Writing for Consumption by AI not Humans
The Batch from @AndrewYNg and @DeepLearningAI is one of the better AI newsletters out there.
In this issue, he talks about how “people are posting text online that’s intended for direct consumption not by humans, but by LLMs (large language models).”
I find myself doing this all the time with our internal documentation. People aren’t going to read it when I write it (or probably ever). But when they need the answer, they’re going to ask our AI RAG librarian.
So, I’m writing the content for the AI and not for the human. The AI will then write for the human.
The AI needs to be able to find the right answer and package it properly for the human.
When I create new content, I write it with the AI in mind. When the AI gives the wrong answer but the content is there, I rewrite the content and test it until it gives the right answer.
Writing for the AI is different from writing for a human. It needs clarity, simplicity, and structure.
If you give the AI what it needs, then its outputs are more likely to be what you need.
https://t.co/krLYKP49QD
#AI #LLM #RAG #artificialintelligence
So much good content always, but especially recently, from @a16z on AI, crypto, manufacturing, and operations.
1. On AI and Operations: This is exactly what we're seeing in the professional services industry. So far, we've automated inputting unstructured emails and contracts and outputting structured data for our profile, contract, bookkeeping, and other databases, saving us many hours of boring work and avoiding errors.
"As AI turns labor into software, the opportunity to productize external professional services (e.g., in legal or accounting) has become a hot topic. However, we believe there is also substantial opportunity in productizing internal work within organizations. These responsibilities often fall under the umbrella term of “operations” and can range from full-time data entry and front desk roles, to routine operational tasks embedded in every other role. This work generates fewer media headlines, but it is the internal stitching that holds companies together."
https://t.co/dz2rOw0d5p
2. On Manufacturing and Operations: Interesting podcast on "Rebuilding America's Industrial Backbone", including lots of good tips on how to build and scale an efficient organization even outside of the manufacturing space.
https://t.co/UyTLYAuXzG
3. On Crypto: Several good posts. It will be exciting to see what happens with the potential change in regulatory approach.
https://t.co/81exPlGZca
#AI #automation #operations #efficiency #crypto #blockchain
Writing is thinking.
We launched @meetgranola because we don’t want meeting bots to think for us. Turns out, a lot of people felt the same way.
Excited to announce our $20 million Series A led by @nabeel at @sparkcapital, along with @mignano, @natfriedman, @danielgross.
Think better with https://t.co/DG71Z5mEPZ
An early version from @AnthropicAI of AI using your computer to do your work for you, based on your prompt. You tell it what to do and it does it.
https://t.co/aDu6IZeO6N
Chat interfaces are fun and easy to use, but to get the most value from AI, you need to build a system with various subsystems that use the right tool for the task, including using different AI models and fine-tunings, RAG, traditional code, and more.
We've seen this when building internal tools at NU. Before we employ AI, we're creating, parsing, and editing data, and using traditional code and approaches, to narrow and simplify the task for AI to help the AI succeed.
Great podcast from Latent Space (@latentspacepod) with Mike Conover (@vagabondjack) of Brightwave (@brightwaveio) on how AI is changing financial research and how to build AI products generally.
https://t.co/oAs8EilECS
Really interesting podcast from Practical AI (@PracticalAIFM) on building Perplexity (@perplexity_ai), an AI search and answer engine, with Co-Founder & CTO Denis Yarats (@denisyarats).
https://t.co/FAgcp6VOuM
Writing is thinking.
If a bot is writing notes for you, it’s thinking for you. Instead, Granola starts with your notes, and makes them better.
Introducing @meetgranola , the first AI notepad for back-to-back meetings. Don’t stop writing, don’t stop thinking.
Download at https://t.co/9tGdvoQzN9
Are most new AI products just a front-end website with a top secret, proprietary AI prompt behind it? It sends the prompt and my data to an AI API, receives the response, and then presents a pretty, formatted response back to me?
Will every prompt become a company the same way that every unix command became a company?
#AI #unix #product @OpenAI
During the creation process, your website says: "We do NOT refund anyone who breaks these terms." And: "We do not refund orders that do not meet the requirements." So people probably assume that you're going to be difficult about refunding and jump directly to a chargeback. If you really want to lower the chargeback rate, you may have to make it clearer in the process, which will also probably cause more refunds, but still might be the best practice.