hedgehog

https://knowyourmeme.com/memes/inb4--2

Ars points out that these findings contradict those of other experiments and then goes on to postulate as to why. I clicked on the link to the other experiment:

when data is combined across three experiments and 4,867 developers, our analysis reveals a 26.08% increase (SE: 10.3%) in completed tasks among developers using the AI tool

By comparison, this experiment considered 16 developers. That’s 0.3% as many as the experiments its findings contradict. Fortunately, the authors don’t claim their findings are broadly applicable. They even have a table that reads:

We do not provide evidence that | Clarification —- | —- AI systems do not currently speed up many or most software developers | We do not claim that our developers or repositories represent a majority or plurality of software development work AI systems do not speed up individuals or groups in domains other than software de- velopment | We only study software development AI systems in the near future will not speed up developers in our exact setting | Progress is difficult to predict, and there has been substantial AI progress over the past five years [2] There are not ways of using existing AI systems more effectively to achieve positive speedup in our exact setting | Cursor does not sample many tokens from LLMs, it may not use optimal prompting/scaffolding, and domain/repository-specific training/finetuning/few-shot learning could yield positive speedup

That said, the study has been an interesting read so far. I highly recommend reading it directly rather than just the news posts about it. Check out their own blog post: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

I personally find the psychological effect - the devs thought they were 20% faster even afterward - to be pretty interesting, as it suggests that even if more time overall is spent, use of AI could reduce cognitive load and potentially side effects like burnout.

I’d like to see much larger scale studies set up like this, as well as studies of other real world situations. For example, how does this affect the amount of time this takes 10,000 different developers to onboard onto an unfamiliar repository?

There’s a whole history of people, both inside and outside the field, shifting the definition of AI to exclude any problem that had been the focus of AI research as soon as it’s solved.

Bertram Raphael said “AI is a collective name for problems which we do not yet know how to solve properly by computer.”

Pamela McCorduck wrote “it’s part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, but that’s not thinking” (Page 204 in Machines Who Think).

In Gödel, Escher, Bach: An Eternal Golden Braid, Douglas Hofstadter named “AI is whatever hasn’t been done yet” Tesler’s Theorem (crediting Larry Tesler).

https://praxtime.com/2016/06/09/agi-means-talking-computers/ reiterates the “AI is anything we don’t yet understand” point, but also touches on one reason why LLMs are still considered AI - because in fiction, talking computers were AI.

The author also quotes Jeff Hawkins’ book On Intelligence:

Now we can see the entire picture. Nature first created animals such as reptiles with sophisticated senses and sophisticated but relatively rigid behaviors. It then discovered that by adding a memory system and feeding the sensory stream into it, the animal could remember past experiences. When the animal found itself in the same or a similar situation, the memory would be recalled, leading to a prediction of what was likely to happen next. Thus, intelligence and understanding started as a memory system that fed predictions into the sensory stream. These predictions are the essence of understanding. To know something means that you can make predictions about it. …

The human cortex is particularly large and therefore has a massive memory capacity. It is constantly predicting what you will see, hear, and feel, mostly in ways you are unconscious of. These predictions are our thoughts, and, when combined with sensory input, they are our perceptions. I call this view of the brain the memory-prediction framework of intelligence.

If Searle’s Chinese Room contained a similar memory system that could make predictions about what Chinese characters would appear next and what would happen next in the story, we could say with confidence that the room understood Chinese and understood the story. We can now see where Alan Turing went wrong. Prediction, not behavior, is the proof of intelligence.

Another reason why LLMs are still considered AI, in my opinion, is that we still don’t understand how they work - and by that, I of course mean that LLMs have emergent capabilities that we don’t understand, not that we don’t understand how the technology itself works.

We are. Why do you think we stopped?

It may be aware of them, but not in that context. If you asked it how to solve the problem rather than to solve the problem for you, there’s a chance it would suggest you use a reverse image search.

LLM image processing doesn’t work the same way reverse image lookup does.

Tldr explanation: Multimodal LLMs turn pictures into a ~~thousand~~ 200-500 or so ~~words~~ tokens, but reverse image lookups create perceptual hashes of images and look the hash of your uploaded image up in a database.

