I find this new trend quite amusing. Not the lawsuits, but the influencer one. Few years ago I did a #NowAboutThis video called "Influencers are just sales people" . In short, it was about the new wave of sales folks, the influencers. And if you think about it, if you copy a window or a product from some other brand, or even a campaign, then you might see yourself in trouble. Why can't this be the same for sales people - I mean, influencers? I understand there's a fine gap between originality and trend adoption, in particular with aesthetics and styles that sooner or later everyone adopts, because it sells better. I've been lucky enough to not have the #NowAboutThis copied or even taking inspiration from others :) The difference between industry norms and individual rights is going to be interesting, in particular when brands are now more focused on buying those rights than rather create content that inspires others.

Remember cyber cafes? I do! In Portugal, they were one of the few trends we adopted quickly in the ’90s. Cities like Lisbon had a handful where you’d pay by the hour, enjoy a coffee, and - here’s the shocker - actually meet people. I even know couples who first met at these cafes! It’s fascinating to look back at how they fostered knowledge sharing, community building, and even rivalries between spots. I vaguely remember Cyberia from a trip to London; I dropped by once and loved the vibe. Now, we call them co-working spaces, but sadly, the magic of spontaneous human connection seems to be missing.

Back in June I posted a comparison between Arima vs Transformers for a Time Series LLM and this Google research is quite interesting. I like how their visual-first approach simplifies model reasoning and aligns with human intuition, avoiding additional training demands, but I think that Time-Series LLMs offer scalability and flexibility, capable of fusing structured and unstructured information for better decision-making. Their approach depends on effective data visualisation tools, which may not always be accessible or scalable for vast time-series datasets, whilst the Time-Series LLMs face issues like data sufficiency and risks of hallucinations, requiring careful grounding in real-world scenarios. But bottom line is good to see we're moving towards more context-aware and flexible time-series analysis methods.

I know that network topologies aren’t exactly dinner table conversation, right? But this caught my attention, and it feels like a bridge between old-school efficiency and modern AI demands. The way it blends torus simplicity with the flexibility of switched networks is impressive, especially for managing sparse models like MoEs. The promise of high bandwidth at lower costs is great, but I wonder how scalable it really is as models and data keep growing. Also, integrating it into existing HPC systems might be trickier than simulations suggest. It’s an exciting step, but the real test will be whether it can evolve as fast as AI workloads do.