Postdoc @MSFTResearch — exploring human-AI interaction and digital experimentation.
NYC
Joined November 2010
- Tweets220
- Following2K
- Followers1.6K
- Likes9.9K
Pinned Tweet
🚨 Out in Science this week, with @DG_Rand @duncanjwatts and Abdullah Almaatouq 🚨
We apply an integrative approach to a classic question in behavioral econ and cooperation research: *when* does peer punishment help or hinder collective welfare?
Mohammed Alsobay retweeted
PSA: If you're a student and need @ExpectedParrot credits for your research, don't create a bunch of fake accounts w/ emails and make self-referrals to get credits. Just email us! We run a very loose, highly progressive fiscal policy & can make credits by fiat. We also have a soft-spot for students!
Mohammed Alsobay retweeted
We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI.
Post-training results coming soon -- stay tuned!
Lots of details in our technical blog post.
microsoft.ai/news/frontier-p…
Mohammed Alsobay retweeted
I think I wrote exactly that, and it is still right. Organizations are a form of narrow superintelligence. My university, or Walmart, can do things that a human cannot, often using complex processes no one designed explicitly.
I will go further: the failure to deeply consider how to make AI work together with existing organizational superintelligences is a big reason why we have not yet seen any large impact of AI on scientific discovery or economic productivity despite high levels of AI ability.
I think a big part of why talks with AI-generated slides tend to suck is that there are two kinds of decks: one intended to be shared as a standalone artifact, and another that serves as a prop in a human performance.
Models are trained on this task under the implicit assumption that the two kinds are similar enough, but when slides aren’t built with effective presentation flourishes in mind (eg gradual buildups, pauses, balance of text/icons/speech), talks that rely on them fall flat.
It’d be cool to see systems that, instead, jointly propose the slides and choreography of the talk, then help users iterate on both in lockstep.
12pm: got off the Instinct waitlist
3pm: wife texts me that the cat puked on the couch
“Find highly reviewed couch cleaning services, get me a few quotes, and help me understand how they’re different”
These assistant apps have always felt like a miss, but it nailed it
“What counts as participation in the Hugging Face attack is sometimes hard to define, and was determined by an Al agent grader.”
A useful approach where some error is acceptable since all agents represent the same party (OpenAI), but will likely be a future battleground in agent-focused insurance and criminal law.
they should also ask investors how they’d feel if it suddenly went to zero
According to WSJ, Anthropic will likely tell investors that its potential revenue is above $30 trillion. wsj.com/tech/ai/anthropic-ex…
A lot of agentic SWE stories are of the form “we took a human-written codebase in X and had an agent port it to Y” or “Claude built a Node app that connects to A and uses B”.
After this transition period, when all coding is eventually done by agents, who decides what the reusable building blocks look like, and by what process? What do software libraries written entirely by agents, used exclusively by agents, even look like?
Mohammed Alsobay retweeted
Our new work on “Physics of Agents” arxiv.org/abs/2608.16578 lead by Batu El and Jinhee Paeng in collab w/ @james_y_zou
The outcome of many interacting agents seems hard to reason about. Yet we were able to study the opinion dynamics of 10,000 different LLM agent communities as they communicated with each other to solve both objective and subjective questions.
Remarkably, we could account for their opinion dynamics through a simple Ising model that involved minimizing an energy function corresponding to social pressure to conform. Our dynamics could explain the build up of consensus, polarization, and societal correction of initially incorrect majorities.
Lots more to do on statistical mechanics of interacting agent dynamics!
I want my coding agents to proactively report progress in numbers rather than telling me that a script is running in the background, e.g. if it can tell from a script that there are N inputs to process, pipe the progress from tqdm into the CLI.
Or if count is unknown upfront (e.g. scraping), show me how many items have been retrieved/generated so far. Just figure out when numbers are useful and show them to me.
Do any harnesses/agents do this well?
This is great. In reviewing a draft of mine, I was surprised that Refine caught a subtle yet important logical error in a *conceptual* figure just by parsing the image, no accompanying text.
Excited to be back at IC2S2 after missing last year’s; be sure to check out these works by @csspenn and @MSFTResearch colleagues!
📍 CSS Lab at @IC2S2 2026
The 12th International Conference on Computational Social Science (IC2S2 2026) is taking place in Burlington, Vermont, from July 28–31, 2026.
Looking forward to sharing this work with @DavMicRot, @jakehofman, and @dggoldst — see you in Philly!
Come see _the_ Mohammed Alsobay ( @malsobay ) present at AOM Tuesday August 4th in the session "From Tools to Social Actors: How Conversational AI Is Reshaping Work" 8-9:30 in 107A.
events.aom.org/og/sessions/8…
Mohammed Alsobay retweeted
just as the prophecy (our peer-reviewed research) foretold…
With 5.6 Sol, a lot of people are still prompting the model exactly as they did 5.5
It's important to note that 5.6 Sol is a lot more tenacious and thorough than previously models.
Check out the guide I wrote here to get better outcomes
learn.chatgpt.com/docs/promp…
If voice becomes the dominant mode of interaction, will we finally be rid of open floor plan offices?
