Mastering social ads & decoding reasoning LLMs: the edge you need today
👋 Hi and welcome to Brighteye Bulletin - your fortnightly take on all things in Edtech, HRTech and Talent Empowerment, both from a Startup & VC perspectives.
Articles and Resources of the week
Understanding reasoning models - LLMs for complex problem-solving
In 2024, LLMs shifted toward specialization, including reasoning models tailored for multi-step tasks like advanced math, puzzles, and coding. These models complement general-purpose LLMs but are better suited for complex use cases.
What are reasoning models?
Reasoning models solve complex queries by generating intermediate steps, unlike standard LLMs that provide direct answers. They excel at advanced tasks but can be costlier and less efficient for simpler applications.
Key approaches to building reasoning models
Inference-time scaling: Techniques like chain-of-thought prompting and majority voting improve reasoning without retraining the model.
Pure reinforcement learning (RL): Demonstrated by DeepSeek-R1-Zero, reasoning emerged as a learned behavior without supervised fine-tuning.
Supervised fine-tuning (SFT) + RL: Combining instruction fine-tuning and RL, as seen in DeepSeek-R1, enhances reasoning performance.
Distillation: Smaller, efficient models like DeepSeek-R1-Distill are trained on outputs from larger models, offering strong performance at lower costs.
Lessons from DeepSeek R1
Emergent reasoning: Pure RL revealed reasoning capabilities during training.
Cost efficiency: Distilled models balance affordability and performance, showcasing potential for low-budget research.
Future innovation: Techniques like "journey learning," which includes incorrect solution paths during training, may improve model reliability and self-correction.
These models align with the growing trend of domain-specific optimizations. For Brighteye Ventures, this means we will have more personalized tools across education, workforce development, and operational productivity - key drivers of our investment thesis. Building something in these verticals? Reach out! 🎯
Master your social ads - a data-driven guide to creative testing
Are you overlooking the most critical factor in your social ads? Most marketers focus on targeting and objectives, neglecting the real game-changer: creative performance. This week, we're diving into data-driven creative testing - the ultimate method to optimize your campaigns and scale winners.
Key takeaways:
Focus on the right KPIs
Track spend, conversion volume, and CAC to identify ads to scale.
Use secondary KPIs like frequency, CPM, and CTR unique links for deeper insights into why an ad works (or doesn't).
Get advanced data insights
Dive into audience demographics, device usage, and placements to uncover trends.
Monitor hook rate (first 3 seconds of a video) and hold rate (15-second engagement) to enhance attention-grabbing content.
Optimize your actions
Low hook rate? Rethink your opening visuals or messaging.
Poor hold rate? Ensure your storytelling is compelling and concise.
Misaligned targeting? Let the data guide your audience adjustments and creative tweaks.
Ready to upgrade your ad game? Dive into Jordan’s benchmarks, actionable advice, and pro tips for smarter ad performance today. 💥
Investment news
Copenhagen-based Pandektes has secured 2.9M€ in a seed investment round led by People Ventures and Interface Capital to revolutionize legal research with AI.
Paris & SF-based Riot has secured 30M$ in a series B investment led by Left Lane Capital, with existing investors Y Combinator, Base10, and FundersClub participating once again. Originally focused on educating employees about cybersecurity risks, the company now wants to nudge employees so that they minimize their attack surface.
Freiburg-based Prior Labs has secured 9M€ in a pre-seed investment round led by Balderton Capital along with XTX Ventures, Hector Foundation (SAP), Atlantic Labs, and Galion.exe. The company builds tabular foundation models that supercharge data science teams working with spreadsheets and databases.
Dublin-based Output Sports, a company that disseminates athletic data through wearable sensors to enhance coaching, has raised $4.8M in pre-Series A funding. The round was led by Uni.Fund and includes investment from Apex Capital and Dopamine Sports, as well as existing backers Elkstone, Atlantic Bridge and Enterprise Ireland.
Jobs
Ornikar is looking for a Business Strategy Manager (Paris/Hybrid).
Zen Educate is hiring a Senior Marketing Executive (London).
Twenix is looking for an Operations Lead (Almeria/Hybrid).
UpHill is hiring a Senior Product Designer (Lisbon/Hybrid).
Riot is looking for a Chief of Staff (Paris).
Thank you for reading! If you enjoyed it, please share with friends and colleagues interested in European Edtech, HRTech and Workforce Empowerment. We also explore the AI space - stay tuned for more updates!
Onward & upward!