Our agent harness makes models inside Cursor faster, smarter, and more token-efficient. Here's how we test improvements to the harness, monitor and repair degradations, and customize it for different models.
LLMS
-
Critique of GDPval-AA AI Benchmarking and Artificial Analysis Transparency
By
–
i started writing about it in september, see my essay called Peak Bubble.
-
Distillation Practice in AI Should Be Covered by Fair Use
By
–
What people call "distillation" is a super common practice (you use other models to benchmark your model, to evaluate your inputs or to add a little bit to your datasets) that in my opinion should be covered by fair use (just like using public data is), especially when the
-
Falkor’s Open-Source GraphRAG SDK Tops GraphRAG Benchmark
By
–
Falkor's new GraphRAG SDK just took #1 on GraphRAG bench and yes it's open source
-
Researcher Claims Priority on World Models Concept Since 2016
By
–
I started using the concept in 2016 (e.g. in my NIPS 216 keynote, in which I called it a "world simulator"). I published papers on video prediction in 2016.
This was meant to be a key step to train world models.
Ha&Schmi appeared in 2018. The slide below is from a talk I gave -
AI Performance Is Now a Data Movement Problem Not Compute
By
–
Everyone is scaling compute. That’s not the bottleneck anymore. AI performance is now a data movement problem. Every token =
read → fetch → compute → repeat And that loop is inefficient on traditional GPUs (like NVIDIA Blackwell B200). More FLOPs won’t fix it. Dataflow -
Frequent AI Users Can Spot AI-Generated Prose by Its Tells
By
–
"Load bearing," "I keep coming back to," "Not X, but Y" A curse of using AI a lot is that you realize how much of the writing around you is just AI, now People who don't use AI have been unable to identify AI prose on sight, but those who use it a lot can spot the tells easily
-
Claude-Jupiter May Be Anthropic’s Next Sonnet-Class Model
By
–
Could "claude-jupiter" be Anthropic’s next Sonnet-class model? Sonnet 4.8 surfaced in the Claude Code source leak alongside Opus 4.7 and the Mythos/Capybara references — so this might be another internal codename rather than a random Discord rumor.
-
Smarter models make verification more valuable than prompting
By
–
The models are getting smarter and more confident at the same pace. Your verification system is now more valuable than your prompting system.
-
GPT-5.5: Smartest model but most confidently wrong benchmark results reveal flaw
By
–
GPT-5.5 is the smartest model ever tested. It's also the most confidently wrong. That's not an opinion. That's what the benchmarks say when you read both columns. Artificial Analysis runs AA-Omniscience, a benchmark designed to penalize models that guess instead of saying "I