I previously reported a result of 74% on BU bench v1, with the open-source BrowserAgent harness, but I forgot to set the temperature to the default specified in the model card... Now the model performs neck to neck with GPT 5.6 Luna (xhigh) and beats all other affordable models that I tested.
Qwen3.8 27B is insanely good value!
This wine is incredibly special to me for several reasons.
First, the winery that produces it is based in Armenia—my wife’s homeland—directly facing Mount Ararat. (See the second photograph, taken from their vineyards. It’s a gorgeous place.) They produce only a small number of bottles each year—around 50,000—so it can’t be found everywhere in the country. We always have to visit specific supermarkets to stock up when we’re travelling in Armenia.
The grape variety, Haghtanak—meaning “victory” in Armenian—was developed in 1977 by P. K. Ayvazyan by crossing two varieties: Sorok Let Oktyabrya from Russia and Saperavi from Georgia. (The latter is another fantastic grape variety, especially for semi-dry wines.) It produces a deep purple colour that the photograph doesn’t do justice to, with rich notes of vanilla and pronounced tannins. Overall, it offers a drinking experience unlike anything I’ve encountered in another wine , and I’m a French guy, so I’m used to drinking many different wines.
If you ever get the chance to get your hands on a bottle, I highly recommend buying it!
For this batch, I used lavender honey from my home town in South of France:
4L Water
1.5 Kg of Lavender honey
2.5g 71B yeast
I wanted to make an open source browser harness that's cheaper, more accurate and faster to run than available alternatives, without requiring a large, expensive model. Using 5.6 Luna, it outperforms the next best option (BrowserCode) on BU Bench v1 and BrowserWebApp bench. Full benchmark results available on the repository.
You can use it with your codex subscription or directly using the OpenAI API. And of course, it was built with Codex itself :)
Hi guys! I'm a computer scientist who recently found out by talking to architect friends that the tendering process requires manually aggregating data from hundreds of scanned documents, often in poor quality, originating from multiple contractors with different standards.
What are some examples of mundane, back office tasks that you'd like to automate? I'd love to build something that's useful for people
Hi guys! I'm a computer scientist who recently found out by talking to architect friends that the tendering process requires manually aggregating data from hundreds of scanned documents, often in poor quality, originating from multiple contractors with different standards.
So i built software to make it easy for people in the field to automate this sort of tedious, back office work. The video attached explains how it works.
What are some examples of mundane, back office tasks that you'd like to automate? I'd love to hear your feedback, so I can build something that's useful for people
Hi guys! I'm a computer scientist who recently found out by talking to architect friends that the tendering process requires manually aggregating data from hundreds of scanned documents, often in poor quality, originating from multiple contractors with different standards.
So i built software to make it easy for architects to automate this sort of tedious, back office work. The video attached explains how it works.
What are some examples of mundane, back office tasks that you'd like to automate? I'd love to hear your feedback, so I can build something that's useful for people
I'm a developer, so automating repetitive tasks to increase productivity comes easy. However, a lot of non technical people I know seem to be stuck doing the same repetitive tasks, without finding ways to improve productivity.
What recurring task do you still do manually even though you suspect/wish it could be automated? What has stopped you from automating them so far? (e.g restrict company policies, budget, not feeling like it’s worth the effort, other reasons?)