djray

Unrelated to the subject matter, but something that I never fail to appreciate is just how well-written the majority of Cloudflare's posts are. (There was an article on Merkle Tree Certificates which was phenomenal.) There is real skill in creating an informative work which condenses complex technical subject matter into something that a technically-inclined non-expert can understand.

Could someone knowledgeable please explain why an AI-focused CPU is superior to a GPU-based solution with an ARM CPU driving the high-level operations, please, especially when the RAM isn't on-chip?

I know that for GPU's the model weights have to be transferred over the bus initially, but that only has to occur once for inference use cases, so is the Fujitsu system more about training scenarios? Or is the focus more about efficiency, as these are ARM-based cores with AI additions?

They misspelled their own company name:

"...working together to bring Mistal’s scientific innovations to consumers around the world"

A quick proof-read wouldn't have gone amiss here.

I also fail to see how privacy is enhanced by sending private browsing data to a cloud service which then forwards it to an LLM, in addition to storing 'memories' and dredging through my browser history.

No, the model has answered correctly. Your question is poorly phrased (possibly deliberately).

Your question would correctly classify the user's input as requesting a human support agent, but at an indeterminate time.

If you wanted to determine whether the user wants to speak to a human support agent immediately, you would have to correctly qualify your question, e.g. "Does the user want to speak to a human support agent now?". You could have another question which is "Is the user requesting a call-back from a human support agent?". Or you could have a multiple choice query which would filter the conversation into one of a number of pre-written possibilities.

This is nothing to do with accuracy or hallucination. It's a different method of interacting with the model where you are relied upon to be precise.

Your default should be WCAG-compliant. That's the whole point! A visually-impaired user shouldn't have to customise your site for it to be readable. Your site will be failing some accessibility metrics until you do that, and that will affect your search ranking etc. Plus it's generally a good audience-friendly thing to do.

I was aware of the poll, but putting that up after the fact and it being only about the timestamp is a bad signal, too. It says that you're willing to pull the rug on whole sets of features at a whim and just because _you_ think it's the right thing to do. If you're trying to foster a community, I'd argue that this is not the best approach. These sorts of things could be reversible experiments - such as a banner at the top of the page saying "We're trying out removing timestamps and counters like likes! Click here to give us feedback or to switch back to the old UI." And then you keep track of everyone who switched back or re-enabled the feature you're fiddling with. Or you let people switch for a while and give their opinion after the test period.

Her playing at Glastonbury here in the UK only seems like the other day, but I just checked and it's been _12 years_! She played her most popular songs and the audience treated her like the star she was. What an incredible, talented, kind-hearted woman. An inspiration to so many for all the best reasons.

I agree with removing the obvious dopamine-related signals, but removing timestamps is a mistake, as they are vital for technical and news-based posts. They're a valuable relevance signal in many cases, not 'noise'. At least make them available on-hover or as a user option. Yes, you could try with a 'limit to historical posts' option, but that's going to be difficult as some topics move a lot faster than others, e.g. current affairs versus web standards.

This sort of unilateral action-before-asking is imo not a great signal to send out if you're trying to grow a platform. If you're going to run experiments, let people know first.

Could I also suggest making the comment text size larger (it's smaller than CSS 'small'), changing the typeface to be proportional, and most importantly making sure the contrast between foreground and background elements is WCAG-compliant? Most of the text is less than 3:1. Currently the site may be quite challenging to access for a visitor with visual impairments, and I think many such users would enjoy a site such as this if it were more accessible. Thanks.

"If the online check fails and you’re otherwise online, it gets revoked"

That's not being offline ;)

> If you are online, and the rights service is failing, it will revoke your key.

Do you have evidence of this? I've played Xbox for 25 years and never seen this.

This is false. You don't need an internet connection to play from disc:

https://support.xbox.com/en-US/help/hardware-network/connect...

You can play digital titles, too, as long as your console is set to "Home".

I played Forza Horizon 3 for days without internet a couple of months ago.

One would hope that you change minds and differentiate yourself through the strength of your argument, rather than the anfractuosity of your prose or a propensity for sesquipedalian verbiage.

Clarity does not equate to genericity. I am regularly struck by clear writing that is both lucid and powerful without resorting to ten-dollar words or tortured syntax.

