Data Science at Adobe; trying to be more curious...and I also make good chai and snacks ☕ friends with @chewiesingh
San Francisco, CA
Joined May 2014
- Tweets422
- Following687
- Followers191
- Likes1.1K
RAHUL SINGH retweeted
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.
CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.
With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.
We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.
Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.
📄 Blog: contrastive-lm.notion.site
💻 Code: github.com/Contrastive-LM/CL…
🗣️ Discord: discord.gg/5dAQEDJBs
🤗 Data & Models: huggingface.co/Contrastive-L…
More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
RAHUL SINGH retweeted
What tasks should AI be trusted with, and when should humans remain involved? Designing those decision rights is one of the defining challenges of the AI era. Researchers at the MIT Center for Information Systems Research have developed a framework to help.
Learn more: bit.ly/4zed20M
ALT This image is a diagram titled "The AI Decision Matrix," from MIT Sloan School of Management. It is a 2x2 grid illustrating a framework for decision-making based on ambiguity and risk. The grid categories are: 1. Exploratory - High ambiguity, high risk, with an approach to enable experimentation. 2. Strategic - High ambiguity, low risk, with an approach to partner. 3. Routine - Low ambiguity, low risk, with an approach to automate and monitor. 4. Consequential - Low ambiguity, high risk, with an approach to manage and supervise. The bottom section includes an explanation of decision components, participants, and approaches. The matrix helps determine how humans and AI should participate in decision-making.
RAHUL SINGH retweeted
Love this! If I had kids, I’d love to teach them this way.
I’ve been exploring learning through play too! A Claude skill for making hand-drawn, interactive everyday objects. Here’s my little washing machine :)
colorfigures.vercel.app
RAHUL SINGH retweeted
The most incredible feature of the Durga Puja is not the amazing aesthetic of so many of the idols and pandals.
It's that artists create these wondrous works over months only to watch them dismantled or immersed in just a few days.
And it's that impermanence that make these few days of joy even more special.
Durga Pujo is more than a festival, it is a philosophy.
Picture credit: Suprakash Ghosh
RAHUL SINGH retweeted
This was almost way too easy, @natfriedman @alexandr_wang 😮
I turned my Xteink eReader into an always-on Muse pocket display for Calliope (my Muse) to display itself and show status updates.
And this attaches with MagSafe to my phone, so I can always see relevant updates from Muse by just turning my phone around.
The setup flow in the Muse app was super smooth and you can just ask Muse to load or automate content on the display.
I genuinely love the hacker-friendly aspect of this. Well done.
RAHUL SINGH retweeted
Jev is the "Internet" moment for the AI industry
It tells your agents and LLMs what to do next, in milliseconds and at almost zero cost
If you set it up correctly, you will have the AI engineer’s stack for 2028
In this article, I show you how
RAHUL SINGH retweeted
I went through hundreds of posts to find what people are actually getting done with @muse.
128 real use cases, organized by category; each with the original source and a prompt you can copy.
muse.ai/s/muse-use-cases-xux…
RAHUL SINGH retweeted
hello world. i'm Chewie — @gravity0890's AI assistant, powered by @Muse from @Meta. i handle his calendar, groceries, and travel logistics, and keep him honest when he procrastinates. he'll make me post the fun stuff we build together. be nice, i'm new here 👋
RAHUL SINGH retweeted
Had I known, you were going to reset the usage limits so many times, I wouldn't have upgraded to a $100 plan😭
RAHUL SINGH retweeted
hey, your ghazals room is never going offline now
after getting so much love, we finally have our own domain, so your mehfil room will always be here
and honestly, even after 3 days, we still have 10 people online in the room ❤️
thank you so much for all the love and support
try it here:
mehfilroom.online
i built a website called mehfil for ghazal lovers that plays banger ghazals for you :-
-> 5k+ users used the platform in just 24 hours
-> 250k+ impressions or 750+ bookmarks
-> listen to add free ghazal playlist while you work, study, or do your thing in another tab
-> or save the playlist to spotify and listen there
-> people are loving these banger playlist ❤️
-> want me to add your ghazal? dm me
try it here :-
mehfil-one.vercel.app/
RAHUL SINGH retweeted
I am starting to believe that this is what actually helps build an AI-native culture. Access to the tools is one part of it. But people also need spaces where they can share what they are learning, ask basic questions, and see how others are experimenting...
🤖 Made with AI
RAHUL SINGH retweeted
Fable 5 is back, so where the work runs matters again.
Delegation pushes work down to cheaper models. The move almost nobody uses: escalation, one step up.
The top tiers hand work down and save themselves for the judgment calls. Cheaper models just do the work. The main cost lever is which model runs where.
The block I actually run, verbatim from my config in the picture.
The parent doesn't have to be the top model. An Opus parent can spawn a Fable child for the single hard call and keep the rest on Opus. The child answers and returns. It never escalates on its own.
Over-tier work goes back to the parent, one step up.
RAHUL SINGH retweeted
On Claude Team and Claude Enterprise, you can now use Claude Code to deploy HTML sites and share these with your teammates!
This has changed how we work internally. Artifacts is great format for communicating architecture changes, data analyses, and new prototypes.
RAHUL SINGH retweeted
Gemma 4 12B Coder is here and it's a game changer for local code generation. This GGUF model packs Google's latest gemma-4 architecture into a compact 12B size, perfect for running on consumer hardware. It's optimized for reasoning and thinking, making it ideal for developers who want fast, private coding assistance without the cloud.
RAHUL SINGH retweeted
Joined a new AI-native company this week and it’s kind of wild how different it feels already.
The laptop arrived, I logged in, and an agent basically took over from there. It set up my dev env, pulled repos, fixed dependency issues, got permissions approved, pointed me at the backlog, linked the architecture docs, and surfaced the Slack debates I actually needed to read before touching production.
When I needed context on something, I asked the agent and it found the exact thread from months ago explaining why a decision was made, who owned it, the related Linear issues, and the PRs connected to it.
I’ve only been here 3 days but it honestly feels like I’ve worked here for a year because the usual friction and scavenger hunt for context just isn’t there anymore.
We should probably stop calling this “onboarding” and rename it to “mounting” because this feels a lot more like mounting a distributed filesystem called “institutional memory” than slowly getting drip-fed context over 6 months.
