DougBTX

> Are the AI companies really living in an echo chamber?

The author tested 12 models, and only one was consistently wrong. More than half were correct 100% of the time.

A better conclusion would be that there’s something in particular wrong with GPT-5 Chat, all the other GPT 5 variants are OK. I wonder what’s different?

Absolutely fine, as long as the success flag is predicted by the model ensemble under test. That’s how Claude Code works for example, it will continue to iterate until success (or it will give up with failure at a certain point).

Agreed, this is just a quote on his link blog. Better to post Reddit thread directly. I wouldn’t expect a plain Reddit link to get to the HN front page though.

My hot take is that your friend should show you what they’re using, not just dismiss Copilot and leave you hanging!

Yeah, the chess example is interesting. The best specialised AIs for chess are all clearly better than humans, but our best general AIs are barely able to play legal moves. The ceiling for AI is clearly much higher than current LLMs.

The once-per-millennium marks are captivating. They feel almost within our understanding, but not quite.

Recent LLMs have larger context windows to process more data and tool use to get new data, so it would be surprising if there’s a fundamental limitation here.

For a little while someone added Claude’s noreply email address to their account, so their profile started appearing alongside private repo commits!

Agreed, a screwdriver isn’t broken just because it isn’t a good hammer. The title seems misleading, I was expecting a bug, memory unsafe etc.

Allowing more safe uses seems OK to me, but obviously expanding the functionality adds complexity, so there’s a trade off.

Not the prompts exactly, but these commits should give you a good idea about what they were: https://github.com/steipete/steipete.me/pull/106/commits

Model weights are code, for a dive into that see [0]. That shows how to encode Boolean logic using NAND gates in an MLP.

The expressivity is there, the only question is how to encode useful functions into those weights, especially when we don’t know how to write those functions by hand.

[0] http://neuralnetworksanddeeplearning.com/chap1.html

> the more general view that Rust is attempting to be as close as it can get to a language that reads like it has a garbage collector without having one.

I've used Rust a fair amount, and I've never seen that expressed as a goal.

A couple of general principles followed by Rust are to prefer explicit code over implicit conversions and to support local reasoning. Those are both present here: the borrow needs to be made explicitly, rather than implicitly based on code later on.

Do any of these support streaming dictation, to show text progressively while dictating?

From the article:

> A 2024 GitHub survey found that nearly all enterprise developers (97%) are using Generative AI coding tools. These tools have rapidly evolved from experimental novelties to mission-critical development infrastructure, with teams across the globe relying on them daily to accelerate coding tasks.

That seemed high, what the actual report says:

> More than 97% of respondents reported having used AI coding tools at work at some point, a finding consistent across all four countries. However, a smaller percentage said their companies actively encourage AI tool adoption or allow the use of AI tools, varying by region. The U.S. leads with 88% of respondents indicating at least some company support for AI use, while Germany is lowest at 59%. This highlights an opportunity for organizations to better support their developers’ interest in AI tools, considering local regulations.

Fun that the survey uses the stats to say that companies should support increasing usage, while the article uses it to try and show near-total usage already.

> For example, if 'you' agreed to develop a piece of software 'exclusively' for the customer, and then use AI to create substantial parts of the software, then neither it was 'you' who developed that, nor was it 'exclusively' for the customer

On the other hand, if ‘you’ had taken no action at all, then there would be no software at all. The actions by ‘you’ are necessary for the software to exist, so the argument must be about whether those actions count as development or not. Is the definition of development written down anywhere?

AFAIK none of those models have been trained to produce bounding boxes. On the other hand Gemini Pro has, so it may be worth looking at for your use case:

https://simonwillison.net/2024/Aug/26/gemini-bounding-box-vi...

