I'm extremely proud of what we've been able to achieve with Huff. Thank you to all of the hard-working contributors who have helped turn this into something special.
We're just getting started, I can't wait to see what the future holds for us.
1/
🔥 Today we're thrilled to release Huff, the most gas-optimized smart contract language to ever touch the EVM
🏎️ Huff enables you to optimize your contracts by up to 80% compared to other languages like Solidity and Vyper
👨💻 Check it out at huff.sh/
Two reasons Cignetti would say this:
USC as a team is better built than last year and the best talent are freshmen.
Riley has had two timelines at USC. Quick build around Caleb and rebuild from ground up afterward.
I gave a talk at 200ms (formerly MEV Day) in London yesterday. Here are my 2 most important slides.
Nobody has any business building an exchange without understanding its economics, and the most important prerequisite is internalizing that not all activity is equal.
🚨GAME CHANGER🚨
Introducing “SMART” — the new air traffic management tool that saves YOU time, cuts back DELAYS, and makes your flying experience BETTER
MORE TO COME ✈️
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
some thoughts on writing software in nov 2026:
- it feels like magic
- the first 80% is 100x faster
- rewriting systems is also 100x faster
- opus 5.5 is very good at a lot of things
- "deep flow state" is much easier to reach
- almost anything can be built in typescript and rust
- building secure things is very, very hard
the optimization function is how to maximize agents working on goal with the fewest interruptions needed. some things that help:
- know where to trust vs guide (and keep retesting)
- give them agency like "i'm going to bed, get this done"
- if there is a reward function, ask them to hillclimb
- the socratic method works wonders
- skills and memories can often hinder
- asking agents to review themselves still helps
- cowriting a plan/spec/invariants helps sometimes
managing your own psychology is equally important:
- be a lot more ambitious about what you can achieve
- avoid dopamine traps / false signals of progress
- get good at parallelizing and context switching
there are also some things that don't quite work fluidly yet, but will get unlocked in the very near future:
- software factory or "find and fix the bugs"
- multiplayer agentic workflows
- agents managing other agents for high level outcomes
- organizational memory
Today, 1 in 2 Ethereum blocks will be protected.
With Phylax, apps can set circuit breakers that Titan Builder enforces at block construction, excluding transactions that violate app-defined rules before settlement.
It's a simple idea: hacks should not settle.
@HyperliquidX's IOC speed-bump mechanism has a flaw. Normally designed to protect makers in the interests of a more liquid market, it breaks down in periods of high activity. When blocks become full, pending modifies and cancels queue up for later blocks, so their latency increases. However, IOCs behave as though they have a separate blockspace dedicated for them, and their latency barely budges. You can see this in the lower left chart, showing latency of different order types during 19 activity bursts.
The result is that under relatively modest load (25-30k tps), IOC + 8 bps (priority fee) orders can pick off resting liquidity before it has a chance to get out the way. Meanwhile, under higher load (35-40k tps), even IOCs without any priority fee can pick off resting liquidity.
This, in my opinion, is antithetical to the ethos and design of the ALO and cancel prioritization mechanism and needs investigation by the team. It exists to reduce toxicity from HFT takers, and is breaking down precisely when most needed.
It may also be exploitable. If IOCs really are insulated from the congestion affecting cancels and modifies, a malicious user could spam orders, delaying makers’ cancels while their own IOCs go through, and taking the liquidity left stale.
One single example that occurred on AVAX today caused 1.3M$ of volume to be adversely selected for a total loss of ~13k$ over the subsequent 10 seconds. This affected us materially. Faced with such episodes of adverse selection, spreads and liquidity are undoubtedly worse than they could be otherwise.
Would appreciate if anyone with a contact at HL could share this with the team.
and look how it’s gone for them:
- 95% of nonhuman mammals are livestock
- insect biomass reduces by 2% every year
- farmed animals live in crowded cages worse than any prison
- wild animals have their hunting, migration, and habitats disrupted by human light, sound, and trash pollution
outside of house cats and dogs, the industrial revolution has been disastrous for animals. even pets may live “happy” lives but are neutered, separated from their families, housed indoors, and fed slop.
this is despite humans generally being “aligned” with animal wellbeing.
From the perspective of virtually every other species on Earth, they’ve already been sharing the planet with a superintelligent, recursively self-improving x-risk for millions of years…
us.
Kalshi and Polymarket’s fees are now so high you can’t actually trade. If you actively trade, you’ll erode any edge.
Consider the economics of trading a football game.
At 50¢, Kalshi charges 1.75¢ per taker execution.
A trader entering and exiting a position 3 times over the course of a game makes 6 executions.
That’s 10.5¢ in Kalshi fees, before the spread.
Assuming a 1¢ bid/ask spread, total execution costs across those 3 round trips are 13.5¢ per contract.
27% of your committed capital.
On Kalshi, a taker trading near 50¢ must capture more than 4.5 percentage points of favorable repricing per round trip just to break even.
Suppose a trader can consistently identify 2% mispricings in live sports markets.
This trader has genuine predictive information. In a frictionless market, exploiting that information would be profitable.
On Kalshi or Polymarket, the same strategy is negative EV if executed through taker round trips.
A trader who continuously incorporates new information into prices pays repeatedly.
The fee structure therefore selects against active price discovery.
Bullish on lower fee markets that enable better price discovery. Bearish on exchange extraction.