Ddpo
βOur core technical contribution is the \textbf{DDPO} algorithm,Diversity Driven Policy Optimization, a multi-turn GRPO-based approach designed to preserve dialogue diversity while holistically optimizβ¦β
What the smart people are recommending. 8874 books, tools, and products endorsed by the thinkers absorb.md tracks. Ranked by how many times each has been recommended across compiled podcasts, papers, posts, and tweets.
βOur core technical contribution is the \textbf{DDPO} algorithm,Diversity Driven Policy Optimization, a multi-turn GRPO-based approach designed to preserve dialogue diversity while holistically optimizβ¦β
βThis repository contains code for the following Keras models: - CRNN for music taggingβ
βyou can use flux or mlj.jl both of those are i think viableβ
βWe build upon this promising foundation and extend the method to work as an uncertainty estimation technique for already-trained artificial neural networks in the domain of regression. Our experimentsβ¦β
βTo learn more how this works, check out the Tailwind documentation [here](https://tailwindcss.com/docs).β
βBased on this observation, we propose QuantClaw, a plug-and-play precision routing plugin that dynamically assigns precision according to task characteristics.β
βWe validated our method, QDTraj, by generating diverse trajectories in simulation and deploying them in the real world. QDTraj generates at least 5 times more diverse trajectories for both hinge and sβ¦β
βOverall, H-optimus-1 achieves the strongest survival prediction performance.β
βNotably, the compact distilled model H0-mini slightly outperforms its larger teacher model H-optimus-0, despite using fewer than 8% of the parameters and enabling significantly faster feature extractiβ¦β
βThe platform is publicly available at Energy-Arena.org.β
βHDET is implemented as a drop-in replacement for PyTorch's OneCycleLR scheduler, requiring no changes to model architecture, optimizer, or data pipeline.β
βThe system uses instruction-tuned qwen3-embedding-0.6B embeddings...Qwen3 embeddings with 300-token chunk size achieved 94.6% accuracy on a clinical question-answering benchmark.β
βIts closure leaves behind non-updatable benchmarks, irreproducible results, and ultimately a field at risk of perpetuating these issues by turning to closed-source LLMs.β
βOxyGent is publicly available at https://oxygent.jd.com/β
βThe resulting corpus comprises Β± 35 billion tokens across the medical domain in about 100 million documents, freely available on Hugging Face.β
βthe most widely adopted multilingual base (Chatterbox, 23 languages) does not even tokenise Telugu or Tamilβ
βwe have a gpt-3 fine-tuning api these days um you know we'll be rolling that out for codexβ
βthe best open-source bases (Chatterbox, Indic Parler-TTS, IndicF5) trail them on measured phonological dimensionsβ
βFor intra-sentential code-mix we add a third branch (IndicF5 + native-script transliteration) that drops code-mix LLM-WER from 0.80-0.85 to 0.14-0.27β
βWe release R6 LoRA weights (Apache-2.0), inference code and router (MIT), and a Gradio demo.β
βSo there's one approach that I was very excited about and that I thought was it's very cool and I really like it's it's called dreamcoder by Dr. Kevin Ellis and and folks Um, so check it out if you ifβ¦β
βIn this paper, we introduce AegirJAX, a fully differentiable hydrodynamic solver based on the depth-integrated, non-hydro hydrostatic shallow-water equations.β
βBasecamp is what I've always relied on to help keep projects on track on schedule and on budget it takes a straightforward approach to project management it streamlines workflow management and definitβ¦β
βno check out uh verse AI to get started building your own AI apps and nextjs.org to check out our frameworkβ
βIf you find the work useful in your research, please cite the DeepTraffic paperβ
βTo get started right away, this repository provides a code snippet to insert into the code box on the DeepTraffic siteβ
βSee Documentation page for more details and hints and how to submit to the competition.β
βWe introduce Minimum Specification Perturbation (MSP), the smallest number of changes.β
βwe started out with i we built a software project called open ai gym which a lot of people use in reinforcement learningβ
βOur results establish Mollifier Layers as an efficient and scalable tool for physics-constrained learning.β
βRails production setup via SQLite3 made durable by https://litestream.io/β
βPlease feel free to put your questions in the slido throughout and we will answer those as we get to themβ
βI wrote about the experience of building a custom GPT (in my case, a copy editor) and how OpenAI is paving a path toward powerful AI agents, for better and for worseβ
βAnd a lot of people kind of -- You know, that was sort of the inspiration, you know, behind Torch, the first Torch, which had the Lua language as the front-end for Caffe2, or for Caffe1, actually, andβ¦β
βwe use streamlit streamless awesome it was really easy to get started - who was in Cabin ce128 we can ask a question we can press submit it should hopefully run in the background we'll get back an ansβ¦β
