smart people.
Applied spaces demand creativity and cleverness just as much as the pure theoretical spaces. And you can get paid handsomely for your creativity; after all, clever applied processes will either directly save money or make more money on the same amount of resource inputs. Even if you already have money or don’t care about money, you can impact the world really positively with your work. If you find a way to do tumor mutation testing for cheaper or manufacture CAR-T cells for cheaper, you will help a lot of people.
If you are good at designing systems, you’re in a lot of luck! Real world systems are messy. If you are smart, you can design rigorous systems and know when to make tradeoffs between doing things quick and dirty and when to systematize your processes. The world needs smart people to reason through trade-offs and allocate limited resources efficiently. There’s no shortage of meaningful problems to work through.
For promising talented people, I encourage you guys to discern through what fields excite you, what fields give you meaning in life, and be adaptable. Life is long, fear and anxiety won’t get us anywhere. But optimism and curiosity will get us a long way.
Counterpoint: AI has made the education that I received in grad school far more valuable. Granted, I only spent a year doing a professional course-based masters at UIUC.
The basic computer science principles that I learned in undergrad, I don’t think would have been as helpful in getting me a job especially compared to before AI. But I think grad school actually taught me how to think, because all of the classes were applicable to industry, after taking classes like:
- computer security
- distributed systems
- advanced algorithms
Agents can write code. But it takes humans and academic knowledge to discern what code to write, and whether the code is right for the problems you’re solving. And it is very difficult to get this fundamental understanding in industry, outside of classrooms.
High trust societies are a consequence of competent governance.
For places to be governed competently, however, you need people who are willing to take on risks. Not everything will be done perfectly; but enough people will take risks that you feel progress around you.
One thing that the solved Millennium problems (and now the alleged race to solve whether P=NP) tell us is that even after a landmark problem is solved, there is plenty of mathematical knowledge to research.
There are plenty of cases and things we don’t know, many of them in useful domains, especially considering many of these breakthrough solutions have been counterexamples. We still don’t know that much about the “under which conditions, whether something is true” cases, which will be quite interesting to explore.
We're a frontier lab attacking hardware development from all angles, and formal is another one of our directions. Excited to sponsor this hackathon as well!
This quoted post is unavailable.
Economics is so interesting because for the same issue you can have two extremely smart people who completely disagree with each other, both for valid reasons
I’m very curious how AI in biology would shape circulatory diseases.
It seems like in oncology, there is a more defined playbook today: find mutations, try tune the microenvironment, use targeted inhibitors on immunotherapy, and voilà, you have new drugs and therapies.
Circulatory diseases, such as atherosclerosis, feels like it is much more of a physics problem and less of an evolutionary behavior problem. Will be interested to see where this leads us.
This is a formulation of another relatively well-known problem in theoretical computer science, called maximum flow.
However, even though max flow algorithms run in polynomial time, they generally are quite inefficient to compute. Heuristics can probably get us most of the way there without incurring huge computational costs and poor asymptotic time complexity.
For six months, Google Maps sent a slice of drivers in ten American cities down deliberately slower routes.
About 30 seconds slower on average, and under 2% of trips were touched. The experiment ran on roughly 100 of the most congested road segments in each city, switching on and off day by day so each city acted as its own control. The results are out now in Nature Cities, from Google Research with collaborators at Berkeley and Stanford.
The reasoning behind it is old and slightly counterintuitive. Every navigation app does the same thing: it finds you the fastest route right now, for you alone. When millions of phones do that at once they all pour onto the same handful of arteries, and those arteries stop being fast. Transport economists have been writing about the gap between what's good for one driver and what's good for the network since the 1950s. Nobody had tested a fix at city scale with real drivers, because you'd need to steer routing for a large share of a city's traffic, and about three companies on Earth are in that position.
So Google put a penalty on the worst segments during their worst hours, which nudged the routing engine towards alternatives of similar road class and comparable travel time, then measured what happened across the whole network on weekdays between 7am and 8pm.
On the targeted segments, traffic moved roughly 2% faster in the median city. Los Angeles got 4.56%, Atlanta 3.30%. Fuel burn on those stretches dropped between 0.5% and 1%. Across every road that saw a change in traffic, which covers around 80% of each city's driving, speeds rose 0.35%, reaching 0.5% during the morning and evening peaks. Total travel time on affected trips fell 0.69%. Their Bayesian model put the probability that the speed effect was genuinely positive at 99.8%.
That adds up to more than 1,000 tonnes of CO2-equivalent saved per year, per city, in most of the places studied. Cars and vans account for around 10% of global carbon emissions, and the average driver spends about 2.6 years of their life behind the wheel, so a fraction of a percent across an entire road network is a serious amount of fuel.
Per driver, the saving comes to about 0.25% of an average journey, roughly one fortieth of the normal day-to-day wobble in how long the same commute takes. Nobody in those ten cities noticed anything, in either direction, including the people sent the long way round.
