SAIL: Kendrick is winning, AI agents, Elections, AI & Nukes
Sensemaking, AI, and Learning (SAIL).
May 6, 2024
Agents are all the rage. Russell & Norvig describe agents as something that can perceive and act in an environment. Much of what is loosely called agents today (in the GPT frame) is prompting at best. There is limited autonomy. We don’t have true agents yet. But progress is being made, in the world of LLMs, as Andrew Ng describes here “Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output.” While work here is still building (and it’s not a Russell/Norvig agent), it gets at the ideas of asking an agent to “order me a pizza” and it does the rest, building tech solutions to route around any concerns it encounters, primarily through LLM prompting. Regardless, agents are the future of what’s happening with AI and it will have significant education implications.
What AI can Do
Like the rest of you, I have been fixated on the Drake/Kendrick feud - because it somewhat started with AI. Since the space has been rapidly escalating (Kendrick 100% dominating, btw), I decided I needed to get on top of it to retain my coolness status (with myself - no one else has ever seen me that way). After a Google search took me down many rabbit holes, I tried Perplexity and You. AI as a sensemaking agent…
We’re increasingly going to be thinking and working with AI as it becomes more integrated with daily cognitive tasks. What does it look like in practice? Here’s a good overview of promising capabilities in programming. “English is becoming a programming language.”
Propaganda, AI, elections. Buckle up. Humanity isn’t ready for this.
There are too many uneducated males out there. Not directly AI related, but we need to do education differently to reverse this decline.
GPT-4 will be the dumbest model we’ll use. I’m happily optimistic about AI. There will be an interesting race between LLMs getting better and the support infrastructure around AI (platforms, models, tooling) getting better.
An AI Diplomat “The AI spokeswoman has been created to deliver pre-pared official statements on behalf of the Ministry, and this is done using a visual avatar generated by AI.”
General AI
LLM benchmarks are starting to be less effective since the models are starting to fit the training data too well. This means that the model isn’t generalizing but rather matching to data in the training set. ScaleAI released a recent report on “data contamination”: “where data closely resembling benchmark questions leaks into the training data, instead of true reasoning ability”. In response, there are calls for new approaches so that benchmarks can be trusted to accurate reflect a model’s capabilities.
LLMs are getting smaller. We’ll start seeing them on our phones and other wearables. Meta is one example as they are rolling out Llama 3 for all their services, including Ray-ban glasses.
Killer bots should be banned: “the biggest revolution in warfare since the invention of gunpowder but feels it is far more dangerous”.
Also, only humans, not AI, should control nuclear weapons.