Mark Pollack

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

TL;DR Mean matters in GRPO-style RL training – use geometric mean (GM) instead of arithmetic mean (AM) in calculating the advantage of rollouts in GRPO. GM is inherently less sensitive than AM to outliers, which stabilizes the RL training for LLMs. == Introducing GMPO, a stabilized variant of GRPO. Instead of optimizing the arithmetic mean, GMPO maximizes the geometric mean of token-level rewards, which is inherently less sensitive to outliers and maintains a more stable range of importance sampling ratio. We provide comprehensive theoretical and experimental analysis to justify the design and stability benefits of GMPO. Beyond improved stability, GMPO-7B outperforms GRPO by an average of 4.1% on multiple mathematical benchmarks and 1.4% on multimodal reasoning benchmark, including AIME24, AMC, MATH500, OlympiadBench, Minerva, and Geometry3K. https://lnkd.in/grAeW9Hp

310

3 Comments

Andrew Ng reported from Davos, that AI deployment for transformative impact will require workflow redesign. It is "theory of constraints" all over again (remember Eli Goldratt's "The Goal"?), we have to understand that real business benefits come only from working across silos, from analyzing flow bottlenecks, from understanding that optimizing non-bottlenecks creates waste. As Ng noted, letting a thousand (AI) flowers bloom has failed to lead to significant payoffs. Obviously! We in the silicon industry are no strangers to this. Workflows have been stagnant for decades. Technologies for high level specification, high level synthesis, and formal verification are not yet integrated into the full silicon development flow. They are often done piece-meal, and enthusiasts' pilots are dropped at the hint of a failure or delay. People say they want formal verification in their flow, but want it yesterday and do not have the time to invest in training engineers for its rigorous usage. Ashish Darbari, responding to a question on "if formal verification is so powerful then how come it is not widely used", noted that the problem is in the name - the actual task to be done is formal analysis while the bug detection is only a corollary, but the industry puts the cart before the horse. Based on my experience with industrial practice of high level specifications and formal verification, I tend to agree with his succinct comment. Jan Rabaey, in a DAC keynote speech, said that someday he expects to see a comic book series entitled "adventures in formal verification". AI deployment faces the same problem, if not actively worked on by business leaders. To be fair, deployments of all disruptive technologies face this mental barrier, we are so used to thinking in terms of the old flow (using the old technology) that thinking of a new flow literally means us having to begin again from first principles - few leaders have the courage and the time to do that. Goldratt mentioned that he himself had faced the same mindset problem when deploying the early ERP solutions and was surprised why he wasn't getting the expected large benefits, and he said it took him many years before he could identify the core problem and define solutions for it, which he named (rather academically) as "theory of constraints". AI leaders would do well to rework the conventional workflows in a holistic manner in order to get the real benefits from this new technology. As Henry Ford had said, "If you always do what you've always done, you'll always get what you've always got".

21

I've just added the latest models to this small tool, which helps in understanding the context window sizes of various LLM models. https://lnkd.in/eiyMFNWm

2

The main reason why transformer-based sequence models have become ubiquitous is that they can be trained more efficiently than RNNs. But there is a flipside: On many simple tasks, they require significantly more training data than RNNs. A likely reason for this difference is that RNNs have a very natural, human-like inductive bias that transformers lack. They must learn to compress the past to predict the future. https://lnkd.in/gTGW66sa

91

8 Comments

My easy hypothesis on LLM inference variability was always around floating-point precision — but this excellent article from Thinking Machines Lab helped me see why that’s only part of the story (https://lnkd.in/gg4MZste). The real cause? Batch size variability affecting GPU kernel operations, causing nondeterministic outputs depending on system load. Remarkably, fully deterministic inference is achievable, though at about 2X slower performance. For those in regulated industries or managing multi-agent operations where precision is critical, balancing how much determinism you need versus your budget may become a vital new consideration—and a potential new service offering. Would love to hear how others approach this tradeoff. #reproducibility #ai #llm

10

A new study from Anthropic explores whether large language models can actually notice what’s happening inside their own reasoning. By injecting known concepts directly into a model’s internal activations and observing how it responds, the researchers found that advanced models like Claude Opus 4 and 4.1 sometimes recognize and describe those injected “thoughts.” They can even tell the difference between their own internal representations and external text, and in some cases, identify when a previous output wasn’t truly theirs. This doesn’t mean AI systems are conscious or self-aware, but it suggests a small step toward something like introspective awareness — the ability to monitor and report on internal states. The finding could make AI behavior more interpretable and transparent, though it also raises new questions about control and reliability. In short, today’s most capable language models might be beginning to “notice themselves,” even if only faintly.

3