benkaiser

Often there are "abliterated" or "uncensored" tuned models that suppress the rejections. From my high level understanding it is performed by finding which weights activate for the rejection and lowering those so the model is less likely to reject. It doesn't fix if the model doesn't know what you're asking it though (i.e. if the model never actually learned about meth production in the first place).

For those looking for mobile support connecting to remote MCP servers, I maintain a source-available chat app that does exactly that (uses OpenRouter for inference): https://benkaiser.github.io/joey-mcp-client/

Would be nice if the MCP implemented the Streamable HTTP MCP spec instead of the CLI one. I know this is already a HTTP API, but making it available as an MCP server that clients like Joey[1] can consume easily over network would be nice.

[1] https://github.com/benkaiser/joey-mcp-client

One key aspect that is missed here, is mobile users.

iOS don't have a CLI, Android phones... kinda have a CLI, but it's not really usable by all apps / universal like it is on desktop.

I've been putting together a MCP client for mobile recently [1] since to me it seems apparent that Remote MCP is a great utility but none of the major players offer it without a paid subscription.

As other commenters have mentioned, the use case here is really not for software developers, it's for everyone else. It's bringing that agentic loop to do all those "personal assistant" type things in a connected way with all of your services. We're not there yet, but maybe some day.

[1] https://github.com/benkaiser/joey-mcp-client

Am a software engineer at Microsoft using a M3 MBP, opinions are my own and all. Honestly one (of many) reasons I opted to go through the exception process to request a macbook was the screen brightness. The fact you can run software to boost the screen to HDR brightness levels for SDR content is insanely useful for working outside.

I think the parent may be referring to the fact that safari/webkit will evict all localstorage/indexeddb/caches etc after 7 days of not visiting a site. And apparently this now extends to PWAs making it a pretty big blog to building any infrequently accessed PWA that needs to persist user data locally.

Counter-example... some people have no idea that SLS is causing these issues and just chalk it up to how things are.

I recently switched to an SLS free toothpaste but it was really hard to track one down that still had flouride in it!

This is super nice as it's a native unencumbered linux environment.

I recently blogged about trying to do web development on the Quest 3, and although it kind of works, it's way more hacky (and performance is still lacking). The nice thing however about the Quest, is still being able to leverage virtual window placement in the space around you vs. the fixed 2d monitor approach here.

https://benkaiser.dev/web-development-in-vr/

I actually tried doing web dev on a Meta Quest 3 recently, but found the biggest limitation to actually be the lack of devtools in Android.

Here is the blog I did discussing the limitations: https://benkaiser.dev/web-development-in-vr/

I wonder if something like this running on the quest could technically work, but I suspect it would be too heavy running Linux chrome in a chroot. You also lose the cool "place and resize your windows anywhere" if it's all stuck inside one window for a desktop.

I'm quite curious how they go about licensing the content, maybe they just pay the creators a cut of the subscription fees. Or is it really just scraping youtube in some form?

As for the search you mentioned, that might come into it for an older range, for my little ones they still can't read or spell yet, they just want to click on the thumbnail that looks the most engaging at any random time.

On this same topic, I just launched an app that lets you use your offline videos in an interface that is easy to use for kids.

https://play.google.com/store/apps/details?id=com.kaiserapps...

I stuck to a one-off payment, rather than the garbage subscription models all the other parenting apps use.

Thanks for sharing, I wasn't aware of that blog post. That said, their approach sounds kind of disheartening. I love being able to use extensions within my PWAs (primarily uBlock Origin, obviously), and some Android Chrome forks (Kiwi/Mises) let me do this, while still letting me feel like I'm in a dedicated app (i.e. no browser chrome at all). The firefox team really seem to stress here that they want to keep the chrome around the app (albeit different), which really feels like it goes against the grain of what I expect a PWA to be in the first place (a chromeless website).

As a developer, the one feature I really love in Chrome is PWAs. But Firefox abandoned PWA support years ago, and seems to have no appetite for adding PWAs back[1]. Maybe I'll just have to split my usage across PWAs in Chrome (since I trust those apps/websites anyway) and Firefox for general browsing.

[1] https://connect.mozilla.org/t5/ideas/bring-back-pwa-progress...

Brightintosh is the same as Vivid (roughly) and is actually open source, but the author ships the binaries as a Mac app store purchase for $2:

https://github.com/niklasr22/BrightIntosh

Yeah there's a way to do this with Calibre + some plugins on windows. I tried on my Mac but it was unsuccessful.

