Posts by AdamWill 1727 publicly visible posts • joined 4 Nov 2010
I love the examples. I don't call support to ask about my package if it's out for delivery; I'll already have got an email telling me it's out for delivery. Every competent company had that set up a decade ago. I call support to ask about my package if it said it was out for delivery two days ago but I never got it. Or it hasn't moved from Fairbanks, Alaska for the last three days. Or it says it was delivered but I don't have it. You know, all the stuff it turns out a chatbot can't actually handle, and has to go to a human instead.
Similarly, I would love to meet the company that would trust a chatbot to issue a refund to a customer with no human signoff. I bet there aren't any. I bet OpenAI doesn't.
this article has several hundred/thousand words in it and the sole bit of meat is:
"And that's when I became a Quick Desktop convert: it flagged an email buried forty messages deep in my inbox that I'd mentally filed under "dealt with" - but very much was not. My own inbox had given up on me like everyone who's ever tried to love me, but Quick Desktop hadn't."
That's *it*. The lot. It found one email. Great?
Jeez, what a waste of space.
No, not really. all forges have some kind of workflow automation system and they are all vulnerable to exploitation to some degree.
Forgejo has Forgejo Actions, which mostly copies github actions' interface and can similarly be subject to attack, if you configure vulnerable workflows or unwittingly merge a PR that contains one.
On most forges (including both github and Forgejo) you can disable actions (or whatever it's called) entirely for a repo. But then, people use it because it's useful, not just for giggles. It's the old security/convenience tradeoff again.
Sorry, new projects is also now a junk metric.
I was cleaning up an agentic AI mess this morning, see https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.org/thread/SFVETHOYKQAO7KKLEXCK4IBT4WVPRE6F/ . One of the many artifacts of that system that I came across is https://github.com/nathan9513-aps/whiskered-wise - an entire project apparently hallucinated by whatever system this was (I'm guessing OpenClaw). It created it, inundated itself with weird PRs, triggered other agentic systems to do nonsensical actions on the PRs, and merged them. What does the project *do*? Who knows, but the AI reckons it's very important to a (probably non-existent) barber shop in Marrakech:
https://github.com/nathan9513-aps/whiskered-wise/commit/67107b6c3d9edb5256444153ab9e3b8472d16f33
what a world.
"Our government is as frustrated as you are by the fragmentation and silos, the complexity, the status-quo thinking and the dangerously slow take up of digital and AI technologies,” she added"
I doubt anyone in New Zealand is frustrated by any of those things. They might be frustrated by governments constantly starting quarter-baked IT modernization projects that inevitably cost huge amounts of money and fail, though.
Updates fixing this for Fedora went stable yesterday. F42 - https://bodhi.fedoraproject.org/updates/FEDORA-2026-ec1c523fdb , F43 - https://bodhi.fedoraproject.org/updates/FEDORA-2026-cccb681166 , F44 - https://bodhi.fedoraproject.org/updates/FEDORA-2026-4462efc052 .
I preferred Ed's take on this. It's snappier.
"For example, Kevin O’Leary — known as “Mr. Dogshit” to his friends — is allegedly building a 9GW data center in Utah, but he may as well say that he’s building a unicorn that shits Toyota Tacomas, as doing so is far more realistic than a project that will likely cost $396 billion, assuming that locals and bankers don’t drag him to The Other Side like Dr. Facilier."
https://www.wheresyoured.at/where-are-all-the-data-centers/
We (RH) were working on this under embargo till some idiot disclosed it prematurely. Unfortunately, they did so while we had an s390x builder outage, which rather complicated matters in terms of building fixes. That's resolved now.
Updated kernels are now stable for all Fedora branches. RHEL is coming, the RHEL folks are working on it as fast as they can. I believe we're recommending the mitigation suggested by Hyunwoo Kim until then, the one mentioned above:
sh -c "printf 'install esp4 /bin/false\ninstall esp6 /bin/false\ninstall rxrpc /bin/false\n' > /etc/modprobe.d/dirtyfrag.conf; rmmod esp4 esp6 rxrpc 2>/dev/null; echo 3 > /proc/sys/vm/drop_caches; true"
the nice thing about it is it doesn't require a reboot.
Not having the modules currently loaded does not render you immune. The exploit causes them to be loaded.
[adamw@vaioz dirtyfrag (master %)]$ lsmod | egrep "(esp4|esp6|rxrpc)"
[adamw@vaioz dirtyfrag (master %)]$ ./exp
[root@vaioz dirtyfrag]# lsmod | egrep "(esp4|esp6|rxrpc)"
esp4 32768 0
blocking the modules from *ever* being loaded prevents the exploit.
That's not what they're admitting, though it's not totally clear unless you go read the original blog post.
