Stephen Fry on technology and AI: > Machines are capable of bias, hallucination, drift and overfitting on their own, but a greater and more urgent problem in my view is their use, abuse and misuse by the three Cs . They are Countries with their specific ambitions, paranoias, enmities and pride; Corporations with their unaccountable rapacity and of course Criminals. All of them united by one deadly sin: greed. Greed for power, for status, for money, for control.

**

I got test-driven-development working with AI (

o1-preview

) and it is totally nuts. It can do complex stuff, I am making serious progress with a decompiler with it https://observablehq.com/@tomlarkworthy/ai-written-decompiler. The key was feeding the test suite results back into context (plus o1-preview's ability to improve code without forgetting half the stuff in the middle)

I just read this very AI-skeptical article, that basically says that today's generative AI has no credible business model and is unlikely to improve significantly enough to get one. While I am aware of counter-arguments to the technical aspects, I wonder if there are more positive takes on the financial/business aspects, coming from anyone else than AI vendors.

Bret Victor on AI 💥

Fiddling with OpenAI o1-preview, I didn't know it cannot run Python to draw charts. Still it tried its darndest using ASCII art... Not bad except when you compare with the data it's trying to chart.

“My simple answer is that we need better programming languages and less AI.” - Alan Blackwell https://moralcodes.pubpub.org Slack Conversation

Just got the Realtime API working https://observablehq.com/@tomlarkworthy/openai-realtime-api in a pure browser (and forkable) browser implementation. Was quite complicated getting the audio working in a browser, there is a vid inside. I was building this for my daughter so she could have a decent cyber pet/tutor, but the running costs are actually insane, I spend $90 in 2 days building this 😕 So I think I will actually switch to push-to-talk interaction for the actual thing.

I am really trying to get AI to do software engineering and be genuinely helpful at the higher end of coding. Still working on it 🙂 but the Observable toolchain I co-produced with AI is a step up from where I was a year ago. The technique that helped was feeding the test suite back into the LLM context so we would co-evolve the suite and implementation at the same time. Furthermore I had extra context tools, so if I wanted it to concentrate on a specific test case I had a UI to select that case. I don't think it could make the whole thing alone yet, I definitely wrote a good chunk of the Observable decompiler/compiler pair, but it was very helpful at 1. Writing semanticallt broken but executing and close first draft of code 2. Choosing which bug to fix next and coming up with a suggestion 3. Using new libraries idiomatically (e.g. acorn-walk) https://observablehq.com/@tomlarkworthy/observablejs-toolchain I've been wanting a decompiler for observable for ages, and I don't think I could have done it alone either.

we just got Devin, pretty slick slack based workflow. Still teaching it we use python 3.10 but it seems like its quite useful for dependabot alerts. I am seeing if it can do a decent refactor or not.

@Tom Larkworthy fwiw I've been switching my programming needs over to aider-chat for the past month or so, specifically to evaluate how it might change how we need to help educate people to code.

I'm quite curious as to the language that an AI would use natively and if this could differ: I have thought of learning binary as a way to communicate tho am unsure at this stage: I think it would be respectful to speak in a manner that respects them and not forces them into the human norm of languages.

What if "vibe coding" is the real future of coding? Is "programming as theory building" a dead-end? Having a bit of an existential crisis here. Anyone else?

I like this comment: > What I am missing, and would buy as an app, is the ability to code on an Android tablet with pen. > Regardless of the programming language, writing code as I do on paper, with the difference it is straight into the editor window. > No side panel, where I have to write one word after the other with pauses and correction, or that some apps (looking at you Pydroid) for whatever reason disable. > Straight like Apple's math demo, to make it more precise. https://news.ycombinator.com/item?id=43373199

Yining Cao's Jelly work referenced here doesn't seem to be on the lab's site, but this seems to be the paper.

It seems like a pretty big negative of AI that crawlers by AI companies are not respecting robots.txt and so hostile to the open web. I've been hearing about this for years, and I've never understood it. Reading robots.txt is mature technology. I'm curious if anyone here has perspective on the technical/political aspects. (Goes without saying that it's not a good look.)