The fact that AI text is often so easily spotted is a clear sign that so-called 'slop' is not the same as a writing style focused on brevity and understanding over florid meanderings.

This pertinent paper appeared more than a decade ago about the flaws in the amyloid plaque hypothesis: https://pmc.ncbi.nlm.nih.gov/articles/PMC4207354/

Many people without dementia show amyloid plaques in their brains in autopsies. It's becoming more accepted now that there are multiple interrelated causes after decades pursuing the simplistic amyloid plaque theory.

The article is bordering on irresponsible.

They aren't equal, as the faster version does an unconditional memory write instead of only writing to the array if the condition is satisfied. The compiler is strictly forbidden from turning a conditional write into an unconditional one.

Large organisations have advocacy, support and friendship groups for many underrepresented or protected groups, such as those for religions, mental health issues, physical disabilities etc. I imagine this is simply one of those internal group's organisational repos.

Just fucking use the language appropriate to the task you're trying to accomplish.

Just fucking use the language your colleagues can understand and support for the next few years.

Just fucking use the language with the framework and tooling you need to get your job done efficiently and effectively, and one at the appropriate level of abstraction for the project.

Just fucking use the language which AI agents can read and write well, because we're in the End Times and this stuff matters.

Just fucking use the language with great testing and CI/CD support because you'll be spending longer supporting your code than writing it.

The skill is choosing well, and a key realisation is that it's never a one-size-fits-all thing. That's why an article like this is less than helpful, gosh darn it.

"Skirmish" or "Conflict" or "Action". It's less illegal if you don't call it a war.

It says "approx. 5 hours" in the "Power & Batteries" section.

I said in my comment that it was GPU bound. Task Manager was reporting 100% GPU usage from the browser, which went down to 0% after closing the tab.

It looks to be that the demo has been fixed since I initially looked. Testing today showed much improved performance.

The demo runs at 5 FPS on my laptop (11th gen Core i5 and Iris Xe graphics, Chrome Latest as the browser, with the GPU being the bottleneck). I was hoping for something rather more efficient given the write-up saying it ran at 60 fps on mobile.

The maps are pretty, but the per-tile build constraints of the WFC build approach means that pretty unnatural generations end up happening because non-local influence is difficult to take into account. I think this may be OK for games where you discover tiles one at a time, but for a full map generator it's not great, and better solutions exist. Red Blob Games did a writeup of a noise-based method which looks superior imo. You can use moisture-tracking approaches for rivers, lay roads, bridges and other artificial elements in a separate pass, and it will likely end up faster and more robust. I think WFC is an interesting programming problem, though, so it was likely fun to implement.

Nonetheless, this was an excellent write-up and impressive demo.

This is quite insensitive and a poor analogy. Telling someone who's struggling to "Get your head in the game" is unhelpful in the extreme. It's like telling someone who's drowning to "Just swim better".

Free climbing is a high-risk pursuit. A free climber goes on occasional expeditions where they solo a route, but before then, they train with others, have spotters and partners during route practice, and whole communities to provide help, support and emergency contact should things go wrong. Alex Honnold doesn't just rock up to a mountain (if you'll forgive the pun) and solo it every day for months on end.

It's very difficult to just put everything aside and focus on a singular task for the amount of time it takes to complete a significant software project. Once the initial enthusiasm wanes, keeping going can be a real challenge when you're on your own. The OP's concerns are absolutely valid and deserve respect.

I think the key here is the ability to focus on the privacy-first nature of local LLMs. A cloud-based service will always be more powerful (and markedly so), but Apple is very cautious about pursuing cloud-based solutions when user data is involved - privacy is a selling point of their products, after all. This is a double-edged sword, as you get to sell your services as privacy-friendly, but your offerings can be significantly less capable than your competitors' (see the iOS messages summary debacle, for example). The advantage of waiting is that smaller AI models are becoming much more powerful all the time.

Of course, Microsoft is also at this with its Copilot programme for laptops, where an onboard Neural Processing Unit has to be a particular speed to qualify. This lets you do local AI things like content-aware image snipping, text summaries and...er, Recall.