They're not saying the problems are caused by *GitHub* making "a rapid change in how software is being built". They're saying the problems are caused by *other people* doing that, and slapping the results on GitHub. There are some pretty gnarly graphs in the blog post showing the massive acceleration in repo, commit and PR creation.
I think anyone who's trying to be a sysadmin in the current age could sympathize to a degree, TBF. Just trying to keep a simple web service up under the weight of scraper spam is a challenge. I would not be particularly keen to volunteer to run an open-access git forge right now. (Of course, they did volunteer for the job, and their parent company is one of the primary drivers of AI hype, soooo.)
I don't fault you or Liam at all for not finding it - you're definitely not the only ones - but it *is* possible to do custom storage configuration in webUI. On the 'Installation Method' page, click on the 'kebab menu' - three dots in a vertical line - in the top-right corner, and click 'Launch storage editor'. That's the custom storage thing.
Why is it there? Eh, it's a long story. Roughly, the *original* plan for webUI was not to have it, which is why it's not in the main UI flow. The idea was that, if you wanted custom storage, you could configure it outside the installer using whatever tool you like, then just assign the mount points in the installer; that's why one of the options that *are* obvious on the 'Installation Method' page is assigning mount points. But for various reasons it was eventually concluded this wasn't enough, and the storage editor got retrofitted in - it's actually Cockpit's storage module, customized somewhat to the installer environment. But it was sort of stuck in a corner I guess because it was a later addition and still not really part of the "intended design'. I guess we could talk about making it more obvious somehow, because a lot of people do miss it.
The bit about RH applies to RHEL / CentOS. All Fedora branches are already fixed. This was fixed upstream in 6.19.12, which has been stable in F42 and F43 for two weeks. F44 has 6.19.10 in the shipped images, but 6.19.14 is available as a 0-day update (i.e. it is available as an update immediately after install).
"As an example, he said that Citi Sky can take market moving announcements from Washington DC and deliver a simultaneous portfolio review to all of their clients based on unfolding events."
This is exactly how you're *not* supposed to invest, isn't it? Panic responses to breaking news announcements?
"How would this be verified?"
It won't be. The relevant laws (at least the ones I've seen) don't require anyone to verify it.
The scenario the legislation (at least the California one) envisages is parental control. A parent sets up an account for a child and enters the child's date of birth. They enter it accurately because they *want* age controls to be applied to their child's usage. The whole mechanism is designed around the parent-as-sysadmin, effectively. There's no envisaged external/government verification of the DoB.
If you are your own sysadmin, or your sysadmin doesn't want age restrictions applied to you, you enter whatever the heck you like. It seems pretty clear by implication that the law thinks this is fine.
"In the meantime, though, there's another issue. These kinds of laws don't work. They've never worked. Prohibition failed in the United States. I kept reading Playboy"
They *did* work, though. For a law to "work", it does not have to prevent every incidence of the act it prohibits. By that standard, *every* law fails. People get murdered every day; should we repeal the law against murder?
Age restrictions "work" in several ways. Primarily, they *significantly reduce* the incidence of the prohibited behavior in the prohibited group. Yes, you kept reading Playboy, but it probably wasn't *easy*, right? You had to sneak it from a parent or get it from a friend or try to blag your way through buying a copy. or whatever. Your (and my) access to age-restricted porn when we were young was hugely different from the present day, where anyone can load an internet full of porn with no practical restriction whatsoever (it's quite easy to find it *by mistake*).
An important consequence of this is they make *not* doing the thing socially acceptable (and hence normal). This is the point of social media restrictions. Nobody is expecting them to keep every 15 year old off Instagram. But they do at least reframe things so that you're not The Weird Outcast if you're 15 and *not* on Instagram.
I don't think these are good laws, in various ways, but "age restrictions never work" is a very bad argument against them.
Unimproveable comment from The Verge: https://www.theverge.com/ai-artificial-intelligence/883243/anthropic-claude-deepseek-china-ai-distillation?commentID=7abca920-4830-4bb7-9a84-b00b7f54f5f7
"Much easier to point the finger and scream "Hey, they're stealing my theft machine!" than admit they're getting lapped, I guess."
"The bot, designated MJ Rathbun or crabby rathbun (its GitHub account name), apparently attempted to change Shambaugh's mind by publicly criticizing him in a now-removed blog post that the automated software appears to have generated and posted to its website. We say "apparently" because it's also possible that the human who created the agent wrote the post themselves, or prompted an AI tool to write the post, and made it look like it the bot constructed it on its own."
It looked to me like the human set things up so the bot would go out, scan github for issues it thought it could fix, send PRs, *then blog about it*. The blog had existed for several days before the Matplotlib Incident, and the second post - https://crabby-rathbun.github.io/mjrathbun-website/blog/posts/2026-02-09-post.html - looks a lot like a prompt the model was supposed to fill in.
