Justin Waugh

@LLMJunky @mweinbach Dell via call for quote (the video above is of the Dell version, hence the "them") I saw the cdw thing as well, but not clear that that is real / actually available. I haven't heard back directly from any other supplier (there are 6 that were on the NVIDIA website)

@JustinWaugh @mweinbach who is them? they are ~$97K generally cdw.com/product/msi-nv…

Does anyone have pricing on this I? A couple months ago I was betting like $50-75K

@Ex0byt @MichaelDell Quoted at 169k + 14k tax, and 60 day to delivery date.

Take my money @MichaelDell

@mweinbach (called them and got this number, precisely: 169,005.32 base, 183,029.39 with tax) i was seriously ready to buy at ~80k-100k price point, but was shocked to hear the price they are asking.

@mweinbach I got quoted at 170k + 13k taxes (60 day delivery)

@thsottiaux Codex stops early way too often, and i have no way to meta-prompt it to wake itself up on a schedule / keep iterating like i can with claude code

What are we consistently getting wrong with codex that you wish we would improve / fix?

@StirlingForge @NVIDIAAIDev I got a quote: 170k (+14k tax) w/ 2 month delivery time

@NVIDIAAIDev HOW MUCH?!

NVIDIA DGX Station is now available to order from select OEMs🔥 Powered by the GB300 Grace Blackwell Ultra Desktop Superchip, DGX Station brings data-center-class AI performance to the desk — enabling developers to build and run autonomous AI agents locally. ⚡ 748GB of coherent memory ⚡ Up to 20 petaflops of AI compute performance ⚡ Run large open models up to one trillion parameters Together with NVIDIA NemoClaw — an open source stack that simplifies running OpenClaw always-on assistants, more safely, with a single command, we are delivering a full-stack platform for secure, long-running agentic AI. Learn more: #dgx-spark-station" target="_blank" rel="nofollow noopener">blogs.nvidia.com/blog/gtc-2026-…

@towheretobegin Tokyo transit isochrone I made

When I moved to new york, I found it hard to visualize what commute times actually looked like. The same dilemma occurs every time you move, or even book a hotel: what's actually accessible in 20 minutes of public transit? Deployment link below

@_chenglou Very cool~ I wonder if it would generalize well across more puzzle varieties (eg. those in pencil puzzle bench ppbench.com )

I’m very happy to present my toy research project: Sotaku! It's a neural net that automatically discovered the rules of sudoku and learned to solve them, achieving a new state-of-the-art score of 98.9% on one of the hardest sudoku datasets, while being agnostic to the game, and beating all other sudoku-optimized neural net architectures* Read more for fun motivations, plus some extremely unconventional discoveries, e.g. reverse curriculum consistently beating curriculum (!), emergent reasoning-like capabilities, and the future of traditional programming

@carterwmckay @scaling01 ppbench.com

@scaling01 Where are you finding the puzzles, all the sites I tried had some broken puzzles, 0 and 1 points directly at each other.

felt good about myself after I solved 3 Yajilins in 5-9 minutes each for puzzles that models couldn't solve then encountered thanos on the next one aaaand im down to 40 minutes again it's over

@AndilesAnthony @scaling01 For this puzzle specifically here's the trace: @xhigh&puzzle=yajilin_de2cf706b2ff47627cc6ded790ff3de4" target="_blank" rel="nofollow noopener">ppbench.com/replay.html?mo… it took ~30 minutes thinking and then one-shot it, haha.

I wrote the benchmark and ran gpt-5.4-xhigh test shown here. Paper is available on arxiv: arxiv.org/html/2603.0211… website has full traces of the model on all the puzzles @xhigh.html" target="_blank" rel="nofollow noopener">ppbench.com/model/gpt-5.4@… (full details available on huggingface) For your actual questions: GPT-5.4@xhigh was run against these puzzles using the API and with this basic-agentic harness, defined here. github.com/approximatelab…

just tried another level of this benchmark that only GPT-5.4-xhigh could solve i solved it, but it took me 44 minutes GPT-5.4-xhigh did it in 38 minutes ppbench.com/share/QVBZEEtF…

@scaling01 That was a fun one one for sure! Took me 17:05 (and I consider myself pretty good at these / have done many yajilin before) (also, thanks for posting this, prompted me to try it, led to learning the open-graph preview for share had a bug on X!) ppbench.com/share/QVBZEEhD…

Exponential improvements* everywhere for those with the eyes to see them. This is a cool benchmark, and was impossible for early non-reasoner LLMs to do at all. * Okay, technically "logistic improvement" because the maximum score is bounded at 100 (and logistic has a lower AIC)

@ChristosTzamos I recently released pencil-puzzle-bench. Awesome to see so many steps / decoding as a computer. Would be interested to see if it can adapt solutions for many puzzle types, not just sudoku as shown. ppbench.com

1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy

@YafahEdelman pencil puzzle bench recently crossed 30%, but each LLM sucess is still expensive/slow ($5+, 10min+), especially compared to CPU based SAT solvers (custom to problem) that are ~7-8 orders cheaper and ~3 orders faster Huge efficiency gains still remain ppbench.com

Okay, what benchmarks are still under 30%?

Just released the harness and code I used for Pencil Puzzle Bench! (ppbench.com) Also has a gym environment and a verifiers environment Check it out on github github.com/approximatelab…

@eigenform Took this picture the other day

latest installment in "is this a factorio screenshot or old intel p-core layout"

@hardmaru theres still time

In an alternate timeline we’d be using Evangelion GUI designs rather than CLIs

@j_dekoninck Saw something similar with pencil puzzle bench: gpt-5.2@xhigh out performed gpt-5.2-pro. (gpt-5.2-pro scored closer to 5.2@medium) ppbench.com

I know about some other benchmarks where GPT-5.4-Pro seemingly does not outperform GPT-5.4 by all that much, but this clearly shows it's at least better in some areas :)

One more: We now added GPT-5.4-Pro to ArxivMath February and Apex. Extremely expensive, but it is SoTa on both by quite a significant margin

Just added GPT-5.4 to pencil-puzzle bench results! Large uplift (70% solved now). Longest single success for a puzzle took 95 minutes and $17.90 in inference alone See the full breakdown, play the puzzles, and look at the 5.4 traces here: ppbench.com