News for hackers. Stuff that matters.

jmoore15

Hi,

John here from the team that built Microsoft Decision 1 and 1.1.

Microsoft Decision 1.1 is on the Artificial Analysis Decision Index v2.1 leaderboard, where score 68, above Gemini Decision 1, GPT-6 Luna Decisions, and Jev 1.14. [1]

We also compare ourselves to other decision models in our blog post on Microsoft's Command Line blog from November 17 [2]. We name all of the public benchmarks we evaluated against in the post, and provide a breakdown by category.

[1] https://artificialanalysis.ai/evaluations/artificial-analysi... [2] https://commandline.microsoft.com/microsoft-decision-model-g...

Hi,

John here from the team that built Microsoft Decision 1 and 1.1.

We compare ourselves to other decision models in our blog post on Microsoft's Command Line blog from November 17 [1]. We name all of the public benchmarks we evaluated against in the post, and provide a breakdown by category.

We also are on the Artificial Analysis leaderboard, which is an independent third party assessment. On AA, we score 68, above Gemini Decision 1, GPT-6 Luna Decisions, and Jev 1.14. [2]

Our previous MD1 model from October is also benchmarked on JevBench, another independent benchmark for decision models. [3] On JevBench, we are 9th place for API models, ahead of Gemini Decision 1 and Jev 1.14. We look forward to seeing how the new MD1.1 scores on JevBench.

With regards to cost: Our models are the same price as Jev on a per-token basis. However, on both JevBench and Artificial Analysis benchmarks, our models use fewer tokens per task than Jev. For example, our cost score for MD1.1 on Artificial Analysis at $140 per 1M tasks compared to Jev's $156 and Gemini's $345. There are similar stats for latency, where our new MD1.1 model is 2x faster than our previous one.

[1] https://commandline.microsoft.com/microsoft-decision-model-g... [2] https://artificialanalysis.ai/evaluations/artificial-analysi... [3] https://benchmarkheaven.com/jev-models/api

There was a thoughtful reply here with some feedback on the scores. It seems to have been deleted while I was writing a reply. In the interest of substantive discussion, I'm posting my reply below, since I think it's still valuable information.

--

Yes, I noticed there were a few things off about this example (e.g., certain questions feel inapplicable for certain examples, or certain scores feel too optimistic for the bad example), and I intentionally left them there so as not to window dress things too much.

To add to some of your observations, I'll note that: 1. Automatic question generation runs before I've given any examples in this particular chat. This can be a positive (in that you can get started without even providing an example of your own data), but it also means we sometimes add questions that don't make sense for the data you actually have. The co-pilot is meant to be iterative for that reason. (As an example, towards the end of the chat, I do ask it to remove some questions that don't feel applicable).

2. The model still has to output a score for all questions, even if they don't apply to a particular input. We're working on a new system that will understand which questions actually apply, and can turn certain questions off if they're impossible to answer given the inputs provided.

3. We do get feedback from users that the scores feel off sometimes. In some cases, they're too high; in others, they're too low. We're working on an interface for calibrating the scores to your own preferences, e.g. with a small amount of thumbs up/thumbs down data. There's a tradeoff here, though, because we're also trying to make the evaluation process a lot easier than today's "prompt an LLM as a judge" paradigm where writing a prompt with a rubric can take a substantial amount of time, and any kind of calibration adds friction for users.

Overall, we release new models every other week and track ourselves on internal benchmarks to see improvements to both question generation and question scoring. If you play around with the system more and find other ways it's not working as you'd expect or like, please feel free to email me your examples and we'd be happy to prioritize looking into them.

Hey HN: I'm John, an engineer from Pi Labs. Despite lurking on HN daily for over 10 years, I've never posted, and now feels like a good time to change that :)

I joined Pi 3 months ago after a decade at Google. It was partly the HN community that inspired me to make the switch to a smaller company where I could have more direct impact. Working at a start-up has been quite an adjustment: while the work is extremely rewarding and fun, the pre- product/market fit phase is challenging in ways I've never experienced before in my career.

That's why I asked the team to post here, and am excited to show off this launch to see whether it meets a need that developers have (or learn why if not!)

Just for fun, I spent the last 5 minutes making an evaluation system for Show HN posts, so you can look at a real example if you'd rather not make your own [1]. If you sign in, you can fork and modify it, but you can also go directly to the homepage to try your own hand at it without any sign-in.

[1] https://withpi.ai/project/Xxyhrg2UR8kZHeNmbdV3