I assumed I’d provide just a few takeaways and reflections primarily based on final week’s first AI Codecon digital convention, Coding with AI: The End of Software Development as We Know It. I’m additionally going to incorporate just a few quick video excerpts from the occasion. If you registered for Coding with AI or when you’re an present O’Reilly subscriber, you may watch or rewatch the entire thing on the O’Reilly studying platform. If you aren’t a subscriber but, it’s simple to begin a free trial. We’ll even be posting extra excerpts on the O’Reilly YouTube channel within the subsequent few weeks.
But on to the promised takeaways.
First off, Harper Reed is a mad genius who made everybody’s head explode. (Camille Fournier apparently has joked that Harper has rotted his mind with AI, and Harper really agreed.) Harper mentioned his design course of in a chat that you just would possibly wish to run at half velocity. His greenfield workflow is to begin with an thought. Give your thought to a chat mannequin and have it ask you questions with sure/no solutions. Have it extract all of the concepts. That turns into your spec or PRD. Use the spec as enter to a reasoning mannequin and have it generate a plan; then feed that plan into a special reasoning mannequin and have it generate prompts for code era for each the appliance and checks. He’s having a wild time.
Agile Manifesto coauthor Kent Beck was additionally on Team Enthusiasm. He instructed us that augmented coding with AI was “the most fun I’ve ever had,” and mentioned that it “reawakened the joy of programming.” Nikola Balic agreed: “As Kent said, it just brought the joy of writing code, the joy of programming, it brought it back. So I’m now generating more code than ever. I have, like, a million lines of code in the last month. I’m playing with stuff that I never played with before. And I’m just spending an obscene amount of tokens.” But sooner or later, “I think that we won’t write code anymore. We will nurture it. This is a vision. I’m sure that many of you will disagree but let’s look years in the future and how everything will change. I think that we are more going toward intention-driven programming.”
Others, like Chelsea Troy, Chip Huyen, swyx, Birgitta Böckeler, and Gergely Orosz weren’t so positive. Don’t get me flawed. They assume that there’s a ton of fantastic stuff to do and be taught. But there’s additionally a variety of hype and unfastened pondering. And whereas there can be a variety of change, a variety of present expertise will stay necessary.
Here’s Chelsea’s critique of the latest paper that claimed a 26% productiveness enhance for builders utilizing generative AI.
If Chelsea will do a sermon each week within the Church of Don’t Believe Everything You Read that consists of her displaying off varied papers and giving her dry and insightful perspective on how to consider them extra clearly, I’m so there.
I used to be a bit shocked by how skeptical Chip Huyen and swyx had been about A2A. They actually schooled me on the notion that the way forward for brokers is in direct AI-to-AI interactions. I’ve been of the opinion that having an AI agent work the user-facing interface of a distant web site is a throwback to display scraping—certainly a transitional stage—and whereas calling an API can be the easiest way to deal with a deterministic course of like cost, there can be a complete lot of different actions, like style matching, that are perfect for LLM to LLM. When I take into consideration AI searching for instance, I think about an agent that has discovered and remembered my tastes and preferences and particular targets speaking with an agent that is aware of and understands the stock of a service provider. But swyx and Chip weren’t shopping for it, a minimum of not now. They assume that’s a great distance off, given the present state of AI engineering. I used to be glad to have them convey me again to earth.
(For what it’s price, Gabriela de Queiroz, director of AI at Microsoft, agrees. On her episode of O’Reilly’s Generative AI within the Real World podcast, she mentioned, “If you think we’re close to AGI, try building an agent, and you’ll see how far we are from AGI.”)
Angie Jones, alternatively, was fairly enthusiastic about brokers in her lightning discuss about how MCP is bringing the “mashup” period again to life. I used to be struck particularly by Angie’s feedback about MCP as a sort of common adapter, which abstracts away the underlying particulars of APIs, instruments, and information sources. That was a robust echo of Microsoft’s platform dominance within the Windows period, which in some ways started with the Win32 API, which abstracted away all of the underlying {hardware} such that utility writers not needed to write drivers for disk drives, printers, screens, or communications ports. I’d name {that a} energy transfer by Anthropic, aside from the blessing that they launched MCP as an open customary. Good for them!
Birgitta Böckeler talked frankly about how LLMs helped scale back cognitive load and helped assume by way of a design. But a lot of our day by day work is a poor match for AI: massive legacy codebases the place we alter extra code than we create, antiquated know-how stacks, poor suggestions loops. We nonetheless want code that’s easy and modular—that’s simpler for LLMs to know, in addition to people. We nonetheless want good suggestions loops that present us whether or not code is working (echoing Harper). We nonetheless want logical, analytical, essential eager about drawback fixing. At the tip, she summarized each poles of the convention, saying we’d like cultures that reward each experimentation and skepticism.
