Conference/ORCwAI/QnA
Conference/ORCwAI/Pt1
For someone currently exploring opportunities in the job market (USA) for sofware engineering opportunities, how much AI fluency is needed?
- Denis, a good idea would be to read Steve Yegge's post, "the death of the stubborn developer," in which he argues that senior/junior is less important than the willingness to dive in and learn.
My question to Tim. In 2019 at EdCrunch Conference we were talking about AI biases. Today AI is everywhere. Do you see its biases influencing coding and software development in general?
- Yes, that's part of why we need to understand that we're building hybrid systems, and why "eval" is a critical part of sw development now, as Lili Jiang will be discussing later in the program.
From the group chat: I'm currently a student at UNF, and I'm wondering if I should continue studying Computer Science — especially with how fast AI is advancing and changing the industry. Is it still worth pursuing?
- Definitely, as long as you orient your curriculum towards the latest tech, and spend time on real world projects. I studied classics in college, and learned everything I know about tech outside of school. But another good option is to jump ship if you have an opportunity to enter the workforce now. You can continue learning with platforms like oreilly.com
3 best books from the Tim's bookshelf he would recommend :)
- Albert O. Hirschman, The Passions and the Interests; my wife Jennifer Pahlka's Recoding America; Chip Huyen's AI Engineering; Addy Osmani's Managing Effective Engineering Teams; Brian Christian's Algorithms to Live By; Dan Davies' The Unaccountability Machine
I appreciate Gergely's book recommendation (A Philosophy of Software Design). What other books do you recommend, technical or high-level/philosophical?
- One of my touchstones is Kernighan and Pike's The Unix Programming Environment. Very out of date, but describes the fundamental architecture of participation that was more important than licenses in making open source and open systems work. AI so far is totally lacking in the architecture of participation, though the various agent protocols are trying to figure this out. But it's one of the big missing pieces.
Not a question but rather than an idea. To mitigate some basic skill erosion we could introduce a notion of vibing credits. As with any credits, you can earn and spend them. Basically, some kind of feedback is needed to keep engineer's skills sharp.
Given that LLMs are trained on existing data and code, how do we envision their evolution without the ongoing input of software developers? Considering that developers innovate to address specific challenges, can we expect AI systems to independently advance software development without the creative problem-solving of humans?
- It's certainly possible. There are multiple possible paths. In some of them, AI generated coding drives us into a localized fitness peak that is actually an evolutionary dead end. In others, it opens up new frontiers. As I said in my book, the future is up to us, and the choices we make about how we use technology.
The real attraction of software engineering/programming has been its low financial barrier to entry compared to classical engineering which has created a large, open and active community. Isn't software engineering at risk of becoming a more expensive exercise with the advent of paid and pervasive AI tools (APIs, IDEs etc) and thus might discourage new entrants into programming?
- I don't agree. I think the opposite is happening.
Conference/ORCwAI/Pt2
Is this workflow Harper is describing written down anywhere? He's moving too fast to take notes or digest it.
Has Generative AI Already Peaked? Computerphile discussed a paper (https://arxiv.org/abs/2404.04125), No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance which found consistently that, models require exponentially more data to achieve linear improvements. If AI has to resort to feeding on their own output, has AI peaked?
- It is possible that gains from model performance per se may have peaked, but it's like Moore's law. Once simple transistor density peaked, companies found new says to drive performance.
What would you say to business leaders that are making interesting staffing and product decisions to optimise for AI usage and perhaps over-indexing on some of the buzz around AI?
- Test but verify before going all in.
12:50