This week’s we focused on AI capabilities—not as an end in themselves, but as a way of grounding how we later think about risk. It started quite concretely, with capabilities like language models, image generation, and then this idea of multi-modality. At first, I thought multi-modal systems were simply about translating between formats—image to text, text to image. But what began to stand out is that this is only a small part of it. A system really becomes multi-modal when it can take in many kinds of inputs—text, images, audio—and produce multiple kinds of outputs, all while keeping them in a shared context. Most of the tools I already use fall into this category, which makes the capability feel both ordinary and quietly powerful.
State of the art AI
Interestingly, what made me pause was the idea of multiple-image visual dialogue. The example of comparing images—holding them both in context and reasoning across them—felt like a qualitative shift in AI capability. It’s not just recognising what’s in an image, but relating images to each other, to a human prompt, and to a generated output. When I read that similar techniques are being used to compare medical scans over time, it clicked for me that these capabilities aren’t just creative or aesthetic. They have implications that spill directly into domains where stakes are high, even if the same underlying capability looks playful elsewhere.
This theme of generality showed up again with models like Gato. A single network, with the same set of parameters, performing radically different tasks—playing games, chatting, manipulating objects with a robot arm. What struck me wasn’t just the technical achievement, but the direction it points toward. If one fixed set of parameters can handle such varied tasks, then the boundary between “this model does X” and “that model does Y” starts to dissolve. It feels like a step toward systems that are less about task-specific tuning and more about broad competence.
Robotics added another layer to this. The distribution of where robots are actually being deployed—transportation, logistics, hospitality—was not surprising, but the decline in medical and healthcare applications caught my attention. I don’t know why that is, and the material doesn’t really explain it, but it made me curious about where capability meets constraint in the real world. The discussion of models like PaLM-E reinforced how much progress has come from integrating vision, language, and embodied action into a single system. At the same time, there was an almost deflating honesty in acknowledging how slow and impractical many of these robots still are. The promise feels real, but not immediate.
Some sections felt more convincing to me than others. The discussion on game-playing capabilities, for instance, initially struck me as somewhat superfluous. Yes, AI can play Go or chess or Atari games without being taught the rules—but why does that matter? It only began to make sense to me when I connected it to strategic reasoning and negotiation. If systems can learn strategy, deception, and planning in games, it’s not a stretch to imagine those capabilities migrating into domains like defence or policy. That’s when the “toy problem” framing started to feel inadequate.
Foundation models
The discussion on AI capability landscape was followed by an introduction to the concept of foundation models versus models designed for specific tasks. Foundation models were where my thinking shifted more explicitly towards safety. The mechanics of pre-training, self-supervised learning, and fine-tuning are familiar to me, but reading them again in this context made something clearer: when a model is trained broadly (foundation model) and then adapted narrowly (model designed for specific task), we don’t get to neatly isolate what it has learned. Biased learned during the foundation model training stage could get carried forward into different models being designed for diverse and specific tasks. Capabilities—and risks—come bundled when a model becomes a general foundation; any bias or misalignment now has many more places to surface.
Intelligence
This is also where the discussion of AI intelligence started to matter to me. The material argues that defining intelligence in AI systems isn’t just philosophical—it’s operational. What you deem as intelligence in one domain of work like scientific research could be very different from what would be deemed as intelligence in film production. Therefore, it is difficult to have one concrete definition of intelligence. If you can’t define it, you can’t measure it, and if you can’t measure it, you can’t track risk.
This is why I found myself persuaded by the capabilities-based approach to tracking AI progress, not because it resolves deeper questions about consciousness or understanding, but because it sidesteps them. From a safety perspective, it doesn’t really matter how a system reasons if what it can do poses risk. That realisation helped me separate my interest in AI ethics from the more urgent, and perhaps more pragmatic, concerns of alignment and capability.
Scaling
We then moved on to a discussion on the concept of scaling in AI systems development. Scaling tied many of the above threads together. The “bitter lesson” was uncomfortable but difficult to ignore: progress has come less from clever human-encoded rules and more from scaling compute, data, and learning. At the same time, it’s not that humans are no longer needed—it’s that their role has shifted. Architectural innovations like transformer models made scaling possible in the first place. Still, the emphasis has clearly moved away from hand-crafted intelligence toward systems that learn broadly and at scale.
The discussion of scaling laws and hypotheses left me with more questions than answers. The idea that performance doesn’t always improve smoothly—that there can be plateaus, regressions, sharp transitions—felt especially important. It complicates any clean narrative of linear progress. And the debate between strong and weak scaling hypotheses made me aware of how much uncertainty there still is, especially around data limits. If high-quality human-generated data runs out, what does that mean for continued progress? Synthetic data might help, but it might also change what models learn in ways we don’t fully understand yet.
Forecasting
Finally, forecasting brought the stakes into focus. The point isn’t to declare when transformative AI will arrive, but to assign probabilities and timelines so we can decide when safety interventions need to happen. Five years versus fifty years isn’t just a numerical difference—it determines whether we act urgently or deliberately. Breaking bigger prediction problems (like the rate of progress of AI systems) into smaller components—rate of progress of data, compute, algorithms— and then averaging out the results to predict the bigger picture felt like a sober way to think about the future without sliding into either hype or denial.
We were also introduced to different ways of forecasting AI progress, along with the strengths and limitations of each approach. What stayed with me was the idea that, despite having increasingly sophisticated prediction and forecasting tools, some AI capabilities may be so novel—or emerge in ways we haven’t yet learned to anticipate—that forecasting alone isn’t enough. In those cases, we still end up relying on the judgment of experts and scientists to make sense of where progress might head. This connected to a brief discussion on takeoffs: the idea that AI capability growth might be gradual, sudden, or uneven across domains, which further complicates how confidently we can predict timelines or risks.
What I am left with, at the end of this week, is a growing sense that capability is the thread tying everything together. Ethics, alignment, safety, forecasting—all of them seem downstream of what systems can actually do, and how broadly those abilities generalise. I am still interested in the ethical questions, but I am beginning to see why safety work keeps circling back to capability as the starting point. It’s not a conclusion so much as a reorientation, and I am aware that my thinking here is still very much in motion.
Have a lovely day!
Trishna
*These posts are sometimes shaped with help from AI tools, but the thoughts, interpretations, and views are wholly my own.
Sahitye Jeevanam 