Sean Falconer

The third blog post in our AI and metadata workflows series is out -- this one is written by my colleague Brian Lavoie and focusses on workflows for institutional repositories! Read and find out -- can AI help to manage institutional repository metadata? From improving deposit workflows to cleaning up legacy data, explore how AI might support IRs while keeping human expertise front and center. Our final blog post will be out next week! Stay tuned. https://lnkd.in/gMaQjTyy

12

As COO at www.jaiinfoway.com this is an important signal for where agentic AI is heading. Non technical implementations are catching up fast but the real story is iteration speed and learning loops. Building Cowork entirely with Claude Code shows how agents are becoming first class workers not just assistants. At www.jaiinfoway.com we see this shift clearly as enterprises move from copilots to autonomous task execution with guardrails. UX will evolve tooling will mature but the direction is set. Curious how others here see non technical users adopting agent driven workflows this year. Are we ready for agents as coworkers. #Jaiinfoway #AgenticAI #Claude #AIAgents #EnterpriseAI #FutureOfWork #AIProductivity #AITransformation #AILeadership

2

New work on Rethinking Thinking Tokens: LLMs as Improvement Operators: https://lnkd.in/errhNuuz Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost. The paper develops a simple yet effective Parallel-Distill-Refine (PDR) procedure: Generate diverse drafts in parallel, Distill them into a compact textual workspace, and Refine conditioned on this workspace. This decouples context length from total token count, allowing control over compute via parallelism. PDR yields higher accuracy than long CoT at lower latency. Training an 8B model with RL to align with PDR further shifts the Pareto frontier. On math benchmarks, PDR achieves +11% (AIME 2024) and +9% (AIME 2025) over single-pass baselines. With Lovish Madaan, Aniket Didolkar, Suchin Gururangan, John Quan, Ruan Silva, Manzil Zaheer, Sanjeev Arora, and Anirudh Goyal.

78

1 Comment

Barbarians at the Gate (https://lnkd.in/gkEBX79s) is a very interesting new paper from some great folks at Berkeley, with some exciting results showing the potential for AI in systems research. But I think the authors aren't quite asking the hardest problem about where this takes systems as a field. I wrote a new blog post about it: https://lnkd.in/gxS3s4v4 The core point of the paper is that AI can be effective at automatically driving the 'number go up' part of systems research. Given an evaluator, AI can automate progress on implementation. I believe this is 100% correct. The paper also talks about how more focus will have to go into developing evaluators. Again, I 100% agree, but think they're understating their case. Building evaluators (basically tests) is, I believe, going to become the focus of systems and software over the next decade. Which leaves researchers with the hard problem of choosing which problems to go after. That's never been easy in academic systems, and is only going to become more of a challenge. But the best systems research is driven by choosing problems well! This vision for AI could lead to lots more of the least interesting systems research. The "number go up" kind. Which could make it harder than ever to identify the truly innovative and visionary work. We're already seeing that elsewhere in science. So, mostly, I think the paper is right, but not yet asking the biggest question: what should systems research be?

276

12 Comments

I found this short piece by Lili Jiang on evals very relatable. It applies equally well to physical AI as it does to LLMs: https://lnkd.in/eVCTJpJm

75

3 Comments

One of the most underappreciated pain points in LLM RL training is policy staleness. When your generation and training loops run asynchronously — which they must at scale — the policy that generated your rollouts drifts from the policy you're actually updating. Most teams paper over this with token-level clipping (PPO-style) or just accept the instability. VESPO from RedNote takes a different approach: sequence-level importance weight reshaping with a closed-form variational solution. No length normalization hacks. The key claim — stable training even at 64x staleness ratios — is striking, and they show it holds for both dense and MoE architectures. From my team's work on RL post-training, I can say that the staleness problem is very real and gets worse as you scale. We've seen training runs destabilize when the generation queue backs up and policy lag exceeds ~8-10x. The usual fix is aggressive clipping or just throwing away stale batches, both wasteful. A principled sequence-level solution that handles extreme staleness would genuinely change how we architect async RL pipelines. What I'd want to see next: how VESPO interacts with reward model staleness (not just policy staleness), and whether the variance reduction holds up on tasks with sparser rewards like long-horizon coding or multi-turn dialogue. Paper: https://lnkd.in/g-X2DvGb #LLM #ReinforcementLearning #RLHF #MachineLearning #AIResearch #PostTraining

11

As always, Ben Lorica provides worthwhile deep dive - this time into Ray's distributed computing framework. Buckle your seatbelts for a report on real constraints, and the trade-offs teams are making: "How they’re scheduling scarce GPUs, wiring multimodal data flows, hardening reliability on flaky hardware, and speeding the post-training loop that now drives most gains."

1

Scaling inference for LLMs presents a unique set of challenges, very different from how we've historically scaled traditional applications. This is an amazing post by the llm-d contributors showing how they're tackling this head-on. They detail a smart, AI-aware scheduling approach that considers both KV-cache hits and dynamic server load to dramatically boost performance and reliability.

15

1 Comment

See how the release of Llama4 models, Scout and Maverick, on OCI Generative AI service are optimized for multimodal understanding, multilingual tasks, coding, tool-calling, and powering agentic systems. https://lnkd.in/gRqmRE6b

3

1 Comment