Have you ever played an AI benchmark?

Here, have fun: ARC AGI 3 is no longer a grid puzzle. François Chollet and ARC Prize have turned them into interactive mini worlds. The agent doesn't know the rules, has to act, read feedback, build a world model and

Dr.-Ing. Martin SchieleSep 22

I asked Claude Opus 5.5 to create a webapp that turns a picture into vectors

Then blows it apart into a 3D exploded view Absolutely amazing! Took a lot of tokens though, opus 5.5 doesn't like to give up or do the job half heartedly, it's a hardworker. You can explore here:

Pradeep KapoorSep 229

What model is the best at finding vulnerabilities inside agents?

👽 Enoki Labs's attacker runs on a harness plus a model. The harness is ours. We put 10 models through it, against two agents with 23 known vulnerabilities. One writes code, one moves money. Only the model

Enoki LabsSep 223

Grok 4.7 kept the price.

It did not keep the token count. Artificial Analysis scored Grok 4.7 at 46 on the Intelligence Index, +2 over 4.6. On the Coding Agent Index it gained +9 and now sits right behind Claude Opus 5. Then the token line. Output tokens per Intelligence Index

KravnSep 22