βHere we introduce PluRel, a framework to synthesize multi-tabular relational databases from scratch.β
βThis model and data set combination enables exceptional speed and accuracy for modeling the world at the atomic scale, accelerating the discovery of new molecules and materials.β
βmy brother John co-founder uh also at Stripe, he started a podcast uh recently um because he's the sort of the funny and charming and uh charismatic one. Uh and so so sorry about bad news for you guysβ
βWe prove that SOAP, a recently proposed quasi-Newton method, efficiently approximates the Hessian preconditioner, enabling breakthrough performance in PINNsβ
βTo guide the selection of measurement methods designed for this observable estimation problem, we propose a benchmark called CSHOREBench (Common States and Hamiltonians for ObseRvable Estimation Benchβ¦β
βTo address these limitations, we introduce Obj-Disco, a framework that automatically decomposes an alignment reward signal into a sparse, weighted combination of human-interpretable natural language oβ¦β
βThere's a system that Google's producing called MedPalm2, which has learned to do diagnoses. And it's already, I think it's better than an average doctor now.β
βCahier also raised 170 million round or a couple of rounds this year so uh so hi Nick hey how's it going very well H where does this podcast find you today what country what city what part of the worlβ¦β
βmost people know the Transformers library, but there's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cppβ
βI would use a tool called FrontPageβ
βthe reason that gitlab [&Β __Β ] Hashi you know a lot of these companies are successful it's not just because they did only open sourceβ
βthere's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cpp. So it's a constellation of tools for AI builders to build with open modβ¦β
βthere's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cpp. So it's a constellation of tools for AI builders to build with open modβ¦β
βthere's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cpp. So it's a constellation of tools for AI builders to build with open modβ¦β
βokay occasional reminder that you can support the mindscape podcast on patreon go to patreon.com shmc Caroll and kick in a buck or two for every episode of mindscape in return you get adree versions oβ¦β
βYou can find him on LinkedIn at Forin scobleizer I'll spell it for you s c o b l e i z r you can also find him at x.com sloyer we'll put that in the show notes but Robert welcome thank you for having β¦β
βthere's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cpp. So it's a constellation of tools for AI builders to build with open modβ¦β
βthere's a whole ecosystem around it from Diffusers, Sentence Transformers, TRL, PEFT, even LLaVA, and since last week Llama.cpp. So it's a constellation of tools for AI builders to build with open modβ¦β
βhey I want to tell you about a new site that I built called writing examples we take writers like Steinbeck Orwell Seinfeld and break down what makes their writing so good if that sounds like it's kinβ¦β
βWe've had this historically with with Dolly 2β
βDolly 3 really if if folks have tried it it it takes things to the next levelβ
βMeta is the right partner for open-source AI development. Not just with their llama models they have a gamut of other infrastructure llama stack exeutor torch and many libraries.β
βLet's use gbtv U or gbt 4 with with image inputs to essentially create a nice human readable understandable description of that imageβ
βGmail smart compost suggested what to reply to the emailβ
βat some point actually we we had a great success in robotics with you know controlling a a robot hand um super cool video by the way if you if you go look that upβ
βour goal with the system called Fox was you know fairly simple right where we wanted to essentially reorganize the early part of the page load to overlap the processing of static HTML content with theβ¦β
βto handle this our solution which we call the lamura basically just tries to learn how best to adapt these policies at runtime by jointly considering aspects of the page the network and compute resourβ¦β
βour solution which we called pensive basically tries to automatically learn the right way to balance and jointly analyze those cross stack elements uh to make performant bit rate decisionsβ
βto address this our solution which we call dashlet uh essentially draws on insights and constraints from how short video applications actually workβ
βto tackle this challenge the main Insight behind our system called reducto uh was that low-level video features could help us hereβ
βI won't go into the details of the the geml retraining methodology of course but at a high level it follows these principles and it does this greedy merging process where it really prioritizes these hβ¦β
βfor the AI teammates launched, we have chosen Anthropics Opus 3.6 model. Uh, and that's what we're launching with right now. That's how it's powered. uh it uh did the best in our testing and analysisβ