The authors are careful about what they haven't shown. They measured the immediate effect, not what happens once drivers work out the freed-up route is quicker and pile back onto it, which is the standard way road improvements get eaten. Their penalty scheme was deliberately crude, so the ceiling on the approach is unknown. And the underlying Google Maps data is commercially confidential, so nobody outside can check the numbers.
The finding sitting underneath all this is that a private company's routing algorithm has become a piece of transport infrastructure, tunable in the way a traffic light or a congestion charge is tunable. The obvious follow-up experiments are the ones a city government would want to run, not the ones a mapping company would.
link to full article: nature.com/articles/s44284-0…
Introducing Project Redwood 🚀🚀
@architectlabs is a frontier AI lab bringing together talent from Anthropic, xAI, Google DeepMind, and seasoned leaders across the hardware industry. We've raised a $24M seed round to build AI systems for chip design.
我們 Architect Labs 是一個 frontier AI lab ，由各家 Anthropic, xAI, Google DeepMind 還有各種硬體專家們組成，我們在做的是 ai system for chip design，目前募資 seed round $24M
-
Today, every major hardware company has its own chip design workflow—but these organizations and processes have become so large and entrenched that truly revolutionary change is difficult. We’re starting from first principles to create an AI-native chip design workflow—one that enables chip development to finally move at the speed of AI.
現在每家大硬體巨頭都有自己的晶片設計流程，但很多都大到不能做革命性的流程改動；我們在做的是從零思考，創造 ai native 的晶片設計流程，讓晶片設計真正跟上 AI 的速度。
-
Project Redwood - from a single specification, our AI system designed, verified, and deployed a chip in under two weeks—delivering 3.4× better performance per watt than NVIDIA Jetson on billion-parameter models including Llama, Qwen, and Kimi. It autonomously generated the RTL, verification, firmware, drivers, and kernels, co-designing the model, software, and silicon in one optimization loop. We acknowledge that silicon is the ultimate ground-truth. We’re taking our approach all the way to GDS. We intend to tape-out multiple improved families of Redwood co-designed for various use-cases, on TSMC.
Project Redwood - 一份規格書，我們的 AI 系統在兩週內完成晶片的設計、驗證與部署。在 Llama、Qwen 和 Kimi 等模型，其 performance per watt 比 NVIDIA Jetson 高出 3.4 倍。從 RTL、驗證、韌體、驅動到核心 kernels，全部由 AI 自己寫，並在同一個 loop 中自主設計模型、軟體與晶片。當然，流片才是最終的驗證標準。因此，我們會將這套方法一路推進至 GDS，並計畫採用台積電製程，針對不同應用場景協同設計多個持續改良的 Redwood 晶片系列並完成流片。
-
Full report on Redwood architecture and its autonomous design. Follow @architectlabs on X
我們有公開 Redwood 架構及其自主設計流程，歡迎去看論文～
We gave our AI system a spec, and in under 2 weeks, it designed, verified, and deployed a chip that beats NVIDIA.
It’s built for low-power physical AI workloads. We’re running live inference on >B+ parameter models like Llama, Qwen, and Kimi, serving at 3.4x better perf/watt than NVIDIA Jetson.
From just a specification, our AI system autonomously generated all of the RTL Design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing the model, software and silicon as one optimization loop.
Better AI can now design better chips to run AI, leading to a loop of recursive-self improvement towards our path to abundant intelligence.
Readers added context they thought people might want to know
Architect Labs did not manufacture a physical chip; the design was tested on a commercial AMD Versal FPGA board. Furthermore, the 3.4x perf/watt gain over NVIDIA Jetson is a simulation-based projection for a future 8nm ASIC, not a physical benchmark. businessinsider.com/architect-labs… architectlabs.com/architect-labs…
RT @PeterDAmbrosio: so proud of this team - @axi_master is going to change history
Just getting started and already beating NVIDIA at its…
This quoted post is unavailable.
Here’s a sneak peek into what we’ve been doing. Join us if you want to do meaningful innovation at a frontier lab!
We gave our AI system a spec, and in under 2 weeks, it designed, verified, and deployed a chip that beats NVIDIA.
It’s built for low-power physical AI workloads. We’re running live inference on >B+ parameter models like Llama, Qwen, and Kimi, serving at 3.4x better perf/watt than NVIDIA Jetson.
From just a specification, our AI system autonomously generated all of the RTL Design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing the model, software and silicon as one optimization loop.
Better AI can now design better chips to run AI, leading to a loop of recursive-self improvement towards our path to abundant intelligence.
Readers added context they thought people might want to know
Architect Labs did not manufacture a physical chip; the design was tested on a commercial AMD Versal FPGA board. Furthermore, the 3.4x perf/watt gain over NVIDIA Jetson is a simulation-based projection for a future 8nm ASIC, not a physical benchmark. businessinsider.com/architect-labs… architectlabs.com/architect-labs…