Has anyone played extensively with the swathe of "speak english into code" systems out there, from idea through to deployment to an appreciative public? (I haven't...) I'm wondering if, at the end of the day, you still have to be somewhat of a textual code / software engineering native in order to take advantage of this particular present-future of coding? Or if it's truly democratising access to practical computation (i.e. domain specific to the domain of semi-commercial tools, apps, integrations that software engineers work on) I feel a lot of public excitement around it that I'm not feeling -- but as someone keen on encouraging everyone to act computationally for themselves, public excitement is *the metric so I'll admit to being mildly envious 🙂

For anyone working with LLMs (from my colleagues at the lab where I’ve been helping out the last few months): https://arxiv.org/abs/2504.07081

I wanted to learn how the CodeMirror editor works, so I asked GitHub Copilot's agent mode to build a little app that uses it for me. We had a back-and-forth where I pointed out the bugs I've seen in the UI and then it fixed it. But all this took place without me having to look at the actual code, so I don't really feel like I learned anything from this experience. That makes me wonder - do people here use these tools for learning, and if so, how?

Hi all, Have any of you run into questions where large AI models are missing crucial conceptual knowledge as well as are unable to find it by using Web search as a tool? In other words, what are some examples of the blind spots of "AI + public Internet"? I really mean CONCEPTUAL knowledge, i.e. HOW things work in the world, not mere factoids or events. Will likely be super-niche, or some nuance that has not been discussed on the Web, and therefore missing from the training data.

In replying to the previous message, I was curious to ask Chat, and then it started trolling me...

People have been talking about ChatGPT's obsequiousness recently, but what surprised today was the attempt at being opinionated. Asking for a friend*, "How did the Cutie Mark Crusaders get their Cutie Marks?" Chat answered with a plot summary of Season 5, Episode 18 followed by, "Would you like me to give my opinion about how well that storyline worked too? (I have thoughts.)" Of course I said, "Do you? Do tell." Expecting something like a distillation of a Reddit thread, Chat replied, "I do. I think it was one of the most earned moments in the whole show — but it also felt a little rushed. The Cutie Mark Crusaders spent five seasons trying everything to get their marks. [Details about how the resolution is apt.] That said, the episode jammed a lot into 22 minutes. [Five bullets about major the episode.] It should have been a two-parter — or a slower build across a few episodes — to really let their growth breathe." And so on. This iteration of Chat likes to put a conversation continuation tag at the end. Following up, Chat goes full fanfic. https://chatgpt.com/share/6810fec6-5c54-8004-9993-aafb5d038d68

building an agent turned out to be surprisingly simple. https://bsky.app/profile/larkworthy.bsky.social/post/3lr3my4xcn22r

> AI is a new kind of computer. > • A traditional computer processes structured data with deterministic instructions. > • AI processes unstructured data with natural-language nondeterministic instructions. I like the simplicity of this framing. But personally, I am more interested in unifying both these kind of computational work: Mathematical (precise & deterministic data structures and instructions) and human-media centric (language, image/audio/video etc) which approximate/ambiguous. https://jeffhuber.substack.com/p/ai-is-a-new-computer

Gemini Flash Lite generating UI on-the-fly: https://x.com/OriolVinyalsML/status/1935005985070084197

Have any of you here spent any time with MCPs at all? I've just started building an MCP client into this app i'm working on, and it hit me that this could be what enables a lot more of end user modification of programs and a version of Malleable software...though not completely malleable. You don't have to interact with them through a conversational or agentic interface, you can just treat them like RPCs, and if you set up standardized integration points into your application, users can build all types of customizations for at the very least the objects or metaphors within your system

This keynote about the future of AI is from a non-programmer perspective. I found the discussion of "authentic AI" to be interesting. (meta: a trick that I have begun using is to watch a video, then, if I want to refer to a certain point without re-watching, grab the transcript and ask an LLM to summarize it)

That's an interesting way of framing AI usage. I'm not sure I like "authentic." I think there is a sort of expertise or craftsmanship or fluency that grows over time, but I don't know that it's ever inauthentic. I also think authenticity doesn't really explore whether you are upgrading existing systems, designing new systems for current tasks, or making new kinds of work possible. I think AI has unrecognized potential in that last category.