Jan Žemlička (@JanZemlicka) header
Jan Žemlička (@JanZemlicka) profile picture

Ph.D. candidate at @uzh_df. Macrofinance, computing recursive equilibria with deep learning. Formerly student at @CERGE_EI.

Joined July 2014
Photo by Jan Žemlička (@JanZemlicka)Photo by Jan Žemlička (@JanZemlicka)Photo by Jan Žemlička (@JanZemlicka)
Looking forward to presenting my paper 'Intergenerational Consequences of Rare Disasters' (joint with @MarlonAzinovic) at this year's SED! Friday 28th, 9:30am, Q3/0013, Escola d'Enginyeria
📢Dear SED friends: the 2024 program is online!! editorialexpress.com/confere… Looking forward to seeing you in BCN for a memorable SED conference!🎊 @emanresa82 @SEDmeeting
1
19
3,200
Jan Žemlička retweeted
New blog post: The Wrong Explanation of Preconditioning Here, I show that the classic explanation for why preconditioning works in full-matrix AdaGrad doesn't make much sense. Feedback is welcome as usual. parameterfree.com/2026/11/18…
7
12
120
9,073
Chris Sims Alumni Conference at @PrincetonEcon with Tom Sargent as keynote.
23
1
277
18,515
Most people look at @OpenAI's Jalapeno and think they built a magic black box that replaces GPUs overnight, but this couldn’t be further from the truth. OpenAI's Jalapeno contains a rich set of system architecture tradeoffs and design methodology improvements that went into the chip design that all semiconductor engineers and technical leaders can learn from. Throughout this design, OpenAI leveraged an open source tool Google XLS that aided in block level optimization for PPA and formal verification leveraging their frontier models. Google XLS had its benefits in contributing to a shorter timeline, but I think the iterative, cross-domain (and often messier) nature of work is perhaps the more significant factor. This post is an advanced case study into Jalapeno based on their presentation deck from @hotchipsorg and related material. I will cover the entire architecture, including what motivated it, what tradeoffs and alternatives were likely considered by their leadership, and how AI added value. After writing this post, it makes me wonder whether AI should be best tailored to enhance the steps within the current “waterfall” chip design methodology or enable a more iterative methodology with a new skillset. siliconcodesign.com/p/an-adv…
3
29
3
259
13,089
Jan Žemlička retweeted
i think we are seeing early signs of a converging RL recipe here: >Score centering >Agentic Grader >multi harness training A bunch of recent papers have used similar techniques for post-training, also with this indication that labs are focussing on post-training compute >= pre-training compute! also the model tried to escape the sandbox during training haha
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor. Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon. thinkingmachines.ai/news/ink…
5
18
1
374
16,088
Why do banks use short-term, demandable debt? De Motta and Weitzner argue: banks have private information about asset quality; short-term debt can mitigate adverse selection when markets cannot distinguish refinancing from investment funding
1
5
26
1,786
Jan Žemlička retweeted
Join us at 2 pm ET today (Friday, November 6) for our next IAIFI Colloquium! David Pfau will present "Attention isn't all you need to solve the correlated electron problem." Watch live on YouTube: youtube.com/@iaifiinstitutef…
1
1
3
556
How well do you understand training dynamics for deep nets? Here’s a CS312 (Deep Learning Alchemy) prediction question: midway through training an LM, we temporarily lower the learning rate and then restore it. How does validation loss change relative to the original run? ↓ We train a 35M-parameter LM with a WSD learning-rate schedule (peak LR 0.003, tuned roughly to be optimal for this setting). Consider two different learning rate schedules A short LR drop: between 40-50% of training, dropping to 0.1x the original LR. A longer, shallower LR drop: between 35-55%, dropping to 0.55x the original LR. These LR drops are matched so that the total sum of the removed LRs is the same in both. How does the shape of the dip affect validation loss during the dip and after the LR returns? The image shows four qualitative predictions. A: The deeper LR drop leads to a larger loss drop and a large penalty after B: A deeper LR drop leads to a large temporary loss drop, but both match soon after. C: Both LR drops keep the loss curve monotone, but a larger LR drop leads to slower progress D: The deeper LR drop leads to a larger temporary loss drop, and this gain persists long term Try the full quiz: deep-learning-alchemy.github…
9
37
3
528
39,460
Jan Žemlička retweeted
