That’s an excellent summary of the product.

Armin Ronacher, who is an experienced software dev with a fair amount of open and less open source projects under his belt, was up until fairly recently a keen user of llm coding tools. (he’s also the founder of “earendil”, a pro-ai software pbc, and any company with a name from tolkien’s legendarium deserves suspicion these days)

His faith in ai seems to have taken bit of a knock lately: https://lucumr.pocoo.org/2026/1/18/agent-psychosis/

He’s not using psychosis in the sense of people who have actually developed serious mental health issues as a result of chatbot use, but software developers who seem to have lost touch with what they were originally trying to and just kind a roll around in the slop, mistaking it for productivity.

When Peter first got me hooked on Claude, I did not sleep. I spent two months excessively prompting the thing and wasting tokens. I ended up building and building and creating a ton of tools I did not end up using much. “You can just do things” was what was on my mind all the time but it took quite a bit longer to realize that just because you can, you might not want to. It became so easy to build something and in comparison it became much harder to actually use it or polish it. Quite a few of the tools I built I felt really great about, just to realize that I did not actually use them or they did not end up working as I thought they would.

You feel productive, you feel like everything is amazing, and if you hang out just with people that are into that stuff too, without any checks, you go deeper and deeper into the belief that this all makes perfect sense. You can build entire projects without any real reality check. But it’s decoupled from any external validation. For as long as nobody looks under the hood, you’re good. But when an outsider first pokes at it, it looks pretty crazy.

He’s still pro-ai, and seems to be vaguely hoping that improvements in tooling and dev culture will help stem the tide of worthless slop prs that are drowning every large open source project out there, but he has no actual idea if any of that can or will happen (which it won’t, of course, but faith takes a while to fade).

As always though, the first step is to realise you have a problem.

I’ve thought about jolla, but I’m not particularly interested right now. Their security is unlikely to be anything like as good as ios or graphene, software availability is poor, the hardware quality appears to be ok at best, and so on.

I’m considering various alternative devices, but if it’s effectively a “vanilla smartphone only slightly worse” it doesn’t really appeal to me. If they’d built a modern n900, on the other hand…

This is fun: a zero-click android exploit that allows arbitrary code execution and privilege escalation. Y’know, the worst kind. How did we get here?

Over the past few years, several AI-powered features have been added to mobile phones that allow users to better search and understand their messages. One effect of this change is increased 0-click attack surface, as efficient analysis often requires message media to be decoded before the message is opened by the user. One such feature is audio transcription. Incoming SMS and RCS audio attachments received by Google Messages are now automatically decoded with no user interaction. As a result, audio decoders are now in the 0-click attack surface of most Android phones.

AI, making everything worse, even before it runs!

https://projectzero.google/2026/01/pixel-0-click-part-1.html

Every now and then, I think about going back to android, and then I read stuff like this. FWIW, iOS had a closely related bug, but compiled the offending code with bounds checks, so it wasn’t usefully exploitable (and required some user interaction, too).

Anyway, if you do android, maybe check if automatic transcription is enabled.

Blacksky has delivered on bluesky’s promise of federation by setting up their own app view, creating a complete and independent third party implementation.

https://blacksky.community/profile/did:plc:w4xbfzo7kqfes5zb7r6qv3rw/post/3mcozwdhjos2b

Mcc has an interesting thread on mastodon (https://mastodon.social/@mcc/115918042095581428) which asks a bunch of questions about what the actual consequences of this might be, and no-one really seems to know, but no-one has much faith in the engineering or moderation chops of the bluesky team.

It looks like bluesky is somewhat vulnerable to rich trolls, because the main barrier to entry is cost… blacksky has budget of maybe 80000 usd/year (https://opencollective.com/blacksky) which is well within the reach of a whole bunch of people prepared to spend money to be egregious assholes, especially if they already have access to suitable talent and equipment. It’ll be bleakly interesting to see who tries this first.

A fun little software exercise with no real world uses at all: https://drewmayo.com/1000-words/about.html

Turns out that if you stuff the right shaped bytes into png image tEXt chunks (which don’t get compressed), the base64 encoded form of that image has sections that look like human readable text.