Much longer explanation:

Multimodal LLMs (technically, LMMs - large multimodal models) use vision transformers to turn images into tokens. They use tokens for words, too, but these tokens don’t also correspond to words. There are multiple ways this could be implemented, but a common approach is to break the image down into a grid, then transform each “patch” of a specific size, e.g., 16x16, into a single token. The patches aren’t transformed individually - the whole image is processed together, in context - but it still comes out of it with basically 200 or so tokens that allow it to respond to the image, the same way it would respond to text.

Current vision transformers also struggle with spatial awareness. They embed basic positional data into the tokens but it’s fragile and unsophisticated when it comes to spatial awareness. Fortunately there’s a lot to explore in that area so I’m sure there will continue to be improvements.

One example improvement, beyond improved spatial embeddings, would be to use a dynamic vision transformers that’s dependent on the context, or that can re-evaluate an image based off new information. Outside the use of vision transformers, simply training LMMs to use other tools on images when appropriate can potentially help with many of LMM image processing’s current shortcomings.

Given all that, asking an LLM to find the album for you is like - assuming you’ve given it the ability and permission to search the web - like showing the image to someone with no context, then them to help you find what music video - that they’ve never seen, by an artist whose appearance they describe with 10-20 generic words, none of which are their name - it’s in, and to hope there were, and that they remembered, the specific details that would make it would come up in the top ten results if searched for on Google. That’s a convoluted way to say that it’s a hard task.

By contrast, reverse image lookup basically uses a perceptual hash generated for each image. It’s the tool that should be used for your particular problem, because it’s well suited for it. LLMs were the hammer and this problem was a torx screw.

Suggesting you use - or better, using a reverse image lookup tool itself - is what the LLM should do in this instance. But it would need to have been trained to think to suggest this, capable of using a tool that could do the lookup, and have both access and permission to do the lookup.

Here’s a paper that might help understand the gaps between LMMs and tasks built for that specific purpose: https://arxiv.org/html/2305.07895v7

Thank you! That gives me a starting point that should be easy to look up!

Why is 255 off limits? What is 127.0.0.0 used for?

To clarify, I meant that specific address - if the range starts at 127.0.0.1 for local, then surely 127.0.0.0 does something (or is reserved to sometimes do something, even if it never actually does in practice), too.

Advanced setup would include a reverse proxy to forward the requests from the applications port to the internet

I use Traefik as my reverse proxy, but I have everything on subdomains for simplicity’s sake (no path mapping except when necessary, which it generally isn’t). I know 127.0.0.53 has special meaning when it comes to how the machine directs particular requests, but I never thought to look into whether Traefik or any other reverse proxy supported routing rules based on the IP address. But unless there’s some way to specify that IP and the IP of the machine, it would be limited to same device communications. Makes me wonder if that’s used for any container system (vs the use of the 10, 172.16-31, and 192.168 blocks that I’ve seen used by Docker).

Well this is another advanced setup but if you wanted to segregate two application on different subnets you can. I’m not sure if there is a security benefit by adding the extra hop

Is there an extra hop when you’re still on the same machine? Like an extra resolution step?

I still don’t understand why .255 specifically is prohibited. 8 bits can go up to 255, so it seems weird to prohibit one specific value. I’ve seen router subnet configurations that explicitly cap the top of the range at .254, though - I feel like I’ve also seen some that capped at .255 but I don’t have that hardware available to check. So my assumption is that it’s implementation specific, but I can’t think of an implementation that would need to reserve all the .255 values. If it was just the last one, that would make sense - e.g., as a convention for where the DHCP server lives on each network.

Why is 255 off limits? What is 127.0.0.0 used for?

PSTN is wiretapped.

It’s a good thing that the website itself supports sending and receiving alerts, then.

Current generation iPad Pros and Airs have the same processing power as Apple Silicon Macs. That’s more than enough for Blender. Even the base iPad and the iPad Mini likely have enough processing power - though I don’t think the base iPad has enough RAM.

Does mirroring a screen (or adding a screen) from a computer or connecting to a computer via remote desktop count?