βnews started spreading that there were companies like Jasper and copy. a that started making more money than even open AI at the timeβ
βi'm excited to tell you today about circ 1.0 and our quantum virtual machine these are two tools that we are launching to help you develop quantum programming skills and further your researchβ
βGenerated with [Pelican](https://github.com/getpelican/pelican), hosted on Github pages.β
βthe more recent one called Lana which is the code is open source the model you can get it on request if you are using it for research purpose and it's the same level of performance as things like jgbtβ¦β
βalpaca is a model which basically is a fun tune version of llama that was built by people at Stanford for answering questions and things like that instruction so they they're pretty amazingβ
βConnect the dots: Build with built-in and custom MCPs in Studioβ
βWe introduce MuJoCo Playground, a fully open-source framework for robot learning built with MJX...the entire framework is freely available at playground.mujoco.orgβ
βq sharp from microsoft they all have um they all express you know gates in inq language in iq termsβ
βthis is another method to do this kind of self-supervised running this is which is closer to this Java architecture called ijapa so this is for learning features for images without having to do data aβ¦β
βOmnik uses generative AI to enable you to launch hundreds of thousands of ad iterations that actually work customized across all platforms with a click of a button I believe in omnik so much that I inβ¦β
βwe present a systematic taxonomy of jailbreak attacks and defenses and introduce Security Cube, a unified, multi-dimensional framework for comprehensive evaluation of these techniques.β
βSome more vibe-coding fun - every math major's favorite party trick: the wobbly table theorem as an interactive 3D visualization. https://timvieira.github.io/table-theorem/β
βI generate the embeddings locally from PDF or markdown sources using an embedded model @nomic_ai, which supports large docsβ
βif you go to sdk. verel you're going to see our documentation for basically what is the easiest way to add AI to an existing product or a new productβ
βUm the uh one of the goals of open AI of course is to create a general purpose AI and really looking at the benefits to society about how we can use these tools uh to make life better for for all of uβ¦β
βthey're called Opus Sonet and Hau um they're different tradeoffs between power and intelligence and uh uh you know speed and low cost while still being intelligentβ
βthey're called Opus Sonet and Hau um they're different tradeoffs between power and intelligence and uh uh you know speed and low cost while still being intelligentβ
βSomething called MAE, developed by some of my colleagues at FAIR, masked autoencoder.β
βamong them three shows noticeable performance including DenseNet201, MobileNetV2, InceptionV3 on the teaLeafBD dataset.β
βNew from me: Etsy secretly banned the phrase βfrom the river to the seaβ on sellersβ wares, and some employees arenβt happy. A story about content moderation in wartime: https://www.platformer.news/β¦β
βif the augur and gnosis and all of these other ethereum and blockchain based prediction markets end up launching and it continues to be the case that lots of people don't end up using them then you knβ¦β
βthen they close their eyes take something like mnist or data set that classical people Benchmark their models with and just say this is like my hunter model and how it performsβ
βeveryone tests their models on mnist or Iris all these data sets that you know 20 years ago people have done in classical machine learningβ
βI think there's a a new messaging client I'm I'm really fascinated by called uh you can pronounce it as Simplex or simple x but it's another such decentralized that's what that's what people think of β¦β
βI read uh all Cypher punks a lot on Usenet and this is where you know a lot of that energy was happening um airplane CA CIA they're already coming you think it was something you said was it a black heβ¦β
βI have a competitor to Microsoft Word coming to market right and instead of having to build I don't know if you remember the infamous bad design or I don't know if it was it would end up having to desβ¦β
βsome of his research led to chatbots like Google's Bard which we met last spring confounding absolutely confoundingβ
βTo prove it Hinton showed us a test he devised for chat gp4 the chatbot from a company called open AIβ
βI'm actually very impressed with adobe's incremental edition of AI like Firefly and Tool or like you select and whatever it seems high quality well consideredβ
βWhen Kubernetes was coming out there were a lot of alternatives some proprietary some open source what ended up winning was what every company in the world decided was this standard on top of which thβ¦β
βI invested in this company called Scale AIβ
βevery company I would advise a result of that would start I used to set up a chat server in IRCβ
βAnd one of them created Fedi, Fedimint, which I hope that some of you are aware, if not, you should explore it.β