What are the implications?

Nothing! This was just for fun after a discussion with a colleague whether it might be even possible to make base64 blobs look readable. There’s certainly no poorly coded systems out there which might be hooked up to read emails or webpages and interpret any text they see as information.

No siree I’m sure everyone is keeping the attachments and the content well and truly isolated from each other and this couldn’t possibly do anything other than be a fun proof of concept and excuse for me to play with wasm.

I feel like one day that “no guarantee of merchantability or fitness for any particular purpose” thing will have to give.

There’s room for some nuance there. They make some reasonable predictions, like chatbot use seems likely to enter the dsm as a contributing factor for psychosis, and they’re all experience systems programmers who immediately shot down Willison when he said that an llm-generated device driver would be fine, because device drivers either obviously work or obviously don’t, but then fall foul of the old gell-mann amnesia problem.

Certainly, their past episodes have been good, and the back catalogue stretches back quite some time, but I’m not particularly interested in that sort of discussion here.

My gloomy prediction is that (b) is the way things will go, at least in part because there are fewer meaningful consequences for producing awful software, and if you started from something that was basically ok it’ll take longer for you to fail.

Startups will be slopcoded and fail quick, or be human coded but will struggle to distinguish themselves well enough to get customers and investment, especially after the ai bubble pops and we get a global recession.

The problems will eventually work themselves out of the system one way or another, because people would like things that aren’t complete garbage and will eventually discover how to make and/or buy them, but it could take years for the current damage to go away.

I don’t like being a doomer, but it is hard to be optimistic about the sector right now.

If that won’t sell it to governments around the world, I don’t know what will. Elon’s on to a winner with that strategy.

Ugh, I carried to listening to the episode in the hopes it might get better, but it didn’t deliver.

I don’t understand how people can say, with a straight face, that ai isn’t coming for your job and it is just going to make everyone more productive. Even if you ignore all the externalities of providing llm services (which is a pretty serious thing to ignore), have they not noticed the vast sweeping layoffs in the tech industry alone, let alone the damage to other sectors? They seem to be aware that the promise of the bubble is that agi will replace human labour, but seem not to think any harder about that.

Also, Willison thinks that a world without work would be awful, and that people need work to give their lives meaning and purpose and bruh. I cannot even.

Been listening to the latest oxide and friends podcast (predictions 2026), and ugh, so much incoherent ai boosting.

They’re an interesting company doing interesting things with a lot of very capable and clever engineers, but every year the ai enthusiasm ramps up, to the point where it seems like they’re not even listening to the things they’re saying and how they’re a little bit contradictory… “everyone will be a 10x vibe coder” and “everything will be made with some level of llm assistance in the near future” vs “no-one should be letting llms access anything where they could be doing permanent damage” and “there’s so much worthless slop in crates.io”. There’s enthusing over llm law firms, without any awareness of the recent robin ai collapse. Talk of llms generating their own programming language that isn’t readily human readable but is somehow more convenient for llms to extrude, but also talking about the need for more human review of vibe code. Simon Willison is there.

I feel like there’s a certain kind of very smart and capable vibe coder who really cannot imagine how people can and are using these tools to avoid having to think or do anything, and aren’t considering what an absolute disaster this is for everything and everyone.

Anyway, I can recommend skipping this episode and only bothering with the technical or more business oriented ones, which are often pretty good.

That’s horrifying. The whole thing reads like an over-elaborate joke poking fun at vibe-coders.

It’s like someone looked at the javascript ecosystem of tools and libraries and thought that it was great but far too conservative and cautious and excessively engineered. (fwiw, yegge kinda predicted the rise of javascript back in the day… he’s had some good thoughts on the software industry, but I don’t think this latest is one of them)

So now we have some kind of meta-vibe-coding where someone gets to play at being a project manager whilst inventing cutesy names and torching huge sums of money… but to what end?

Aside from just keeping Gas Town on the rails, probably the hardest problem is keeping it fed. It churns through implementation plans so quickly that you have to do a LOT of design and planning to keep the engine fed.