Not the best reporting, but still interesting (although not surprising) that the Australian gov is cautious about usage of DeepSeek.

Runs on an AMD 7900 XTX at about ~20 tokens per second using LM Studio + Vulkan.

This is not to far from what AdNauseam does today as a simple browser extension.

https://adnauseam.io/

Absolutely.

I recently read "The Life You Can Save" by Peter Singer, and it really does a great job of making the case for generosity even amongst middle class 1st world individuals.

ebook/audiobook are free from their website: https://www.thelifeyoucansave.org.au/the-book/

To add a little context here, I (the author) understand LLMs aren't the right tool (at least by themselves) for verbatim verse recall. The trigger for me doing the tests was seeing other people in my circles blindly trusting that ChatGPT was outputting verses accurately. My background allowed me to understand why that is a sketchy thing to do, but many people do not, so I wanted to see how worried we really should be.

I'm talking about it mostly from an entropy standpoint here. With larger size, you can represent more information.

Think of it like how you have a JPEG, you can compress it further and further, but eventually you lose the ability to understand the original image.

With the models, if they had infinite size, I imagine they could recall values in their training data extremely accurately. But as you compress down further and further to smaller and smaller models, you are trying to distil the same amount of information in less space, and so things cannot be perfectly recalled.

We have people tuning these smaller models to squeeze every ounce of what we consider meaningful out of them (passing certain benchmarks, seeming coherent in dialogue, etc), but in the process the things we don't tune them for, i.e. accurate recall of scripture (not that we should) they lose that ability.

I will say with all of that though, that I only have a high level understanding of LLMs, I've integrated them into products on the job, but I am by no means an ML engineer.

I can speak to a couple of perspectives I have seen other people use it for, ranging from valid to somewhat scary.

1. Preparing a sermon for Church, I don't advocate for this, but it's definitely being done out there. Here, the pastor may know the topic they are speaking on, but want the LLM to help them plan out the message and structure it. 2. Preparing lesson plans for Sunday School. This seems reasonably fine to me, but I would still err on the side of not trusting the raw scriptures output as evidence, and instead look them up separately before reading them out.

The above examples may particularly come into play when English is not a first language, since although they can understand and express their faith easily in their native language, ChatGPT can help them represent it in English well.

Personally, I think the use cases are many, but mostly for discussion / personal reflection. These include things like asking for perspectives that other Christians take on certain passages, helping understand how some scriptures link to other scriptures in the Bible, and sometimes even exploring some of the history of the Christian faith through the last ~2 millennia since it was written. Anything meaningful you can manually research further / reference before taking it at face value, but it can work as a great starting point for your search.

I guess the challenge is that the parameters have to encode much more intricacy than just the Bible. Even if one were produced purely with text from the bible, it would likely not be able to converse as well as conversationally tuned ones behave. Perhaps there's a middle ground of a fine-tuned LLM on scripture recall + o1-style background reasoning to produce the best output. Or even just a RAG.

Hence why I marked those cases with warnings, indicating they were accurate if you give leniency for different translations.

Some services may overlay this functionality (e.g. Bing), but in the article I'm making direct LLM calls without any external function calling.

Hah, I did not. Thanks

Fast, easy, less likely to run afoul of copyright.

I agree with the others that ES5 is likely where it came from, but my mind also jumped to The Jackson 5.

It sounds righteous for sure, but the implementation sounds terrible.

Let's suppose kids do get forced off the mainstream social networks with Australia legal entities (Meta, Goog, Bytedance, X?). What stops them from joining fringe social networks operated outside of Australia with even less oversight? Surely nothing bad could happen between kids on those networks.

Yes, the Australian government could DNS block them like they do torrent sites. But it really wouldn't be beyond teenagers to inform their friends about how to change DNS servers...

I can actually weigh in here. Working from Australia for another team inside Microsoft with a large monorepo on Azure devops. I pretty much cannot do a full (unshallow) clone of our repo because Azure devops cloning gets nowhere close to saturating my gigabit wired connection, and eventually due to the sheer time it takes cloning something will hang up on either my end of the Azure devops end to the point I would just give up.

Thankfully, we do our work almost entirely in shallow clones inside codespaces so it's not a big deal. I hope the problems presented in the 1JS repro from this blog post are causing similar size blowout in our repo and can be fixed.