🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

An evaluation system from AI Frontiers folks covering semantic vision & language tests.

New reqs for low to high level researcher positions: jobs.careers.microsoft.com/global/en/job/… , jobs.careers.microsoft.com/global/en/job/…, jobs.careers.microsoft.com/global/en/job/…, jobs.careers.microsoft.com/global/en/job/…, with postdocs from Akshay and Miro Dudik x.com/MiroDudik/stat… . Please apply or pass to those who may :-)

Last year, we had offers accepted from Kwangjun Ahn, Riashat Islam, Tim Pearce , Pratyusha Sharma while Akshay and Miro Dudik hired 7(!) postdocs.

From AI Frontiers, Yadong++ have released Omniparser (microsoft.github.io/OmniParser/) which parses screens better than vision models. The code is open source and the model is on hugging face huggingface.co/microsoft/Omni… .

I voted---and you should to :-)

A general purpose system of agents which roughly doubles perf over GPT4 models on computational productivity tasks. The blog post (microsoft.com/en-us/research… ) has some nice demos.

Our #Microsoft AI Frontiers team is at #NeurIPS2024 this week with several sessions on AutoGen , #Eureka, #OmniParser and #AgentInstruct #AutoGen 0.4: Redefining Agentic AI Systems - Workshop on Tuesday@12pm - Microsoft Booth: Tue@3:30pm, Wed@9:00am and Thu@3:00pm #Trace:

In addition to the workshop, I'll be around all week and am happy to chat with folks.

🚀 Phi-4 is here! A small language model that performs as well as (and often better than) large models on certain types of complex reasoning tasks such as math. Useful for us in Microsoft Research, and available now for all researcher on the Azure AI Foundry! aka.ms/phi4blog

This is a deep re-write of autogen creating modularity which in terms enables, extension, experimentation, and use over the long haul. The event-driven core is probably familiar to many folks looking at high use systems.

It turns out there is an objective-based approach for coalescing states from sequential observations, and doing so allows you to solve new problems.

A new post: Headroom for AI Development hunch.net/?p=13763046 . It's quite interesting to compare biological and silicon capabilities.

I still support Ukraine personally. All the arguments I've seen for betraying Ukraine ($, peace, nuclear war, stalemate) are bogus after looking into the details.

And it's now open source: github.com/microsoft/BST in time for ICLR.

Having tied parameters of forward and backward seems like a potentially good idea that we did not do. I'm skeptical that this alone would solve stargraph since the structure to create belief states does not exist. But please go ahead and experiment.

Agent Objectives will be at the core in the upcoming "era of experience" for working with LLMs - and BST for learning compact belief state representations is an early step towards that. Chat with the authors at #ICLR2025 to hear exciting things we are doing with it!

My understanding is that Phi-4-reasoning is functioning as well as other reasoning models which are roughly an order of magnitude larger.