Apart from a “haha, turns out vide coding isn’t vibe engineering” (because I suspect that “design” and “plan” just mean “write more prompts and hope for the best”) I have to ask again: to what end? what is being accomplished here? Where are the great works of agentic vibe coding? This whole thing just seems like it could have been avoided by giving steve a copy of factorio or something, and still generated as many valuable results.

How about some quantum sneering instead of ai for a change?

They keep calling it a ‘processor,’ but it’s actually a refrigerated probability sculpture they beg to act like it Is a NAND gate for just half a microsecond

“Refrigerated probability sculpture” is outstanding.

Photo is from the recent CCC, but I can’t find where I found the image, sorry.

A lot of the money behind lean is from microsoft, so a push for more llm integration is depressing but unsurprising.

Turns out though that llms might actually be ok for generating some kinds of mathematical proofs so long as you’ve formally specified the problem and have a mechanical way to verify the solution (which is where lean comes in). I don’t think any other problem domain that llms have been used in is like that, so successes here can’t be applied elsewhere. I also suspect that a much, uh, leaner specialist model would do just as good a job there. As always, llms are overkill that can only be used when someone else is subsidising them.

This is a fun read: https://nesbitt.io/2025/12/27/how-to-ruin-all-of-package-management.html

Starts out strong:

Prediction markets are supposed to be hard to manipulate because manipulation is expensive and the market corrects. This assumes you can’t cheaply manufacture the underlying reality. In package management, you can. The entire npm registry runs on trust and free API calls.

And ends well, too.

The difference is that humans might notice something feels off. A developer might pause at a package with 10,000 stars but three commits and no issues. An AI agent running npm install won’t hesitate. It’s pattern-matching, not evaluating.

I suppose it was inevitable that the insufferable idiocy that software folk inflict on other fields would eventually be turned against their own kind.

https://xkcd.com/1831/

Sunday afternoon slack period entertainment: image generation prompt “engineers” getting all wound up about people stealing their prompts and styles and passing off hard work as their own. Who would do such a thing?

https://bsky.app/profile/arif.bsky.social/post/3mahhivnmnk23

@Artedeingenio

Never do this: Passing off someone else’s work as your own.

This Grok Imagine effect with the day-to-night transition was created by me — and I’m pretty sure that person knows it. To make things worse, their copy has more impressions than my original post.

Not cool 👎

Ahh, sweet schadenfreude.

I wonder if they’ve considered that it might actually be possible to get a reasonable imitation of their original prompt by using an llm to describe the generated image, and just tack on “more photorealistic, bigger boobies” to win at imagine generation.

Hah! Well found. I do recall hearing about another simulated vendor experiment (that also failed) but not actual dog-fooding. Looks like the big upgrade the wsj reported on was the secondary “seymour cash” 🙄 chatbot bolted on the side… the main chatbot was still claude v3.7, but maybe they’d prompted it harder and called that an upgrade.

I wonder if anthropic trialled that in house, and none of them were smart enough to break it, and that’s what lead to the external trial.

So, I’m taking this one with a pinch of salt, but it is entertaining: “We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.”

The whole exercise was clearly totally pointless and didn’t solve anything that needed solving (like every other “ai” project, i guess) but it does give a small but interesting window into the mindset of people who have only one shitty tool and are trying to make it do everything. Your chatbot is too easily lead astray? Use another chatbot to keep it in line! Honestly, I thought they were already doing this… I guess it was just to expensive or something, but now the price/desperation curves have intersected

Anthropic had already run into many of the same problems with Claudius internally so it created v2, powered by a better model, Sonnet 4.5. It also introduced a new AI boss: Seymour Cash, a separate CEO bot programmed to keep Claudius in line. So after a week, we were ready for the sequel.

Just one more chatbot, bro. Then prompt injection will become impossible. Just one more chatbot. I swear.

Anthropic and Andon said Claudius might have unraveled because its context window filled up. As more instructions, conversations and history piled in, the model had more to retain—making it easier to lose track of goals, priorities and guardrails. Graham also said the model used in the Claudius experiment has fewer guardrails than those deployed to Anthropic’s Claude users.

Sorry, I meant just one more guardrail. And another ten thousand tokens capacity in the context window. That’ll fix it forever.

https://archive.is/CBqFs