As to whether Apple will come out of this looking good or not, I think they're currently regretting rolling out a shitty initial AI offering, and will get better with the next release. It'll be like Apple Maps. Or the butterfly keyboard. Or any number of other broken version 1 Apple things.

An interesting question is to whether Apple Intelligence can be cancelled or pared back now the landscape is so AI dominated, i.e. will the lack of AI offerings be seen as a competitive disadvantage, or are people so sick of AI by now that it isn't a factor.

The reality is more akin to me agreeing to have my car covered in your bumper stickers (and anyone else's, as long as they abide by my particular set of bumper sticker rules), and then having to remove some without notice at a later date because the chap at the gas station got offended and wouldn't serve me any more.

And there being only one gas station.

And the guy having not objected to the exact same bumper stickers for the last 15 years.

Visa have said this is because of 'enhanced risk' caused by this content, but they've been fine with it up to now. It's only because of the Australian group's censorious actions that they've decided to act. That's the frustrating thing, at least to me.

Whether they have the legal right to do it or not, it's still a dick move.

(If you'll forgive the pun.)

What you gain in performance, you somewhat sacrifice in flexibility, at least in comparison with OpenEXR.

OpenEXR was designed for modularity, allowing efficient access to individual layers or channels. This is crucial in VFX workflows where only specific passes (like normals or diffuse) might be needed at any one time. This access is possible because EXR stores channels separately and supports tiled or scanline-based access.

The custom compression method Aras proposes - using meshoptimizer on 16K pixel chunks, followed by zstd as a second compressor step - achieves significantly faster decompression and better compression speeds than EXR ZIP, HTJ2K, or JPEG-XL lossless. However, it trades off random access and requires decompressing the entire image at once, which increases memory usage. Individual frames for a VFX production can be multiple gigabytes (i.e. dozens of 32-bit layers at 4K resolution).

The author's proposal is still compelling, and I wonder if a variant could find its way into some sort of archival format.

There seems to be a mistaken thought that having an AI (or indeed someone else) help you achieve a task means you aren't learning anything. This is reductionist. I suggest instead that it's about degrees of autonomy. The person you're responding to made a choice to get the AI to help integrate a library. They chose NOT to have the AI edit the files itself; they rather spent time reading through the changes and understanding the integration points, and tweaking the code to make it their own. This is much different to vibe coding.

I do a similar loop with my use of AI - I will upload code to Gemini 2.5 Pro, talk through options and assumptions, and maybe get it to write some or all of the next step, or to try out different approaches to a refactor. Integrating any code back into the original source is never copy-and-paste, and that's where the learning is. For example, I added Dexie (a library/wrapper for accessing IndexedDB) to a browser extension project the other day, and the AI helped me get started with a minimal amount of initial knowledge, yet I learned a lot about Dexie and have been able to expand upon the code myself since. If I were on my own, I would probably have barrelled ahead and just used IndexedDB directly, resulting in a lot more boilerplate code and time spent doing busywork. It's this sort of friction reduction that I find most liberating about AI. Trying out a new library isn't a multi-hour slog; instead, you can sample it and possibly reject it as unsuitable almost immediately without having to waste a lot of time on R&D. In my case, I didn't learn 'raw' IndexedDB, but instead I got the job done with a library offering a more suitable level of abstraction, and saved hours in the process.

This isn't lazy or giving up the opportunity to learn, it's simply optimising your time.

The "not invented here" syndrome is something I kindly suggest you examine, as you may find you are actually limiting your own innovation by rejecting everything that you can't do yourself.

"To err is human, but to really foul things up you need a computer." -Paul Ehrlich

I think your project is at far too early a state to share. You are currently literally using 'fs::recursive_directory_iterator()' to search one file at a time on the selected drives. You aren't even using multi-threading for solid state volumes or caching the file system details during iteration, meaning every search is going to take several seconds on even modestly-sized SSDs. I can't believe this is faster than a comparable Windows 11 search, as you claim on your GitHub - I suspect the Windows search was for a non-indexed location and/or from a cold start, meaning the file system entries weren't cached already by the OS. (By the way, if you're going to make claims like this, you should publish reproducible benchmarks, so people can test your methodology.)

You also say that you want a solution more integrated with Windows, yet your UI is in SFML and your screenshots look nothing like a native Windows application.