The article doesn’t explain why this is interesting. Here’s a different way of putting it:

A classic pinhole camera is a box with a hole in it. The hole is small, so not much light gets in. We replace it with a lens that acts in effectively the same way, but lets in more light. The pinhole is inside the camera, and there is an obstacle in front of the subject, so the camera can’t see it. The fun thing here, is that this camera has a pinhole _outside_ the camera, so the pinhole can be positioned on the other side of the obstacle, and therefore take a picture of the subject.

Maybe the opposite, slow rollbacks cause escalating incidents.

This article claims that 1e2 is interpreted as a string, while this other article on the front page[0] claims that 556474e378 is interpreted as a number. What's correct?

The YAML "Scalars" section[1] says:

> A few examples also use the int, float and null types from the JSON schema.

And includes these examples:

    canonical: 1.23015e+3
    exponential: 12.3015e+02

So, is the "+" required here or not? Is a YAML parser buggy if it doesn't parse all JSON numbers as numbers?

Edit: Ah, further on, it says:

    Canonical Form
    Either 0, .inf, -.inf, .nan or scientific notation matching the regular expression
    -? [1-9] ( \. [0-9]* [1-9] )? ( e [-+] [1-9] [0-9]* )?

The example 1e2 clearly matches this regex, so his YAML parser is broken.

Edit edit:

In YAML 1.1, there were separate definitions of float[2] and int[3] types, where only floats support "scientific" notation, and must have a ".", unlike JSON.

So this article is talking about YAML 1.1, while the other article is talking about YAML 1.2.

[0] https://news.ycombinator.com/item?id=41498264

[1] https://yaml.org/spec/1.2.2/#23-scalars

[2] https://yaml.org/type/float.html

[3] https://yaml.org/type/int.html

Rust has a great package manager, so moving libs into std doesn’t bring much benefit.

On the other hand a change like this perf improvement can be released without tying it to a language version, that’s good too.

The “inherently” means not by default, i.e., the runtime has to support moving green threads between OS threads itself.

Of course logic gates apply logical reasoning to solve problems, they are not much use for anything else (except as a space heater if there are a lot of them).

Inside a constructor you can access a partially initialised "this" value, and even call methods on it, which leads to rules like: "Do not call overridable methods in constructors"[0], as they can lead to surprising, non-local, bugs.

Rust has functions associated with types which are conventionally used like constructors, but critically the new objects must have all their fields provided all at once, so it is impossible to observe a partially initialised object.

[0] https://learn.microsoft.com/en-us/dotnet/fundamentals/code-a...

> Where's the paradox?

Exactly. I propose that the paradox is in first-past-the-post voting, a 5% swing leads to a 100% change in representation. How can that be?

That was a pretty interesting twist in the story. There is a check, but due to a confusion in semantics it checks that the user is root (uid == 0) instead of the pointer!

I’d recommend testing the Apply step in a staging environment before rolling out.

For releases, the plans don’t need to be committed to version control. Instead they can be generated in CI just before release, then put reviewing the plan on the checklist before confirming the release to prod.

FTA:

> The Houthis first attacked commercial ships connected to Israel, expanding to those with ties to the U.S. and United Kingdom following coordinated strikes by the two countries, in partnership with other nations. The Yemen-based group then began to attack any ship going to or from Israel, as well, before declaring they would attack nearly any commercial ship transiting through the Red Sea, as well the Indian Ocean and Arabian Sea, USNI News previously reported. The Houthis have also claimed they would target some ships in the Mediteranean Sea.

> Hiding the truth seems like the exact opposite of that.

It seems like “true defamation” could be a lie of omission. If someone harps on about a past misdeed but omits N years of atonement, that’s misleading, even if true.

"At rest" is used to contrast with "in flight", where the data is being transferred between computers. So data "in flight" is protected by HTTPS etc, once it has been transferred it is "at rest" on the destination computer (even if that computer is still online).

On the other hand, A Pixel Is Not A Little Square[0] would disagree, a raster is a grid sample of a continuous function.

[0] http://alvyray.com/Memos/CG/Microsoft/6_pixel.pdf