Cameron R. Wolfe, Ph.D.

@cwolferesearch

Research @Netflix • Writer @ Deep (Learning) Focus • PhD @optimalab1

following
followers

Image-free runs barely change SWE-bench Multimodal v2 scores in a two-model test

SWE-bench Multimodal v2's issue text and code usually contain enough information to solve tasks, the researcher says.

2 Sources, 6d ago, first seen 6d ago

FineDNA dataset released for training AI models on DNA and text

FineDNA supplied Carbon-X with 302 billion mid-training tokens, plus supervised and reinforcement-learning data for DNA tasks.

16 Sources, 6d ago, updated 6d ago

SWE-bench Multimodal v2 scores barely moved without images in one experiment

Sankalp Shubham found little change in resolved-task rates, while both models faltered on a separate visual test.

5 Sources, 19d ago, updated 19d ago

The value of understanding a hypothetical cancer cure's mechanism

A poster says almost any Stage 4 cancer patient would find human understanding of a working cure's mechanism irrelevant.

20 Sources, 29d ago, updated 29d ago

AI pretraining framed as selecting an action space and initial distribution

In a reply, a user describes pretraining as “just selecting an action space and initial distribution.”

5 Sources, 33d ago, updated 33d ago

How multi-teacher on-policy distillation combines specialist capabilities in one model

A post explains that the student model generates text, then a teacher for that domain scores it token by token.

1 Source, 34d ago, first seen 34d ago

Straightforward multi-teacher distillation may falter when teacher training pipelines differ substantially

A post discussing the Nemotron 3 Ultra tech report suggests a brief supervised fine-tuning warmup to better align the student with its teachers.

2 Sources, 39d ago, updated 39d ago

A characterization of someone described as an evals expert draws pushback

In a reply, a commenter calls a characterization of someone “crazy,” saying the person is widely regarded as an evals expert.

1 Source, 40d ago, first seen 40d ago

Is big Googz back on the menu?

A user floats the question while sharing another post.

11 Sources, 40d ago, updated 40d ago

LLM training trends in NVIDIA Nemotron technical reports

A writer preparing a blog post calls NVIDIA’s Nemotron reports especially useful for following LLM research, citing their open weights, data and code.

2 Sources, 42d ago, updated 42d ago

From policy gradients to GRPO: reinforcement learning for language models

A technical explainer describes how PPO trains a separate value model to help assess actions, while GRPO instead compares rewards across answers to the same prompt.

1 Source, 46d ago, first seen 46d ago

Arcee AI announces Series B funding at a valuation above $1 billion

Arcee says the funding will advance its Trinity models, expand work with the Department of Energy and national labs, and support a platform for building and running open models.

11 Sources, 54d ago, first seen 54d ago

TypeSafe launches Jev, an AI model aimed at decisions rather than text

Creator Diogo Almeida claims Jev is 20–200x faster and 40–400x cheaper, with free output tokens. Investor DCVC says it led TypeSafe’s $40 million seed round.

124 Sources, 45d ago, updated 45d ago

A post breaks down importance sampling in reinforcement learning

The explainer describes a training mismatch: samples can come from a policy that differs from the one being optimized, including when algorithms reuse the same samples for multiple updates.

1 Source, 59d ago, first seen 59d ago

A post explains importance sampling’s role in reinforcement learning

The explainer describes how probability ratios can help correct a mismatch between the policy that generated training examples and the policy being optimized.

1 Source, 59d ago, first seen 59d ago

Notion Page on Multi-Teacher On-Policy Distillation

Notion page outlines multi-teacher on-policy distillation for AI post-training.

1 Source, 68d ago, first seen 68d ago

Kangwook Lee Shares WHALE Paper on Model Training

Lee, CAIO at KRAFTON, links to arXiv paper on joint harness-weight training.

6 Sources, 68d ago, first seen 68d ago

4000 GB200s Arrive in Texas for Horizon TACC Cluster

UC Berkeley professor Alex Dimakis reports the delivery and plans to train open models on the machine.

5 Sources, 69d ago, first seen 69d ago

Cameron Wolfe Questions PPO Ratio Derivation in Tweet

Netflix scientist shares thoughts on common PPO explanation gaps.

1 Source, 69d ago, first seen 69d ago

Nathan Lambert Sees AI Competition Vibes Shift in Four Weeks

Nathan Lambert observes rapid change from Anthropic lead to high competition in four weeks.

14 Sources, 89d ago, first seen 89d ago

Featured in

Image-free runs barely change SWE-bench Multimodal v2 scores in a two-model test

FineDNA dataset released for training AI models on DNA and text

SWE-bench Multimodal v2 scores barely moved without images in one experiment

The value of understanding a hypothetical cancer cure's mechanism

AI pretraining framed as selecting an action space and initial distribution

How multi-teacher on-policy distillation combines specialist capabilities in one model

Straightforward multi-teacher distillation may falter when teacher training pipelines differ substantially

A characterization of someone described as an evals expert draws pushback

Is big Googz back on the menu?

LLM training trends in NVIDIA Nemotron technical reports

From policy gradients to GRPO: reinforcement learning for language models

Arcee AI announces Series B funding at a valuation above $1 billion

TypeSafe launches Jev, an AI model aimed at decisions rather than text

A post breaks down importance sampling in reinforcement learning

A post explains importance sampling’s role in reinforcement learning

Notion Page on Multi-Teacher On-Policy Distillation

Kangwook Lee Shares WHALE Paper on Model Training

4000 GB200s Arrive in Texas for Horizon TACC Cluster

Cameron Wolfe Questions PPO Ratio Derivation in Tweet

Nathan Lambert Sees AI Competition Vibes Shift in Four Weeks

Featured in

Image-free runs barely change SWE-bench Multimodal v2 scores in a two-model test

FineDNA dataset released for training AI models on DNA and text

SWE-bench Multimodal v2 scores barely moved without images in one experiment

The value of understanding a hypothetical cancer cure's mechanism

AI pretraining framed as selecting an action space and initial distribution

How multi-teacher on-policy distillation combines specialist capabilities in one model

Straightforward multi-teacher distillation may falter when teacher training pipelines differ substantially

A characterization of someone described as an evals expert draws pushback

Is big Googz back on the menu?

LLM training trends in NVIDIA Nemotron technical reports

From policy gradients to GRPO: reinforcement learning for language models

Arcee AI announces Series B funding at a valuation above $1 billion

TypeSafe launches Jev, an AI model aimed at decisions rather than text

A post breaks down importance sampling in reinforcement learning

A post explains importance sampling’s role in reinforcement learning

Notion Page on Multi-Teacher On-Policy Distillation

Kangwook Lee Shares WHALE Paper on Model Training

4000 GB200s Arrive in Texas for Horizon TACC Cluster

Cameron Wolfe Questions PPO Ratio Derivation in Tweet

Nathan Lambert Sees AI Competition Vibes Shift in Four Weeks