And we’re back!
Picking up where we left off, lots has happened. For this letter I will focus on some larger developments we’ve seen in AI space the last few months. For even if my typing fingers may rest for a bit, AI certainly does not.
On to the news-
AI & Nuclear energy, a match made in apocalyptic sci-fi novels heaven?
In a quest for ever more energy with the label ‘Clean’, Big Tech is interested in financing the construction and re-opening of (small/modular) nuclear reactors to power their increasingly power-hungry data centers and AI models. Currently, interested parties count amongst themselves Google, Microsoft, and Amazon, though depending on success you can be sure others will follow as clean energy goals worldwide are increasingly under threat.
AI’s all the way down
Ever more agency is given to AI models and chatbots (e.g., Claude now gaining control of your cursor), and they’re increasingly intertwined in opaque interaction loops with unclear secondary effects. From AI-supported scientific articles being reviewed by AI-supported peer reviewers (e.g, for triage), to AI-generated emails being responded to by.. AI tools that generate auto replies.
Times this by 100 for a glimpse of the future and then start to wonder what we’re doing.
Scale is not just the answer?
Yep, diminishing returns seem to finally kick in for at least text-based AI models regarding the amount of data they’re trained on (also because we’re reaching the cap on global amounts of text-data to use). Video and image models seem to still have enough data-runway ahead of them. The ceiling seems high enough I think to still get value from these models though, especially once we start optimising what’s already here, grounding the models in the physical world, and connecting various model-architectures together.
Ah, and what about that Return on Investment (ROI) for training these models? Wonder what that will entail.
Glimpse of the past and future; maybe, perhaps, mayhaps.
Development are going fast, here some thought of mine on what happened and where it’s going. First the aim was great single-modality AI models (text chatbots, image/sound generation, etc.), then we moved multi-modal (combining various modalities such as text AND image), now we’re going agentic (giving AI models levels of agency to decide and ACT on their own volition), and likely after that we’re going embodied (embedding the new multi-modal agentic AI systems into physical entities such as robots or accessories) also to help the models become grounded into the physical realm.
Each step does not invalidate the previous step, but it does build on it and gets strengthened by it.
Once the models are all around, embedded in everyday systems, tying into Internet of Things (IoT) and edge devices like your smartwatch and phone, there will be a wealth of high quality - mixed (non-human) modality - data coming in from every side, leading to a refinement stage of existing systems, and from there who knows. Maybe the Zucc’s AR/VR revolution is then finally done cooking in the oven to be ready for a new interface revolution (I am actually quite optimistic about the tech’s evolution and added value, but the metaverse is a load of nonsense I think).
The future is covered by increasingly dense mist, whilst it’s also coming at us faster than ever. I hope to do my part to help who I can make some sense of it all. Let me know if you have a good/better way for me to contribute outside of this newsletter, always open for input.
Cheers,
Fabian/Ferdinand