Your project website doesn't render correctly on non-wide screens.

You are using a font which won't render many Unicode filenames correctly, e.g. Japanese or Chinese characters.

You have given no thought to accessibility or localisation.

You argue that Everything doesn't work with non-NTFS drives. That's incorrect - you can add them manually. See https://www.voidtools.com/faq/#how_do_i_index_a_fat_volume . You can also add Network Drives, NAS volumes etc.

Everything will run as non-administrator, but it will be unable to index NTFS properties. This is an OS-level restriction. You will face the same when you attempt to read these properties. The same goes for accessing protected folders - you just skip them.

You mention that you also want to feature file encryption and file transfer features in the same application. You are taking on too much here; each of the three features (file search, file transfer and file/folder encryption) is challenging and could take months to years to develop on its own. There are already accomplished apps satisfying these needs. Why would anyone use your solution over 7-zip?

You have no tests. If you want to feature encryption or file transfer, you will need an extensive test suite. Users will not be happy if your tool deletes or mangles their files.

Looking at your "Da-file-encrypter" repo, you don't seem to have any knowledge about how to actually encrypt files - you are doing an XOR against the bytes of the key, which is derived from the same bytes each time. Please NEVER ROLL YOUR OWN encryption strategy. Even the experts get it wrong. Instead, use a proper cryptography library. If you take anything away from this, please let it be that - never implement your own encryption.

Voidtools' Everything has an API and people have used that to integrate with launchers like PowerToys Run. It may be better to use the API and to add the features you think it is missing. It is very unlikely you will be able to beat its performance.

Judging by your licence and the Donate button on your website, you seem to have recently switched licences and want to make money out of this. In my personal opinion, that's not a good look for a project in this state.

These start from 450 euros.

To be clear, I meant the specific issues the OP was talking about in the article, not the overall inefficiencies. I was speaking to the fact that the author had not seemed to have raised the matters with anyone who was actually in a position to do something about it. I apologise for the lack of clarity.

The problem is that this is the first time any of us are hearing of these inefficiencies. Were these concerns raised with the author's state representative? The oversight committee? Why was the press not informed if lives were at risk? Why was pressure not heaped upon senior management and the powers that be to effect actual change years ago when the problems were first observed? Merely griping about how frustrating your job was categorically does no fucking good, and neither does writing about it after the fact. You may think you're fighting the good fight, but this isn't a grassroots protest if the only people you were complaining to were your manager and/or your colleagues.

Yes, inefficiency and bureaucracy suck. No, the answer is not to scrap it all without knowing what the hell you're doing and "just wing it", which is exactly what Musk and co are doing. Strangely enough, the solution is probably a compromise. Compromise takes two sides actually talking to each other, with people familiar with the matter present, and an acceptance that the goal is to make the process better without negatively impacting important things like security, safety and data integrity.

Please forgive the tone of this message, but I can't help wonder how many thousands of people are going to die because of the effective abolishment of USAID and other programmes.

The OP misunderstands (perhaps deliberately or for humorous effect) what a co-pilot is. This is telling:

"I learned that I need to stay firmly in the driver’s seat when tackling new tech."

Er, that's pretty much what a pilot is supposed to do! You can't (as yet) just give an AI free reign over your codebase and expect to come back later that day to discover a fully finished implementation. Maybe unless your prompt was "Make a snake game in Python". A pilot would be supervising their co-pilot at all times.

Comparing AIs to junior devs is getting tiresome. AIs like Claude and newer versions of ChatGPT have incredible knowledge bases. Yes, they do slip up, especially with esoteric matters where there are few authoritative (or several conflicting) sources, but the breadth of knowledge in and of itself is very valuable. As an anecdote, neither Claude nor ChatGPT were able to accurately answer a question I had about file operation flags yesterday, but when I said to ChatGPT that its answer wasn't correct, it apologised and said the Raymond Chen article it had sourced wasn't super clear about the particular combination I'd asked about. That's like having your own research assistant, not a headstrong overconfident junior dev. Yes, they make mistakes, but at least now they'll admit to them. This is a long way from a year or two ago.

In conclusion: don't use an AI as one of your primary sources of information for technology you're new to, especially if you're not double-checking its answers like a good pilot.