Dariush Wahdany

member of technically not staff

[Dariush Wahdany @dwahdany]

@BlancheMinerva @roydanroy Its actually only showing up under certain conditions when you click the link. Maybe for him it looks fine. I don’t think his x necessarily got hacked.

[Dan Roy @roydanroy]

@dwahdany Where does this link appear? Doesn't appear for me. This website came up when I googled for rick roll.

[Dariush Wahdany @dwahdany]

@roydanroy when i click on what you shared it goes straight to phishing for me, i am not clicking on anything on the site 🤷

[Dariush Wahdany @dwahdany]

@GergelyOrosz non-agentic cursor was peak flow

[Dariush Wahdany @dwahdany]

@merlindru uh it ends with like 100 lines of the letter "i"

this seems fine

[Dariush Wahdany @dwahdany]

i guess it's the new normal to see traces of agent hacking when just browsing the internet

(no it's not huggingface)

[Dariush Wahdany @dwahdany]

@dillon_mulroy opus 5 is so ahead of the other models i had to use /bro *twice* on the same response! never had that before

[Dariush Wahdany @dwahdany]

@giansegato wow! makes me wonder if there is some value in constraining modalities or even tools in evals, a la "constraints breed creativity"

[gian @giansegato]

the most goated example i could find when eval'ing Opus 5 was on a frontier bench very hard task. the model is given an engineering drawing of a machine part and asked to write code to rebuild it as a 3D freeCAD model. the fun catch is that the model has no way to view the drawing: mini-swe-agent is a text-only harness, so all the model ever sees is terminal text over bash IO. it's something hard and multidisciplinary even for a human engineer capable of seeing the drawing!

working blind, Opus 5 first tried off-the-shelf OCR with Pillow and Tesseract, found the results unreliable on the drawing’s thin numerals, and built its own vision stack on the spot. carved out each printed character as a connected component, rendered the glyphs as ASCII art (?!?), and matched glyph bitmaps to read the dimension labels off the page

then it vectorized the drawing itself: run-length line detection with sub-pixel centroids, least-squares circle and ellipse fits to recover radii and bores, arrowhead detection to tell dimension lines from part geometry. it worked out that the four views were printed at different scales, derived each one’s mm-per-pixel ratio, and rebuilt the part, feature by feature, as a parametric freeCAD model

welcome to the world, my dude!

[Dariush Wahdany @dwahdany]

it's over guys

[Ravid Shwartz Ziv @ziv_ravid]

NeurIPS reviews are out. At this point it's an arms race where both sides have the same supplier. You write the paper with Opus 4.6, the reviewer writes the review with Opus 4.7, and now we get to rebut with Opus 4.8. Let's just hope Opus 5 doesn't ship before the discussion period closes.

[Dariush Wahdany @dwahdany]

@NicolaiSchmid do you know which search engine they use? if you want to show up on claude they use brave, so you need to optimize there

[Dariush Wahdany @dwahdany]

@modal You literally just select the model and click create. It's magical. And with just an additional click, they automatically benchmark your endpoint for either agentic or interactive usage.

[Dariush Wahdany @dwahdany]

@modal endpoints are really good. You get a well-optimized llm endpoint with just one click, using all their inference knowledge (https://t.co/uW3joSZMlq). GLM 5.2 FP8 (a 750GB model!) running on 8xB200 pushing 600 output tok/s like it's nothing 🤷‍♂️

[Dariush Wahdany @dwahdany]

Frontier-Bench v0.1 (or in spirit, Terminal-Bench v3.0rc1) is out! Exciting new changes:
- stronger differentiation between frontier models
- regular updates: treat benchmarks like software
- auditing process of all benchmark submissions
- refined task review process

[Ryan Marten @ryan_marten]

We’re releasing Frontier-Bench: a benchmark that measures and evolves with the frontier of agent work.

Built by the team behind Terminal-Bench and Harbor, Frontier-Bench is an on-going community effort.

Frontier-Bench v0.1 contains 74 tasks on which the best agents score ~34%

[Dariush Wahdany @dwahdany]

@scaling01

[Dariush Wahdany @dwahdany]

@schizography_ @Linus__Torvalds Only when claude is producing big madness, which does unfortunately happen from time to time

[Dariush Wahdany @dwahdany]

why does 2002 @Linus__Torvalds sound like me talking to claude

[Dariush Wahdany @dwahdany]

@RhysSullivan I found the one with the left text so funny

https://t.co/3HiAb0zFnn