jwilliams

The mentality shift of renting vs. owning the gpus is huge. When renting, each experiment costs money and I had to ask myself is it worth it. When owning, it feels like not running experiments is costing me money.

I feel like there is some very deep generalizable wisdom buried here.

Rovelli is arguing (I think) that we need to fundamentally view consciousness as a natural phenomenon - albeit one that is extremely complex and poorly understood.

So we ditch the philosophical puzzle and focus on the reality we can perceive and reason on. The problem is that consciousness is a philosophical invention (and a slippery one at that).

We're in the wrong frame. If you accept consciousness is a thing you end up in this weird tautological state - it's not special, but we've put it in a special category.

If you view via a grounded, practical frame, you probably don't care about consciousness. The fact that it's undefinable is probably a major clue.

Agree.

We have a bunch of tools - specs, code, tests. All of these really are models of the end outcome we're trying to capture.

You could just build something, see if you're right and then build it again. If that seems ridiculous, what makes a spec special that it can work first time?

Why we've not done this historically is code is annoying and (was) relatively expensive. You can rough out a spec document and get feedback from a wide variety of stakeholders -- after all, they can all read a document.

If you can use AI to explore a problem space and get feedback directly, that's definitely a whole new tool in the kit.

tsz looks impressive. Congrats!

Claude Code navigates a codebase the way a software engineer would: it traverses the file system, reads files, uses grep to find exactly what it needs, and follows references across the codebase. It operates locally on the developer’s machine and doesn’t require a codebase index to be built, maintained, or uploaded to a server....

Agentic search avoids those failure modes. There's no embedding pipeline or centralized index to maintain as thousands of engineers commit new code. Each developer's instance works from the live codebase.

The frame of "the way a software engineer would" and the conclusion seem at odds. I'd love to be schooled otherwise?

I use autocomplete/LSPs all the time and they're useful. That's an index? Why wouldn't Claude be able to use one? Also a "software engineer" remembers the codebase - that's definitely a RAG. I have a lot of muscle memory to find the file I need through an auto-completed CMD+P.

It doesn't need to particularly be real-time across thousands of engineers -- just the branch I'm on.

It's rare that I'd be navigating a codebase from first-principles traversal. It would usually be a new codebase and in those cases it's definitely not what I'd call an optimal experience.

I wish they'd have done this in a separate Codex app. On desktop I greatly prefer having Codex separate from ChatGPT... As compared to Claude, which is growing so fast and adding features so quickly it seems bolted together (I get why they do it, integrations/MCP-wise).

This specific feature is more akin to Remote Control in Claude. You could already kick off Codex Cloud tasks (although it's just a little more fiddly to do so).

If you can move to Codex Cloud (or "Claude Code for the Web"), I think it's the superior approach. Start it there, and just pick it up from the PR if necessary.

I've worked on plenty of code where the original author is long gone. Not sure that holds water at all.

I've personally had people talk about token leaderboards at their work. Amazon and Meta did have ones, but I'd take with a decent grain of salt.

We all know it's such an insanely gameable metric you'd be insane to actually use it...

It’s Still Your Code... AI maximalists will read this section and scoff. They’re already vibe coding everything and have little to no idea what the generated code looks like.

This frames the argument like a dichotomy. And to be honest, using the Social Media "vibe-coding" as a strawman risks anchoring against something that's a mirage.

There are plenty of good engineers getting good results whilst accepting code-ownership as a continuum.

If Claude goes down tomorrow, can you still do your job?

This is a valid counterpoint, but doing software is already a tricky set of dependencies. The answer here isn't automatically "you need to be able to do everything". It could simply be also use Codex.

I think the overall point is well made, I just don't agree with the absolute framing. There are things you can hand over AI safely. Even if you start small and increment it'll have a decent impact.

Are you specifically pointing at a different experience between free + paid? Or just that the free version is unimpressive?

I'm using paid on TypeScript and it's genuinely terrific. Subjectively I think it has the edge over Opus.

I'd be surprised if OpenAI is hamstringing the free version. That would seem crazy from a GTM PoV. If anything the labs seem to throttle the heavy paid users.

