Eric Smith

I've recently taken a 180 w.r.t. agentic coding. It happened when my coworker put up a PR authored by Claude Opus 4.5 which implemented a complicated LLM eval workflow in one shot, and which was ~95% correct. When I leverage coding agents, code review becomes the new time-to-ship bottleneck, instead of code writing. A developer can become an architect, coordinating an army of agentic devs. And its not limited to opening a first draft PR. GitHub MCP supercharges addressing PR comments. Datadog MCP supercharges observability tasks. There are other examples. But a machine can't be held accountable for the decisions it makes. We are accountable for the code LLMs create for us. So its our responsibility to fix, iterate on, and verify what they produce, until we stand by their diffs. This is just the next step for me. Like I said, code review is the *new* bottleneck. Still thinking about how to lift that bottleneck e.g. empirical validation over opaque code via trusted testing and red teaming, etc.

27

2 Comments

Excited to share that the GRAIL-V (Grounded Retrieval and Agentic Intelligence for Vision-Language) Workshop at CVPR 2026 is now accepting submissions. This workshop focuses on advancing grounded, retrieval-augmented, and agentic intelligence across multimodal systems — bringing together research spanning Computer Vision, NLP, and Information Retrieval. 🔥 Keynotes from Kristen Grauman (The University of Texas at Austin) Mohit Bansal (The University of North Carolina at Chapel Hill) Dan Roth (University of Pennsylvania) Scott Wen-tau Yih (Meta) 🎤 Industry Panel Moderated by Sujith Ravi (VP, Oracle AI) 📢 Call for Papers Now Open If you're working on multimodal reasoning, retrieval-integrated systems, vision-language agents, or scalable AI systems, please consider submitting your work here. 📅 Submission Deadline: March 5, 2026 🔗 Workshop Site: https://lnkd.in/g9_r625X 🔗 Submission: https://lnkd.in/gjYjqPpG Proud to see colleagues at Oracle helping push forward research at the intersection of multimodal intelligence and production-scale AI. Amit Agarwal, Hitesh Laxmichand Patel, Jyotika Singh, Vivek Gupta. #CVPR2026 #AI #ComputerVision #Multimodal #Agents #Research

8

The AI industry is starting to look a lot like a food chain. Last month, Anthropic accused Moonshot AI of using hundreds of fraudulent accounts to distill Claude's capabilities to train their Kimi models. Moonshot released Kimi K2.5 as an open-weight model with a license requiring attribution for any product exceeding $20M/month in revenue. Yesterday, a developer discovered that Cursor's brand-new "Composer 2" appears to be Kimi K2.5 with reinforcement learning — the internal model ID (kimi-k2p5-rl-0317-s515-fast) leaked through the API. Cursor's parent company Anysphere reportedly does over $2B ARR. There is apparently no Kimi K2.5 branding anywhere in the product. The entire modern AI stack seems to be built on layers of borrowed intelligence. Pre-training on the internet's corpus, distillation from stronger models, RL on top of open weights and fine-tuning on synthetic data generated by competitors. At every layer, someone is building on someone else's work. These are truly interesting times. PS: Some AI was used to rewrite this.

23

🧩 Looking for a new game to try? Skeptical of AI? See if you can beat my RL agents that taught themselves Monopoly Deal! 🎉 https://monopolydeal.ai Into the technical details? Check out the paper and blog post below, detailing the platform, policy gradient methods, and "tricks" (multiple epochs per update, shared policy–value parameters, entropy regularization and decay, gradient clipping, and more...) we had to use to get this to work! 📜 https://lnkd.in/eacx8eV5 💻 https://lnkd.in/ec2YvJK3

8

OSS Bench for Agentic Coding.. in steps Cline-Bench: "A task that frontier LLMs struggle to complete are ideal candidates for cline-bench. If you find yourself having to intervene, or write code manually because cline was not able to complete the job - your task captures the failure boundary of state-of-the-art models and is precisely what cline-bench seeks to formalize." https://lnkd.in/gHcGYxiZ

6

🤔 Have you wondered which model size to fine-tune for your Agentic RL effort? Our recent paper (https://lnkd.in/gfFE8aUS) suggests that with the right reward shaping, smaller models is the smarter choice. RL fine-tuned Qwen3-8B matched its 32B counterpart while using 3.5× less compute, 1.5× less memory, and lower variance on TravelPlanner. It reached 56.9% final-pass with only 180 training queries — 2.7× stronger than GPT-5. These models also generalized well across out-of-domain tasks including Multi-IF, NaturalPlan, and Tau-Bench.

217

1 Comment