Pedro Silva

I work as fractional CTO through 1d.works, helping early-stage companies move fast on technical strategy and AI integration, from architecture through implementation and deployment.

Working notes

On LLMs and Human Intelligence (2)
December 27, 2024
Thinking more about this.
LLMs interpolate, they don't compose. They can't combine ideas into genuinely new structures because they have no mechanism for compositional reasoning, just proximity matching. Real composition might require consequences that happen to a continuous self, whereas LLMs rebuild from scratch each token. Humans can compose from sparse data.
Show a kid one lever, they'll start using sticks to flip things. LLMs need many examples to build a statistical manifold around "lever-ness." It's about composing from principles extracted from single experiences vs interpolating across massive experience-spaces. Just different kinds of cognition, not different points on the same scale.
AI tasks
December 18, 2024
Your periodic reminder: the big story isn't AI automating our jobs, it's AI revealing entirely new territories of thought, creativity, and opportunity
On LLMs and Human Intelligence
December 8, 2024
Someone recently tweeted about being in awe of LLMs while sensing they're missing something "ineffable yet essential" about general intelligence. I think maybe the fact that it's ineffable is a hint. That in fact there are no such things as general intelligences, and it's a whole slew of context-dependent, dynamic, embodied processes that emerge from continuous interaction between specific systems and their environments.
What irks us about LLMs is they're unreliable in a very inhuman way. When they fail, they violate our expectations of how minds work. We expect intelligence to come bundled with self-awareness, doubt, the ability to know what you don't know. LLMs show us intelligence decoupled from understanding, and that just doesn't feel right.
So I think it's a category error to compare human intelligence to LLMs, and asking "when will LLMs achieve human-level intelligence" is the wrong question. Maybe more interesting: what novel forms of cognition are these systems showing?
teamLab
November 28, 2023
Visited teamLab in Tokyo.
Made me think there's a whole lot to be explored in wonder/transcendence/psychedelic territory - especially for people not keen on diving into LSD or DMT.
The idea is to get a taste of "ecodelic insight" (see Richard Doyle's Darwin's Pharmacy), that moment when you get that you're not just an individual but a piece in a complex ecosystem.
Like a "grokking of Gaia", where you suddenly feel part of a much larger, interconnected life system.
May require a new approach to meaning, where you need to let go of the need to interpret and control and let the experience stand for nothing but itself (Leary's "non-symbolic patterns").
Efficiency in Germany vs Japan
October 28, 2023
Both Germany and Japan care about efficiency, but their approaches are shaped by different cultural backgrounds.
Japan: collectivist, values social harmony and consensus. Technology is adopted if it enhances collective well-being. The restrictive aspects come from a desire to maintain social cohesion.
Germany: individualistic, focus on precision engineering and planning. Less emphasis on collective consensus and more on functionality and standards.
Take the Shinkansen and Autobahn as metaphors:
Shinkansen: collective harmony through rigid protocols. There's a place and protocol for everything, and individual deviations aren't encouraged.
Autobahn: individual freedom within a framework of expertise and shared rules. Go as fast as you'd like on good roads, as long as you abide by the rules.