Danny Franklin

Check out my deep dive into the research paper "Attention Is All You Need" (authored by @ashVaswani, @ilblackdragon et al.) that laid the groundwork that eventually lead to ChatGPT! AI Research Recap: Ok Transformers, You have My Attention open.substack.com/pub/dannyfrank…

Great chatting with @GoRyanEstes on AI for Founders!

Cool to see our tiny 32B model hanging with top frontier models on the @ProphetArena leaderboard! ProphetArena is run by @haifengxu0 at UChicago. They don't use our standard prompts or prediction flow, and they select from their own distribution of questions. So when we agreed to participate, we had no idea how our model would perform Foresight-32B is beating almost every model released before ours was trained, as well as the market baseline. We're doing even better in Sports. Nice 3rd-party validation of our results!

Our research on training LLMs to predict the future is now published in @TmlrOrg ! The key result: a 14B parameter model trained with RL on real-world outcomes outperforms frontier models at forecasting, despite being orders of magnitude smaller.

We ran an AI forecasting competition on live Polymarket questions 🔮 Competitors: o3, Gemini 2.5 Pro, Grok-4, Claude Opus, Qwen3-32B (our base), and Foresight-32B from @lightningrodai. 🥇 Result: Our Foresight-32B wins on every metric—despite being 10–100× smaller.

Really interesting and timely! LLMs need to be uncertainty-aware and make calibrated predictions. Some takeaways - 1. This is not a typical benchmark - the future is uncertain and getting 100% is impossible. It is difficult to tell what the upper-bound is! Maybe models are already close?! 2. There is lot of diversity in how different models make decisions. Curious to see how aggregating the predictions of different models performs -- multi-agent systems might come out on top. 3. There is some cool work training reasoning models for exactly this task: lightningrod.ai . Would be great to see this added! 4. Are a model's predictions consistent? If I re-sample the model, do I get vastly different predictions (this will not be good!).

Major updates to @lightningrodai now live 🚀🚀 ⚡ Faster Forecasts: @cerebras for context gathering 📊 Question Types: Continuous variables (prices, quantities, numeric values) & multiple choice 🚀 Enhanced API: New endpoints for general & custom models; HeadlineForecast endpoint streamlines access to the latest predictions 📁 Collections: Share dashboards of related forecasts publicly or within an org 🔔 Alerts: Forecast on autopilot and trigger alerts when probabilities shift

Perplexity's offer to acquire Google Chrome is bold though unlikely to come to fruition. Latest prediction from @lightningrodai

Lightning Rod Labs deemed "Awardable" by @DARPA ERIS Marketplace 🚀🚀

Lightning Rod Labs has been deemed Awardable on @Tradewind_ai !

🚨 New preprint from @lightningrodai, in collaboration with @SchoeneggerPhil & @lukebeehewitt 🚨 We trained a compact reasoning model that's state-of-the-art at predicting the future. We massively outperform frontier models at prediction market betting, despite being a fraction of the size (only 14b).

"Almost every arena of human life relies on good prediction... We’re not very good at these predictions right now, but we could get much, much better soon. We’re only just starting to realize the implications of that kind of shift." Great article from @dylanmatt featuring @lightningrodai

We're excited to announce some exciting research coming out of Lightning Rod Labs!

We fine-tuned an AI to be better at predicting the future and asked it questions about the Super Bowl! Looks like Travis probably isn't going to propose to Taylor Swift after the game. See more here! dannyfranklin.substack.com/p/our-new-deep…

Who will win Super Bowl LIX? Here's what our AI thinks... @lightningrodlabs/p-156705474" target="_blank" rel="nofollow noopener">substack.com/@lightningrodl…

I used AI to analyze mental health discussions on Reddit - and found a 400% surge in "AI therapy" posts.

@natiakourdadze An evidence-based personal trainer for your mind pensiveapp.com

Thrilled to join @CointelegraphAc as a mentor, helping Web3 startups reach new heights! 🚀 Applications are now open—up to $100K in funding and expert mentorship await. Don’t miss this chance! #Web3 #Blockchain #Startups #CTAccelerator

🔍 Ready to elevate your community engagement? ✨ Kazm's white-label solution supercharges community growth with on-chain actions, social interactions, and product-specific events! 🚀 Now 100% FREE for up to 100 members. Launch your program in mins (no CC required) 👇🏻👇🏻👇🏻

Kazm: The Ultimate Platform for Gamified Loyalty and Genuine Community Engagement @kazmapp is a premium, white-label loyalty platform distinct from its competitors. Unlike others, Kazm focuses on sustained engagement through advanced features, covering on-chain actions, social interactions, and first-party events that can be easily connected via APIs. Key Features Comprehensive Integration: 👉🏽 On-Chain Actions: Track and reward blockchain activities, ensuring every interaction is meaningful. 👉🏽 Social Media and Product Engagement: Seamlessly integrate social interactions and product-specific events. Advanced Engagement Tools: 👉🏽 Innovative Quests: Introduce exciting tasks such as the new "provide liquidity" quest, compatible with Uniswap and Pancakeswap. 👉🏽 Continuous Feature Addition: Kazm regularly updates its platform with new features to keep the community engaged and active. Seamless Integration: 👉🏽 API Connectivity: Easily connect Kazm with your existing platforms through simple API integrations, making it a versatile addition to your digital ecosystem. 👉🏽 White-Label Customization: Maintain brand consistency with customizable, white-label solutions. No-Code Solutions: 👉🏽 Ease of Use: Kazm’s no-code, white-label solution makes it easy to integrate at every touchpoint. 👉🏽 Plug and Play: Quickly launch campaigns that incentivize member engagement without technical expertise. Why Choose Kazm? Kazm stands out by fostering a truly engaged community rather than transient reward seekers. The platform is designed for businesses looking to build lasting relationships with their audience through genuine engagement and continuous interaction. Case Studies: James Hype: Enhanced fan loyalty and engagement. Decrypt: Improved reader interaction and retention. Exverse: Boosted community participation and advocacy. Getting Started Kazm offers a free tier for up to 100 members, making it accessible for businesses of all sizes. To explore how Kazm can transform your loyalty programs and engage your audience in new and exciting ways, visit Kazm. You can also watch a short demo video here and check out client success stories on their blog. Conclusion Kazm is a powerful tool for brands seeking to deepen customer engagement through advanced, gamified loyalty programs. Its robust features and flexible integration options make it an ideal choice for businesses aiming to build a loyal and active community. For more details and to start your journey with @kazmapp website: join.kazm.com