I've been working on JavaScript runner for untrusted code. The whole API is only exposed via messages passed over stdio. Security layers: V8 isolates, two-stage seccomp, frozen globals, mount namespaces, landlock, and more. https://github.com/jonathannen/hermit

Plus it's too early to really show, but also working on a dataflow language (w/ immutable data) that uses some code semantics from Rust/Zig and friends: https://github.com/jonathannen/badger

"AI attribution by default" was an intentional feature, but there were also bugs. This is why the post is so unclear.

Are they apologizing? Was it a bug? Why did they make this decision and what's the end goal? It's so unclear from the message - as evidenced by a lot of the responses.

The other huge problem is you never tell the user what they'll get out of the tour. People will invest in a tour if they understand the reward (and "learning" can't be the reward).

And it's a very weak example in my view.

In fact it's in the article - the reason the Great Wall myth exists is because it's so prevalent on the internet... Presumably because a a lot of conscious people also believe it. Plenty of people walking around today, fully conscious, believe things that aren't factually true.

A child might make the same "seen from spain" mistake, but we would never say the same child wasn't conscious.

It's a tough one to wade into because the definition is so slippery. Most debate seems to focus on the definition of consciousness rather than the evidence... which is a major tell.

To my mind it's better to ask how the definition one way or the other has utility. It's less important to me that Dawkins believes an LLM to be conscious, but more important what specifically he thinks the implications of that are (and equally so, for me to interrogate my own beliefs if I happen to disagree).

Fair call.

I don't think couching it as conspiracy is the right frame either. This is not a one-off. I think a critical eye is warranted.

I don't know what "openclaw" is. It's not something I have knowledge of, and it doesn't appear in your memory or this project's context.

As others have pointed out, Anthropic is allowed to have TOS, even if we disagree with it.

But having Claude deny the existence of OpenClaw is a way more hazardous and likely straight up violates Claude's Constitution: https://www.anthropic.com/constitution

Sad news. I met Craig very briefly at a conference probably a decade back. I pretty much was a self-study in genetics at the time... so let's just say I wasn't in Craig's league. Despite this he was very engaged and took the time for a very thoughtful chat.

Codex is pretty good. Its friction to switch but I think it’s sensible being across multiple AI toolchains.

I do "just use fetch" nowadays -- but I have to say, axios definitely has better ergonomics than fetch, especially for calling APIs.

I think there is more here: Anthropic's whole market positioning is based on trust. It's literally their reason for being.

The Claude constitution has a major section about not being deceptive. Now this is GTM, not the model, but there is clearly a coherence problem here... and if anyone should realize the important of their market positioning it's GTM.

“I don’t have access to X” is only correct after tool_search confirms no matching tool exists.

Yay! This will be a big win. I'm glad they fixed this. The number of times I've had to prompt "you do have access to GitHub"...

I'm surprised it's reporting is listed <5% - I thought it was pretty much ipv6 first?

Is it just me or is this sorely lacking any consideration for inserts? Or durability?

I use pglite for unit/integration tests and it's been fantastic.

Note that it uses "single user mode." This means a single connection at a time, and thus no concurrent transactions. That takes you a little bit closer to SQLite's single-writer.

It’s got plenty of precedent - AutoCAD has a Lisp equivalent since 1986[1]. Arguably it is what made it the powerhouse CAD tool.

Irrespective - This project is pretty cool to see!

[1] https://en.wikipedia.org/wiki/AutoLISP

It’s terrific to see this. I’m definitely going to give it a whirl. I’ve been working on a specific JavaScript isolate[^1]. This is great source of inspiration for it.

[^1]: https://github.com/jonathannen/hermit

GraalVM is genuinely great -- Native Image and the polyglot story are impressive.

I was put off by the earlier licensing - it was confusing, which wasn't great in a license. The GraalVM Free Terms and Conditions "GFTC" now seems better (curious if people agree?), but I wonder if it came too late.

The decoupling from Java SE was good in many ways, but it also made the future a little less clear too.

It's live! If you're on the latest cc you can use /buddy now.