Replying to @johannesmboehm
@johannesmboehm and Chaney combine data on half a million ancient coins with a model of trade and money to reconstruct trade and consumption from AD 325 to 950. The estimates show economic activity shifting away from the Mediterranean from the 6th century. econometricsociety.org/publi…
14
450
74
2,453
719,273
Jan Žemlička retweeted
We formalized PTX in Lean so that we can formally prove GPU kernels correct. This way, we can be confident that the kernels our agents wrote are correct, without ever needing to read them. 🧵
47
91
18
1,329
84,928
Jan Žemlička retweeted
When does Muon beat Adam? We built an exactly solvable theory of Muon vs. SignSGD The answer depends on the data: Muon wins on steep spectra with hard targets, SignSGD on flat spectra or easy targets. (ML THEORY THREAD! 1/10)
5
28
1
284
630,843
Jan Žemlička retweeted
A story about how Jane Street unwittingly reinvented datacenters in 2012, involving office server rooms, natural disasters, market maker obligations, recovery sites, colocation facilities, and old limestone telco buildings in lower Manhattan. When I joined Jane Street from college in 2010, all of the firm’s proprietary trading systems ran out of a 8’X12’ server room on the 33rd floor of 1 New York Plaza, on the southern tip of Manhattan. Jane Street’s ETF market-making, options vol pricing, intercountry ADR arbitrage, and pan-equity stat-arb operated from this small deployment. The room itself was an engineering feat: the Dell server arrays were too heavy for the 1960s building’s floors to support, so the racks were affixed to the ceiling with concrete. By 2012 Jane Street was not yet a member of most major US exchanges, even NYSE, and had virtually no presence in colocation facilities. Nevertheless, it was a ~$2B company experiencing rapid growth, a designated liquidity provider in most listed ETFs, a significant percentage of US and international equity volume, and a critical operator in capital markets. In October 2012, Hurricane Sandy devastated lower Manhattan. 1 New York Plaza, Jane Street’s home, flooded to street level and the basement power generators were fully submerged. The resulting gas leaks made the office building hazardous to enter, and power and water were weeks away from being restored. Unprecedentedly, all major US exchanges closed for two business days after the storm but would be back up by day three. Jane Street was in danger of failing its lead-market-maker obligations if it didn’t relocate its business in time. Jane Street had recently set up a partial office in a central New Jersey office park to comply with FINRA/SEC regulatory obligations for business continuity planning. A few engineers and I showed up to the empty site during the hurricane to begin mass-redeploying the firm’s software and infrastructure for the next critical 48 hours. A truckload of Dell PowerEdge R610s arrived the next morning, some salvaged from 1 New York Plaza’s server room, carried down 33 floors in pitch black by Jane Street’s IT team. For the next few months, everyone at the company from Rob Granieri to new-hires rode rented regional travel buses from Manhattan to NJ to keep Jane Street alive and growing. When we all returned to Jane Street’s repaired 1 New York Plaza office in January 2013, the technology team realized something that’s obvious now: hardware doesn’t need to be colocated with people. The NJ disaster site had accidentally become Jane Street’s first firm-wide datacenter. The site afforded significantly more space, cheaper electricity, ground-floor weight support, and abundant cooling. This decision, to leave Jane Street’s servers in post-hurricane locations, would have profound effects on the firm’s infrastructure footprint and growth. In 2013 it became a top priority for the technology and infrastructure teams to move all latency-sensitive trading into colocation facilities in order to cross-connect directly to exchange matching engine switches. And the search began for a new primary datacenter to house the firm’s research cluster and non-latency-sensitive applications. In 2015 Jane Street found its new electronic home: lower Manhattan, limestone and 32-story windows, originally owned by the New York Telephone Company. It had retina-scan security and neon-colored liquid cooling and everything else we know today. They joined other major financial firms and software companies who had operated there for years. In modeling disaster scenarios, experts tend to underrate the technological redundancies that already exist, or can be willed into rapid deployment, to ensure the US financial system remains secure.