Gergely Orosz weighed in on the continued significance of software program engineering. He talked briefly about books he was studying, beginning with Chip Huyen’s AI Engineering, however maybe the extra necessary level got here a bit later: He held up a number of software program engineering classics, together with The Mythical Man-Month and Code Complete. These books are many years outdated, Gergely famous, however even with 50 years of device growth, the issues they describe are nonetheless with us. AI isn’t prone to change that.
In this regard, I used to be struck by Camille Fournier’s assertion that managers like to see their senior builders utilizing AI instruments, as a result of they’ve the talents and judgment to get probably the most out of it, however usually wish to take it away from junior builders who can use it too uncritically. Addy Osmani expressed the priority that fundamental expertise (“muscle memory”) would degrade, each for junior and senior software program builders. (Juniors might by no means develop these expertise within the first place.) Addy’s remark was echoed by many others. Whatever the way forward for computing holds, we nonetheless must know the way to analyze an issue, how to consider information and information constructions, the way to design, and the way to debug.
In that very same dialogue, Maxi Ferreira and Avi Flombaum introduced up the critique that LLMs will have a tendency to decide on the most typical languages and frameworks when attempting to resolve an issue, even when there are higher instruments out there. This is a variation of the remark that LLMs by default have a tendency to provide a consensus answer. But the dialogue highlighted for me that this represents a threat to ability acquisition and studying of up-and-coming builders too. It additionally made me surprise about the way forward for programming languages. Why develop new languages if there’s by no means going to be sufficient coaching information for LLMs to make use of them?
Almost all the audio system talked in regards to the significance of up-front design when programming with AI. Harper Reed mentioned that this appears like a return to waterfall, besides that the cycle is so quick. Clay Shirky as soon as noticed that waterfall growth “amounts to a pledge by all parties not to learn anything while doing the actual work,” and that failure to be taught whereas doing has hampered numerous initiatives. But if AI codegen is waterfall with a quick studying cycle, that’s a really totally different mannequin. So this is a vital thread to tug on.
Lili Jiang’s closing emphasis that evals are rather more complicated with LLMs actually resonated for me, and was in step with lots of the audio system’ takes about how a lot additional we’ve got to go. Lili in contrast an information science venture she had completed at Quora, the place they began with a fastidiously curated dataset (which made eval comparatively simple), with attempting to cope with self-driving algorithms at Waymo, the place you don’t begin out with “ground truth” and the proper reply is very context dependent. She requested, “How do you evaluate an LLM given such a high degree of freedom in terms of its output?” and identified that the code to do evals correctly may be as massive or bigger than the code used to form the precise performance.
This completely matches with my sense of why anybody imagining a programmer-free future is out of contact. AI makes some issues that was once exhausting trivially simple and a few issues that was once simple a lot, a lot tougher. Even when you had an LLM as choose doing the evals, there’s an terrible lot to be found out.
I wish to end with Kent Beck’s considerate perspective on how totally different mindsets are wanted at totally different phases within the evolution of a brand new market.
Finally, a giant THANK YOU to everybody who gave their time to be a part of our first AI Codecon occasion. Addy Osmani, you had been the right cohost. You’re educated, an ideal interviewer, charming, and a variety of enjoyable to work with. Gergely Orosz, Kent Beck, Camille Fournier, Avi Flombaum, Maxi Ferreira, Harper Reed, Jay Parikh, Birgitta Böckeler, Angie Jones, Craig McLuckie, Patty O’Callaghan, Chip Huyen, swyx Wang, Andrew Stellman, Iyanuoluwa Ajao, Nikola Balic, Brett Smith, Chelsea Troy, Lili Jiang—you all rocked. Thanks a lot for sharing your experience. Melissa Duffield, Julie Baron, Lisa LaRew, Keith Thompson, Yasmina Greco, Derek Hakim, Sasha Divitkina, and everybody else at O’Reilly who helped convey AI Codecon to life, thanks for all of the work you set in to make the occasion a hit. And due to the virtually 9,000 attendees who gave your time, your consideration, and your provocative questions within the chat.
Subscribe to our YouTube channel to observe highlights from the occasion or develop into an O’Reilly member to observe all the convention earlier than the following one September 9. We’d love to listen to what landed for you—tell us within the feedback.