I suspect the human set the blog up and gave the bot credentials to post to it, and instructed it to post every day (or more often) about the PRs it had sent that day. It's not like the bot just *independently decided* to set up a blog and write a hit piece on it. The hit piece is what it came up with when following the 'write a blog post about what happened today' instruction.
"So did AI consulting really blow up, or did salespeople realize they could get a kicker on their quota fulfillment for tagging every possible consulting gig as AI? Not that I'm suggesting that IBM would ever do such a thing, and certainly not that it has ever done so before."
Y'know, I started wondering the same thing. I'm just going to leave this little table I drew up right here (you might have to cut and paste it into something monospace, though):
AI book Consulting Software AI cons AI soft
Q1 2024 >1b 5.186b 5.899b ~0.75b ~0.25b
Q2 2024 >2b 5.179b 6.739b ~0.8b ~0.2b
Q3 2024 >3b 5.152b 6.524b ~0.8b ~0.2b
Q4 2024 >5b 5.175b 7.924b ~1.6b ~0.4b
Q1 2025 >6b 5.068b 6.336b ~0.8b ~0.2b
Q2 2025 >7.5b 5.314b 7.387b ~1.2b ~0.3b
Q3 2025 >9.5b 5.324b 7.209b ~1.6b ~0.4b
Q4 2025 >12.5b 5.349b 9.031b ~2.4b ~0.6b
IBM's said each quarter that the "AI book of business" is 80% consulting, 20% software. Over the period it's been talking about it (last two years), consulting revenue has been basically flat, and software revenue has been growing far *faster* than 20% of the "AI book of business" could possibly explain.
So...yeah, my conclusion is that the "AI book of business" is basically just a label slapped on ongoing consulting revenue. To be fair, in https://cloudwars.com/ai/ibm-genai-business-soars-to-7-5-billion-drives-growth-in-mainframes-consulting/ , the CEO did more or less admit this when asked on a call:
"If I look at consulting, yes, there is a big piece of the AI book of business that is coming because people are directing their dollars towards that kind of consulting as opposed to alternate forms of consulting"
"Ali Sarrafi, CEO and co-founder of Kovant, an enterprise agent platform, told The Register in an interview that the problem with the way people use AI is that they see it as a form of fancy workflow automation
...
"We're working with this big large manufacturing company," Sarrafi explained. "They have about 7,000 suppliers. And every single time they needed to restock something, they had to coordinate with so many suppliers. It's actually the most boring job ever for everyone. But then they deploy this agent worker or a team of agent workers that basically monitors the stock levels. As soon as it goes below the forecast requirements level, it sends a preliminary email to the supplier saying, 'Can you tell us if you can supply this and what price?'""
How, exactly, is that not "fancy workflow automation"? And why the hell does it need AI? AI sounds like a rather terrible way to do it, in fact, since it will probably do what it's intended to do 95% of the time, then suddenly order five zillion widgets or email the supplier a blackmail photo instead of an order. And they'll have no idea why it did that.
For an extremely predictable workflow like this, what you need is a sensible bit of determinitive code to automate it. Not an LLM.
MCPs themselves aren't LLMs, though. They're deterministic code. The flaws here were exactly what the OP said - a failure to sanitize user inputs to code with potentially-dangerous access (to your filesystem and git repos).
The MCPs were *supposed* to be designed so that no matter what the prompt contained, you could contain the execution of anything via the MCP to a specific repository. But they were very badly implemented so they didn't do that.
"They said the volume of scheme members attempting to set up logins and view pension information "just to check" whether their details were correct was a contributing factor. Members have been advised to delay accessing the new portal until early spring when annual statements are due, unless they have a particular issue they wish to resolve or query."
How are they supposed to know whether they have a particular issue to resolve or query if the system can't cope with them logging in to check?
"Vibe coding will deliver a problematic proliferation of personalized software"
The problem I'm seeing with this, so far, is that, yes, people who don't really know what they're doing can vibe code a thing that mostly works!
So...they do.
They *all* do.
Now instead of having three or four tools for doing the thing, each maintained by someone who vaguely knows what they're doing and in which effort to do it well can be centralized, there are 50 tools to do it, all built by people who don't really know what they're doing and have no concept of how to maintain software in the long term, each with their own bugs and (no doubt) security issues.
Can you imagine having to sort this mess out in a few years when companies start realizing there are tens of thousands of vibe coded scripts doing stuff with their data and they have no conception of exactly how many, who is using them to do what, or what's wrong with each of them?
yikes.
Not to mention that ChatGPT is not available in China, so you can drop about a billion from the total.
3 is 44% of 6.82, so if we make that 5.82, they're projecting over half the adult population of the world that *can* access ChatGPT will be using it. (And that's *also* assuming sanctions on Russia are lifted by then). And 10% of those will be subscribing to a paid plan.
Let's...let's see how that goes.
