iamjackg

It depends on what your "currency" is: inference cost vs. models getting dumber/slower with a fuller context.

I find the concept of LLM "brain surgery" fascinating, precisely because of how opaque the network is. One of the first things I did back when llama.cpp first got vision model support was hack the code to zero out (or otherwise modify) random numbers in the image embedding generated by the projector and then ask the LLM to describe the image. It was absolutely fascinating.

It would go from a normal description of the item in the picture to suddenly seeing people clapping in the background that were not there, or making up some other stuff. I kinda stopped after a while, but I should pick that back up and do a more coherent experiment to see if I can find any correlation between vector dimensions and "meaning."

Nobody is promoting a product. Simon is just sharing an experiment he attempted. No products being sold here.

Heck yeah! Love the VisiData shoutout. Echoing other people's desire for a web UI, mostly so I don't have to be the sole Maintainer of the Truth as the only resident household technomancer.

EDIT: alternatively, exposing the data/functionality via MCP or similar would allow me to connect this to an agent using Home Assistant Voice, so anybody in the house could ask for changes or add new information.

Maybe your point was that, but that's not the point the person you replied to was addressing. Nobody was arguing about the specific definition of "modern."

The original commenter made a very clear claim: that the most recent "peak" system was the Xbox, which was discontinued in 2005, and that everything after that has been a rehash.

It's not unsolved, at least not the first part of your question. In fact it is a feature offered by all main LLM providers!

- https://platform.openai.com/docs/guides/prompt-caching

- https://platform.claude.com/docs/en/build-with-claude/prompt...

- https://ai.google.dev/gemini-api/docs/caching

Curious how this will fare when playing Pokemon Red.

Oh absolutely -- this was a no-stakes personal project, so I was happy to rely on pre-made solutions and learn a thing or two along the way.

Somewhat related: a while ago I was working on a project and wanted to use an RS485 to TTL conversion board which came with badly translated instructions. However, somebody had reverse engineered the design and uploaded an EasyEDA schematic. I shoved the raw JSON for the schematic (which looked quite cryptic to me) into Gemini 2.5 Pro and asked it if it could understand it, and it cheerfully responded with:

> Of course, Jack. I can understand the schematic from the provided JSON file. It describes an RS485 to TTL Converter Module. > Here is a detailed breakdown of the circuit's design and functionality

...followed by an absolutely reasonable description of the whole board. It was imprecise, but with some guidance (and by putting together my basic skills with Gemini's vast but unreliable knowledge) I was able to figure out a few things I needed to know about the board. Quite impressive.

Yeah I immediately switched to a Honeywell Z-Wave thermostat as soon as I got the email that they were discontinuing them.

I was pleasantly surprised by how responsive it was, and overjoyed when I clicked back and immediately came back to HN: no messy history. Genius idea!

This seems to be done through the accessibility services, which makes sense: they're built for apps like screen readers that must by nature be able to read the content of other apps. Each app has to be manually authorized by going to the accessibility menu, and there's a big scary pop up that says "enabling this app will allow it to see and interact with everything you do."

This technique was also used in mid-late non-motion-plus Wii games to smooth out the pointer movement! Early games hadan incredibly twitchy pointer because they were simply mapping the IR data 1:1 to cursor movement, whereas later ones have an invisible circle around the cursor and only move the cursor itself once the circle edges start "dragging" it.

It is the best version of fuzzy search I have ever seen: the ultimate "tip of my tongue" assistant. I can ask super vague things like "Hey, I remember seeing a tool that allows you to put actual code in your files to do codegen, what could it be?" and it instantly gives me a list of possible answers, including the thing I'm looking for: Cog.

I know that a whole bunch of people will respond with the exact set of words that will make it show up right away on Google, but that's not the point: I couldn't remember what language it used, or any other detail beyond what I wrote and that it had been shared on Hacker News at some point, and the first couple Google searches returned a million other similar but incorrect things. With an LLM I found it right away.

Technically speaking, Claude Code is an agent, for example. It's just a fancy term for an LLM that can call tools in a loop until it thinks it's done with whatever it was tasked to do.

ChatGPT's Deep Research mode is also an agent: it will keep crawling the web and refining things until it feels it has enough material to write a good response.

I'd love more info about this, because I've been paying for Youtube premium for years and I haven't seen a single ad.

I have very fond memories of using Opera Mini on my N-Gage and on a bunch of later feature phones that only had J2ME support. It felt revolutionary at the time, especially since data plans were still extremely restrictive.

Wifi on phones was also not quite a thing yet, so I had found an application for S60 phones that allowed them to share a computer's connection via Bluetooth. The range was extremely limited, but enough for me to browse the internet from my bed!

It doesn't cure it, the same way wearing glasses doesn't cure poor eyesight.

I'd argue that there isn't one: you have to offer multiple choices. Auth through any TOTP app, Yubi key, pre-generated codes, mailing a physical code generator, etc.

Perhaps I'm ignorant, but I wonder why I haven't seen this aspect explored in post-apocalyptic media more. I mean, beyond the obvious fact that if we can "keep making computers" all of a sudden the apocalypse doesn't sound quite that bad, which would somewhat lower the stakes.

Sorry, I'm a bit confused. That link seem to state the opposite? The user is asking if 911 can still override the location settings on the phone, and the answer seems to be "yes."

Yeah, this one is tough. I tried doing today's puzzle and I immediately knew the full solution, so I guess the designer's dilemma is whether or not to allow Wheel of Fortune-style "I would like to give the full answer", and if so how to score it. Did you do as much "work" as somebody who solved all the puzzles? I would argue not, so it should probably be scored less, but how much?

You might not need to, but LLMs don't have perfect recall -- they're (variably) lossy by nature. Providing documentation is a pretty much universally accepted way to drastically improve their output.

It's often the limiting factor to getting started, though. Idiosyncratic interfaces and control methods make it really tedious to start learning from scratch.

You're right! It looks like I misunderstood the report and the "hidden opcodes" are only accessible to the ESP32 itself, not to connected devices? The article is somewhat confusingly worded.

Great. Time to replace half my home automation devices! This is not entirely unexpected, regardless of whether it was intentional or not, but it still hurts. Although I guess it means it might be easier to take control of existing devices without having to open them up and connect to the GPIOs.

I wonder if this is patchable at all?

Well, in this case it seems like there already is an official Flatpak, but Fedora is overriding it with their own, which is of lower quality. The problem was already solved, unless Fedora is claiming that their Flatpak is better in some specific way.

I think that using Mario as part of the sponsor read is possibly even worse from a legal/Nintendo perspective: you can't even try to make a good faith transformative use argument there. I also subjectively found it very icky. The project itself is very cool though.

I'm very interested in their Jailbreak evaluations: they're new to me. I might have missed previous mentions.

That seems more unlikely to me -- training is not free and takes a long time, so it would not result in "[enhancing] the usefulness of results displayed in the Gemini app" and it being "particularly helpful to our users looking for up-to-date information."

Fine-tuning, which is cheaper and faster, has been proven to not be a good solution to "teach" models new facts.

I think what's most likely here is that Gemini will have access to a form of RAG based on a database of AP articles that gets updated in real-time as new articles are published.