VP, AI Foundations @IBMResearch, IBM Director, @MITIBMLab. Former prof @Harvard, and serial/parallel entrepreneur.
Cambridge, MA
Joined March 2011
- Tweets21.6K
- Following1.9K
- Followers11.7K
- Likes38.5K
David Cox retweeted
What if your language model could reason efficiently in an entirely new language?
We introduce Abstract Chain-of-Thought, a new mechanism which allows language models to reason through a short sequence of reserved "abstract" tokens through reinforcement learning. It is as performant as verbalized CoT at a fraction of the cost, achieving major gains in inference-time efficiency.
Looking forward to this workshop. See you in Korea!
2nd Workshop on Compositional Learning @ ICML 2026
Speakers and panelists have been finalized!
@FrancescoLocat8 @nouhadziri @Kordjamshidi @asalam_91 Sami Bengio @EltetoNoemi @neurobongo
David Cox retweeted
Inference engine is the final stage for GenAI deployment. However, it was not possible to extract or change the internal states (e.g., attention or activation). Our @vllm_project Hook serves as a modular plug-in for programming vLLM models. @IBMResearch
👉github.com/IBM/vLLM-Hook
Introducing vLLM-Hook — a plugin system for interpretable and controllable LLM inference
Check it out: github.com/IBM/vLLM-Hook
Tech report: arxiv.org/pdf/2603.06588v1
LinkedIn post: linkedin.com/posts/chingyunk…
@IRENEKO32 @pinyuchenTW
I succumbed to the temptation to write a (long form 😱) blog to share thoughts on AI. Here is the latest post: "The Ikea Era." open.substack.com/pub/neurob…
David Cox retweeted
this research makes a very strong case that mid-training is becoming its own field and dude there's been so many job listing already for research: mid-training.
good study on how mid-training approaches can be refined with what data to use + when to apply it + interaction with RL + generalization across
architectures.
I was telling my kids about how low-key terrifying commercials were when I was a kid, and my daughter made this:
This.
Cloning any random piece of SaaS is something that could already be done before agentic coding, and the economics of it haven't changed meaningfully. Before, writing the clone would cost 0.5-1% of the valuation of the legacy SaaS company. Now it might be 0.1%. It doesn't make a difference -- if you can pull it off profitably today you could also have done it profitably in the past.
The code is a very small part of the process of making such a clone successful, and the reason legacy software has often bad UX is not because code was expensive to write.
Gemini tells it like it is re: Google. (my primary gmail account is stuck in some weird state where it thinks I am a business)
David Cox retweeted
At IBM Research, we’re always interested in connecting with researchers and engineers who are passionate about building open and efficient LLMs for enterprise. If this aligns with your interests, you can share your background with our team using:
👉 forms.gle/DFoBFTAS7Cb2wsun6
I wanted to send a broadcast message over our Google Home speakers, and the Google Home app told me I had to go to the Gemini app to do that.
That didn't work, but I did get to have this unhinged conversation, so consider me a satisfied customer.
David Cox retweeted
This is Granite 4.0 Nano running on an iPhone 17 Pro with MLX Swift. This is the 350M parameter model and the 1B runs equally well.
@IBMDeveloper
David Cox retweeted
IBM AI Team Releases Granite 4.0 Nano Series: Compact and Open-Source Small Models Built for AI at the Edge
Small models are often blocked by poor instruction tuning, weak tool use formats, and missing governance. IBM AI team released Granite 4.0 Nano, a small model family that targets local and edge inference with enterprise controls and open licensing. The family includes 8 models in two sizes, 350M and about 1B, with both hybrid SSM and transformer variants, each in base and instruct. Granite 4.0 Nano series models are released under an Apache 2.0 license with native architecture support on popular runtimes like vLLM, llama.cpp, and MLX....
Full analysis: marktechpost.com/2025/10/29/…
Model weights: huggingface.co/collections/i…
@IBM @IBMwatsonx @IBMResearch @IBMData @IBMcloud @IBMDeveloper @IBMNews
David Cox retweeted
🏔️Granite 4 Nano quantization support has been added to GPT-QModel! Both H-1B and H-350M W4A16 quantized models now available on HF. 🤗 Dl link in comments. Our eval scores also validate they are indeed tier-one small models. 👇 @neurobongo
Introducing Granite 4.0 Nano, compact and open-source models built for AI at the edge.
Available in 350M and 1B, for building AI on laptops and mobile devices: ibm.co/6013Bzpt7
David Cox retweeted
IBM just released Granite-4.0 Nano, their smallest LLMs ever (300M & 1B)! 😍 The models demonstrate remarkable instruction following and tool calling capabilities, and can even run locally in-browser!
This means they can interact with websites and call browser APIs for you! 🤯
David Cox retweeted
IBM AI Team Releases Granite 4.0 Nano Series: Compact and Open-Source Small Models Built for AI at the Edge
Small models are often blocked by poor instruction tuning, weak tool use formats, and missing governance. IBM AI team released Granite 4.0 Nano, a small model family that targets local and edge inference with enterprise controls and open licensing. The family includes 8 models in two sizes, 350M and about 1B, with both hybrid SSM and transformer variants, each in base and instruct. Granite 4.0 Nano series models are released under an Apache 2.0 license with native architecture support on popular runtimes like vLLM, llama.cpp, and MLX....
Full analysis: marktechpost.com/2025/10/29/…
Model weights: huggingface.co/collections/i…
@IBM @IBMwatsonx @IBMResearch @IBMData @IBMcloud @IBMDeveloper @IBMNews