29
57
6
810
61,050
Jan Žemlička retweeted
Replying to @farmerrf @foscaecon
The confusion that, after 60 years, still prevails about the scientific relevance (= 0) of that debate is disconcerting. Ideology has an incredibly powerful grip on the weak minds.
2
5
35
2,654
Jan Žemlička retweeted
This week in CS 312, we explored one of the most unintuitive concepts in pretraining: sharpness In this post, I'll explain what sharpness is, how sharpness appears in real training, and our unanswered questions about sharpness
12
38
7
392
58,347
Jan Žemlička retweeted
An honor to have @UChicago's @GregWKaplan on the @HooverInst @CapAndFreedom podcast talking history of macroeconomics from Keynes to HANK. We talk about the Hall-Woodford Nobel Prize, the importance of microfoundations, monetary & fiscal, the purpose of models, and Greg's book.
Why do some households spend their stimulus checks while others save them? On @CapAndFreedom, Hoover Visiting Fellow @Jon_Hartley_ and University of Chicago economist @GregWKaplan explore how differences in income, wealth, and access to cash shape the effects of interest rates and fiscal stimulus—and what that means for economic policy. Watch on Hoover+: hoover.org/plus/research/his…
19
1
78
6,119
Jan Žemlička retweeted
ok last one
🤖 Made with AI
7
27
11
186
45,481
Jan Žemlička retweeted
These frontier lab custom chips are also unusable without the internal RLed models that write all the ISA kernels. If they ever sold the chips externally, they would need to also sell API access to a special model that can program it.
Frontier labs making custom chips is not really as hard / as impressive as you might think. 1. They are primarily targeting inference not training. inference is easier 2. They are targeting transformers only. Think large MXU matmul units 3. They are writing all kernels in close to ISA with coding agents RLed for their ISA. 4. They aren’t making high level compilers for codegen A frontier lab vending a true generic chip like a Nvidia GPU or Google TPU is MUCH harder than spinning up an in house inference program. Still I think you will see many more inference chip companies as the cost of the software is just the cost of tokens.
2
8
82
8,167
Jan Žemlička retweeted
The Jalapeño chip is pretty much a systolic array with a bunch of HBM. These designs are not that crazy, but it allows people to squeeze way more performance out of their existing models. This enables two new directions for startups: 1) build models that can modify your workloads to take advantage of every nook and cranny of the available hardware 2) build new chip design workflows that enable fast development and iteration on hardware These should be happening in parallel so that workloads can be accelerated on available hardware while custom silicon is built. At the end of the day, faster chip design has to happen, because we won’t be on the transformer architecture forever, and workloads are just going to grow to be more and more complex as consumer expectations rise
Frontier labs making custom chips is not really as hard / as impressive as you might think. 1. They are primarily targeting inference not training. inference is easier 2. They are targeting transformers only. Think large MXU matmul units 3. They are writing all kernels in close to ISA with coding agents RLed for their ISA. 4. They aren’t making high level compilers for codegen A frontier lab vending a true generic chip like a Nvidia GPU or Google TPU is MUCH harder than spinning up an in house inference program. Still I think you will see many more inference chip companies as the cost of the software is just the cost of tokens.
3
15
109
8,439
Jan Žemlička retweeted
Join us at 2 pm ET today (Friday, October 23) for our next IAIFI Colloquium! Nicole Hartman will present on "Learning the real-space renormalization group via the Jarzynski equality." Watch live on YouTube: youtube.com/@iaifiinstitutef…
1
243
Jan Žemlička retweeted
Formalized textbooks: Dynamic Programming: Sargent & Stachurski (Vol 1 & 2); Stokey, Lucas & Prescott Monetary: Gali (NK); Woodford (interest & prices); Cochrane (fiscal theory). International: Obstfeld & Rogoff 500+ corrections (mostly tiny or typo) github.com/LeanEconomics
1
30
4
183
20,471
Jan Žemlička retweeted
Denmark is raising its retirement age to 70. French election candidates pledge to reducing it to 60 or 62. Denmark's debt to GDP ratio is 27%; France's is 117%. You reap what you sow. cnn.com/2025/05/23/business/…
135
1,472
211
9,934
852,555