LLM Links, 9/19/2024
A simulated society; Ted Gioia makes pessimistic near-term predictions; Spencer Schiff makes optimistic near-term predictions; Scott Belsky makes business near-term predictions
Project Sid: The first simulation of 1000+ truly autonomous agents collaborating in a virtual world, with an emergent economy, culture, religion, and government. https://x.com/GuangyuRobert/status/1831006762184646829
In the fields that refer to themselves as “social science,” I bet that this is likely to be an important application of LLMs.
With so many conflicting AI programs ramping up, the bots will inevitably go to war with each other. As soon as they are given agency and responsibility, battles will ensue.
Like gunslingers in a Western town, they will shoot it out—metaphorically, and possibly in real terms.
…AI pollution is everywhere. Just try finding something real amid the fake images, fake articles, fake books, fake recording artists, fake videos, fake sources, fake quotes, fake people, fake facts, fake everything.
That’s why the word artificial is embedded in the very name for AI.
By the way, this pollution is an inevitable result of the dominant AI strategy which is (always) to flood the market with artificial stuff by the thousands or millions. AI is thus the opposite of the artisan, who takes care in creating each individual object.
While optimists talk about cancer cures and AI tutors, he sees the anti-social and destabilizing applications arriving first.
Most of this drama will have played out within the next 24 months.
Have a nice day.
Nearly everyone is underestimating how big of a deal omnimodality will be… Right now, all we have access to is text and images for input, text for output. By the end of 2025 it will be text, images, audio, video for input, text, images, audio, video for output.
…next year we might have motor control output being part of the omni models as well. So the omni models will be embodied in humanoid robots.
…Many people fundamentally think of an LLM as some kind of interactive encyclopedia, and I understand why they have that misconception. But next year the models will act a lot more like reasoning engines. They won’t have a lot of world knowledge, but they will be very intelligent and creative. That’s basically the opposite of what the current models are like, so it will be a dramatic change that happens in a short period of time.
Pointer from Tyler Cowen.
AI is having a more material impact for builders than consumers. On the consumer front, despite all the glorious use cases of AI making our personal lives more magical, so far we’ve only gotten better search apps, conversational LLMs, better image editing, and incremental improvements in algorithms and voice-based interfaces. Don’t get me wrong: these are all great, but I don’t feel my life has been transformed. However, when it comes to building things, AI is proving to be nothing short of transformative. Just two examples from companies I work with as an investor: A couple weeks ago, network infrastructure startup Meter debuted Command, a product that lets network administrators use natural language to take action and generate real-time dashboards on anything related to their network. And then last week Amjad Masad, CEO/Founder of Replit, debuted an Agent product that creates functional software in minutes - doing anything from setting up a dev environment, configuring the database, installing software, and deploying to production.
substacks referenced above:
@
@
@
I hope Spencer Schiff is right, but he's predicting a lot of improvement in a year. Sora, which he predicts will be far surpassed, is not even available yet.
Belsky seems to be suggesting that AI can reduce production costs as well as transaction friction. 5% to 15% of global GDP is estimated to consist of transaction costs. If AI can actually reduce costs, might it actually reduce GDP? Do economists have criteria for discerning when GDP reduction is a net positive?