Pete Werner @pete.penumbra.software · 19/04/2026A lot of car makers aim to profit off spares and replacement parts. 000
Pete Werner @pete.penumbra.software · 12/10/2025Am I missing something here or did they train a model to spout gibberish after a specific rare token then consider it noteworthy when it works? www.anthropic.com/research/sma...anthropic.comA small number of samples can poison LLMs of any sizeAnthropic research on data-poisoning attacks in large language models 000
Pete Werner @pete.penumbra.software · 02/10/2025Arguably RL has learnt something more general, ie what to do when encountering the plus operator, which can be applied or extrapolated to instances outside its training data. 000
Pete Werner @pete.penumbra.software · 02/10/2025Not familiar with the source that sparked this but take the context of SFT vs RL trying to learn the plus operator. SFT can conceivably rote learn every a + b = c, while RL could learn if a and b are numeric, put the sum after the = symbol. 110
Pete Werner @pete.penumbra.software · 02/10/2025The impressive thing about Gen AI is how often it actually works 000
Pete Werner @pete.penumbra.software · 02/10/2025Is it not about being able to validate a candidate response independent of any initial training data. 120
Pete Werner @pete.penumbra.software · 02/10/2025Great write up on matmuls if you’re into the gory details www.aleksagordic.com/blog/matmulaleksagordic.comInside NVIDIA GPUs: Anatomy of high performance matmul kernels - Aleksa GordićFrom GPU architecture and PTX/SASS to warp-tiling and deep asynchronous tensor core pipelines. 000
Pete Werner @pete.penumbra.software · 29/09/2025I will be visiting Atlassian in a few weeks for a panel discussion on Reinforcement Learning, come along if you’re in Sydney www.aicamp.ai/event/eventd...aicamp.aiAI Meetup (Sydney) with Atlassian - Reinforced Learning for AI ModelsJoin over half million developers learning how to use and build AI through expert-led tech talks, workshops, bootcamps and crash courses. Level up your skills, and stay ahead of the industry | AICamp 000
Pete Werner @pete.penumbra.software · 25/09/2025Feel like I don’t hear AGI as much as I did 3-6 months ago. I guess the checks have cleared. 000
Pete Werner @pete.penumbra.software · 19/09/2025No I hope you talk to someone if you think it might help and are feeling better soon either way 010
Pete Werner @pete.penumbra.software · 13/09/2025Fleshing out a proposal with ChatGPT: 5 minutes Validating the details: 4 hours 000
Pete Werner @pete.penumbra.software · 11/09/2025I block a lot of words like prominent names etc. it’s just not a conversation I can meaningfully contribute to or engage with 010
Pete Werner @pete.penumbra.software · 27/08/2025I don’t mind Gemini but they never listen to their customers. 000
Pete Werner @pete.penumbra.software · 02/08/2025If you feel old, ChatGPT just told me “you’re among the ancient ones of the web.” 000
Pete Werner @pete.penumbra.software · 23/06/2025Startup idea: Secure MCP. It’s just mcp but the logo is a padlock. 000
Pete Werner @pete.penumbra.software · 09/06/2025Hot take: Apple is second only to NVIDIA when it comes to AI. They have been doing it a long time, their own hardware and importantly mature and robust software on top of it. #wwdc 010
Pete Werner @pete.penumbra.software · 04/06/2025I aspire to the level of brazenness whoever makes the marketing charts for NVIDIA has attained 000
Pete Werner @pete.penumbra.software · 23/05/2025Remember in 2016 people were going to hail a self driving Uber instead of owning a car and driving themselves 000
Pete Werner @pete.penumbra.software · 22/05/2025An ablation study is not mathematical rigor. It’s an empirical experiment. 000
Pete Werner @pete.penumbra.software · 21/05/2025It’s gotta happen imo. Book to chapters, chapters to paragraphs, paragraphs to sentences, sentences to words, words to letters. Low frequency to high frequency. 000
Pete Werner @pete.penumbra.software · 07/05/2025Nice looking work on LLM inference arxiv.org/abs/2505.01658arxiv.orgA Survey on Inference Engines for Large Language Models: Perspectives on Optimization and EfficiencyLarge language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the i... 000
Pete Werner @pete.penumbra.software · 07/05/2025If you can’t think of any good use cases for LLMs maybe you’re just boring and uncreative 000
Pete Werner @pete.penumbra.software · 23/04/2025If you are in Sydney this April 30 I will be giving a talk on scaling up AI services at AI Camp in Sydney. How we built and scaled the core AI services that drove our product to over 10 million users. Be sure to come along if it sounds of interest. www.aicamp.ai/event/eventd...aicamp.aiAI Meetup (Sydney): GenAI, LLMs and AgentJoin over half million developers learning how to use and build AI through expert-led tech talks, workshops, bootcamps and crash courses. Level up your skills, and stay ahead of the industry | AICamp 000
Pete Werner @pete.penumbra.software · 23/04/2025Open source is fine but it’s not possible to compete against someone like Google who provide production services at a loss. Unless you have funding and can do the same. Which is still putting control in the hands of the few that can run at a loss for extended periods of time. 010
Pete Werner @pete.penumbra.software · 21/04/2025Fantastic run through of the core pointy end of flow matching youtu.be/7cMzfkWFWhIyoutu.beFlow Matching | Explanation + PyTorch ImplementationYouTube video by Outlier 000