AE Studio

Heard you like your PCs how you like your coffee: custom-built and high performance. Our Same Day Skunkworks team brewed up an AI recommendation tool that gets you Intelligentsia pour-over at gas station prices, but for your custom-built PC: PC Builder AI. Rejoice. #TechRevolution #coffee #onabudget #DreamBigDreamRig https://hubs.ly/Q02HYMZZ0

🚀 Just took Codium for a spin, and it's like having a junior dev who never sleeps or asks for raises. Here's the lowdown: 💻 Codiumate: IDE integration that's basically Copilot's cousin. Available for VSCode and JetBrains, because choosing your IDE wasn't complicated enough already. 🔍 PR-Agent: Open-source tool for PR reviews. It's like having a code reviewer who doesn't judge you for your variable naming choices at 2 AM. 🧪 Test Generation: Automatically creates test suites, including some edge cases. 🗣 Code Explanation: Provides detailed summaries of functions. It's like having a translator for your own code after a long weekend. 📄 Documentation Generation: Creates docstrings for functions. Because we all know documentation is that thing we'll "definitely do later." 🛠 Code Enhancement: Offers improvement suggestions. It's like having a backseat coder, but one you can actually mute. ⚠ Fair warning: The test generation feature was a bit unstable in the latest version. It's going through its rebellious teenage phase. 🤖 Bonus: PR-Agent can even write poems about your PRs. Because who doesn't want their code changes immortalized in verse? Codium's not perfect, but it's a solid suite of tools that could boost your productivity. Or at least give you someone else to blame for those mysterious bugs. Anyone else tried Codium? Share your experiences below. Bonus points if you've named your AI assistant. You can read more about our team and get in touch with them here: https://hubs.ly/Q02GYGyy0 #CodingTools #AI #SoftwareDevelopment #ProductivityHacks

📝 As part of our collaboration with Professor Michael Graziano, the AE Studio research team recently published a paper that shows that adding self-modeling to artificial networks causes a significant reduction in network complexity. These ideas could be key for prosocial AI and AI alignment in the near future. ✅ When artificial networks learn to predict their internal states as an auxiliary task, they change in a fundamental way. 🔷 To better perform the self-model task, the network learns to make itself more simple, regularized and parameter-efficient and, therefore more amenable to being predictively modeled. We tested our approach in 3 classification tasks across 2 modalities.
 ✅ In all cases, adding self-modeling significantly reduced network complexity (smaller real log canonical threshold (RLCT) & narrower distribution of weights). 🔷 These results show that simply having a network learning to predict itself has strong effects on how it performs a task. ✅ The learning has a restructuring effect, reducing complexity and increasing parameter efficiency. 🔷 This self-regularization may help explain some of the benefits of self-models reported in recent ML literature, as well as the adaptive value of self-models to biological systems. ✅ Agents that self-model may restructure themselves to better predict themselves, predict others, and be predicted by others. 🔷 Imagine 2 dancers intuiting each other's moves, lions hunting in cooperative packs, or people who just "get" each other. ✅ Or, Social Cooperation 🐒 It could even be that the evolution of predictive self-models in individual animals allowed for the eventual evolution of ensembles of animals that engage in mutually-predictive, complex patterns of social cooperation. Check out the paper here: https://hubs.ly/Q02Hsg_10

👵 If you haven't noticed, our Same Day Skunkworks team has a sense of humor (they're actually hilarious and very fun people who we can put you in touch with if you have a similar idea). But just in case, they tested Rent Elderly People with individuals identifying as Elderly and they were into it too. So here is an app that lets you skip lines and get senior discounts. Mostly a joke everyone is in on, but some argued that the companionship was worth it. Have a laugh or try it out. https://hubs.ly/Q02HgF9b0

Our July Newsletter is out! We talk about how we built an app to support Eurovision Latinoamerica's live voting, a rare opportunity to book one of our rarely-available dream teams, events and talks our team was part of in June, and a rollup of Same Day Skunkworks faves.

🔬Our team put Microsoft's Florence-2 to the test, and here's our take: Florence-2 is a game-changing vision foundation model that's living up to the hype. After thorough testing, we're impressed by its versatility and performance across various computer vision tasks. Key findings from our assessment: ✅ Exceptional captioning: From basic to highly detailed image descriptions, Florence-2 delivers impressive results. 🎯 Accurate object detection: Consistently identifies and localizes objects with high precision. 🔎 Powerful segmentation: The large model excels at referring expression segmentation, though fine-tuned versions may underperform in this area. 📝 Solid OCR capabilities: Handles both printed and handwritten text, with occasional confusion between similar characters. ⚡ Performance insights: The large model (770M parameters) generally outperforms other versions across tasks. Task completion times vary, with complex tasks like segmentation taking longer (up to ~50s on M2 Pro). 💡 Pro tip: We found the base model sometimes inconsistent with multiple object detection, while the larger model showed more reliability. Florence-2 is a powerful tool for computer vision tasks, but like any AI, it has its quirks. Proper model selection and task-specific tuning are key to maximizing its potential. Have you experimented with Florence-2 or similar vision models? Tell us about it in comments. You can read more about our team and get in touch with them here: https://hubs.ly/Q02GYGyy0 #AITesting #ComputerVision #MachineLearning #TechInnovation #Microsoft

Safe AI may be crucial for mitigating existential risks as AI systems become more powerful. Our AI Safety and research team created this video to make understanding key Safe AI concepts a bit faster and easier. While there are scientific publications about this topic, they have limited reach, so our hope is that this video will be more easily consumed and shared widely. https://hubs.ly/Q02GzY-C0

Our commitment to AI Safety is all part of AE's prioritization of Human Agency. From our research to our Same Day Skunkworks tools--and of course our client projects--we build technology to improve the lives of humans and elevate businesses in a sustainable way. Reinforcement Learning from Human Feedback (RLHF) is the leading technique for fine-tuning large language models to be helpful, harmless, and honest. But RLHF is complicated and requires a distributed workforce of data labelers. To address this, the AE Studio AI Safety Research Team built a scaled-back, single-user implementation of RLHF. This implementation makes it very easy for students and researchers to run RLHF on a single laptop and gain hands-on experience in simple environments with this training technique. For more details see our write up here: https://hubs.ly/Q02G7-nQ0

Our July Newsletter is out! We talk about how we built an app to support Eurovision Latinoamerica's live voting, a rare opportunity to book one of our rarely-available dream teams, events and talks our team was part of in June, and a rollup of Same Day Skunkworks faves.

"Offer to walk a neighbor's dog" and other suggestions from Sharing Your Kindness help users spread positivity daily. A web app built hyperfast by our Same Day Skunkworks team launched on Product Hunt today (link coming in comments 👇) https://hubs.ly/Q02FC2850