Short-form insights you can read in under 5 minutes.
Using a purpose trained models with fewer parameters built of curated training content can out perform mega models for focused tasks. They would definitely be more efficient from a compute perspective. But if you include the training costs and the number of models, I don’t know which approach creates the larger carbon footprint.
Here’s the flow of prompts I used today to convert a code file from one language to another and provide test coverage. I list out the steps illustrated by my actual prompts.
This sounds like a breakthrough advancement in Quantum Computing…
“With (Microsoft’s) qubit-virtualization system, (they) were able to create four highly reliable logical qubits from only 30 physical qubits of the available 32 on Quantinuum’s machine.
Three fundamental criteria to advance from noisy, intermediate-scale quantum computing to reliable quantum computing are:
(They) have demonstrated, for the first time on record, that all three of the above criteria have been met.”
https://lnkd.in/gWyRcenS (via pivot5)
AI is a concept manifested as enabling tools. We can build the tools more or less responsibly. We can use tools to create more or less harm - less or more benefit - for few or many people. It is humans who will be the heroes, heels, and supporting players in this story.
Amazon “Just Walk Out” is an unsettling experience. Turns out it’s not advanced AI. More surveillance and hidden labor. It relied on more than 1,000 people in India watching cctv and labeling videos (theinformation.com) to ensure accurate checkouts.
As software developers, we are paid to apply knowledge practically. The domains we master grow more complex. Our days are overwhelmed with tasks. Use AI automation to reduce time spent on tasks and allocate more for learning. We can produce more value, build our craft, and remain employable.
Working intensively with AI code generation tools, hasn’t changed my basic vision of the relationship between the tool maker, the tool wielder, and the tool. The people need the skill and professionalism to craft good software before their tools will help them create it.
The new OMB policy:
Agencies need to disclose safeguards to ensure AI systems cannot endanger citizens’ safety or infringe rights.
Publish system code (foreseeable loopholes here) and performance metrics.
Create governance boards and an accountable executive with oversight on AI use.
Well done.
One challenge I struggle with is how to build cloud based solutions without generating unnecessary computes in the process. The work requires significant development by configuration. The API’s are abstract. The newer services like the AI suites are sparsely documented and the concrete examples obsolete quickly. So, working with them entails executing against the real resources at almost every step. We build with smaller models (if available), smaller sample sets, etc. but we’re still building without the ability to unit test.
This is going to work… converting a codebase with Generative AI assistance (claude 3/copilot)
Progress over 2 weeks (2-3 hrs/day): • 15 elixir files (33%) converted to typescript/node • 96% unit test statement coverage. • moved react front end and recreated webpack build process. • converted database migrations • 1 of 3 controllers, oauth user authentication with email notification working.
I am an experienced developer rusty in node and unfamiliar with elixir, typescript, and react.
Worth an offhand comment just how helpful coding with an AI assistant facilitates reasonable context switching. I can sustainably balance work on two projects or within two roles in a project in a day. Having a context thread that literally describes what I was working on helps.