This just seems to me like (1) not wanting to learn new things, (2) a logical fallacy because the conclusion that those changes are to "stay relevant" and not to make the life of any future mantainer easier, and (3) fear of the new (https://en.wikipedia.org/wiki/Neophobia)

For point based train control systems, operation of a train is only enforced at every signal (and on the way in approach at specific points), this means e.g. that a train does not run over a red signal, does not drive faster than the maximum speed, and has acknowledged speed limits.

For line based train control systems, that supervising happens continuously over the whole track.

Since the physical signals and sensors are expensive in installation and maintenance and railway capacity depends on the number of blocks on a track, line based train control systems support virtual blocks. This means that the signals and sensors are not physically there, but simulated by computers (on the train and signalbox). From the view of the point based train control that means multiple trains are in a single block, because it does not know about the virtual subblocks. When a train is running that only supports point based train control it can only use whole blocks, so the capacity is simply reduced, but it is still safe, because every physical block is also known in the line based control system.

When the line based train control system stops being in use, the fundamental constraint of the point based train control system is being violated: there are multiple trains that are in a single block, and hence not actually protected by a red signal in between. This means that you must ignore the signals until every train has crossed at least one physical signal.

(This is speculation:) It seems that that the operators assumed that there is still a safety system in place, that they can rely on. Hence they did no tell the trains to use manual operation, (which might also require driving by sight [max. 40km/h], but not necessarily, I don't know the proper procedure in Spain), but told them to fall back to the normal safety system. It's reported that some train drivers and dispatchers questioned that, but that can also be an translation error and instead them repeating what was said, as is proper safety procedure. But the point based train control system will show a green signal, even if there is actually a train in front, because it is in the same block as you, but the signal shows whether you are allowed to drive in the next block, not into the current, as you are already in the current block. This seemed to be what happened here, the second train saw the green signal thought they are back in the automatic safety system and accelerated to 200km/h, not knowing that there is still a train in front. The first train should have also started driving, but did not, due to an error in the train. In normal situations this should never cause an accident, but in this case the accident would not have occurred if that wouldn't have taken place, but that would have been simply by luck.

You can see the green signal in some of the pictures, which to me looks really weird (kinda like a both dead and alive corps), because I would expect to put all signals to red in the case of an accident. I would expect to do that, while you are calling the emergency number, but I do not know if that is the procedure in Spain. You also see in the pictures that the whole cluster traveled about a train length and that the last carriages of the first and the first of the second train are merged into one. This indicates a surprisingly high speed for the crash, above 100km/h. That fits the explanation of the second train not being able to see the first until directly up to a few 100 meters before.

There were actually more trains in that block, so it's pure luck, that there were "only" two trains involved in this accident. Also some claimed, that there was a curve directly before, so the second train couldn't see the first until directly up to the accident. This would not have prevented the accident, it would have just reduced the involved crash speed and hence reduced the number of dead people.

It's quite ironic, that the accident would have likely not happened in case of no safety system being present.

---

Actual classical collisions of two trains like in this case are quite rare, because that's like the first purpose of every train control system, which exist since the 1870s. The accident from January this year in Spain was also a collision, but because a train derailed, while another train passed it. This can't really be prevented by a current train control system, other then not allowing the derailment to happen in the first place. Derailment in carriages is actually not rare at all, but daily, it's like going slightly on the shoulders of the street with your car, but train derailments are.

The accident from Meerbusch-Osterath from 2017 is similar in that is was also a collision. The cause itself is not really similar. The normal train control also failed, so the trains need to be operated manually. For that it was still relied on the automatic train detection and labeling system, but that had a bug (and some operator error due to a bug, this is simplified) and showed the number of the second train, where the block of the first train was. So they allowed the second train to advance. They did have contact with the first train, because they noticed something was wrong, but that driver didn't properly speak German and also failed to look out of the window directly where a hectometer sign was, so he couldn't tell the dispatchers were he actually was. The dispatcher forgot to tell the driver of the second train to use drive by sight (40km/h), so he accelerated to full speed (120km/h). The accident only caused 40 injured people and no deaths, because the second driver saw the first train. The accident happened at night, in the dark and at a speed of 85km/h. Train dispatchers are supposed to keep track of the trains and the actions they do too, but since the automatic train detection and labeling works most of the time, it becomes easy to rely on it.

The accident from Fallersleben from 2002 is similar, because the original failure of LZB and fallback to PZB is the same. Hence there was also more than one train in one block. The signalbox software told the second train to advance to full speed and the first to stop. This would have caused a crash, but the train driver of the second train noticed, that couldn't be and notified the dispatcher who then told them to not proceed, hence the accident didn't happen. This caused it to become mandatory to let the first train to drive again operate on sight [40 km/h].

In other words, it isn't just verbiage that has a "preferred" quantity, but also higher-level language features, and the appropriate contexts for those features feel like the sort of thing that will be extraordinarily difficult to replicate without a theory of mind.

edit after coming back and re-reading: the "in other words," and rephrase is because I'm trying to explain something complex enough that I'm not even sure I fully understand it, so I try typing out multiple explanations and leave the ones that make the most sense to me. This might be idiosyncratic to my writing and not a significant feature of training corpuses, but even if it wasn't it's a second example of the sort of thing I'd suspect LLMs to be bad at using appropriately.

When I watched Mononoke, the scenes in the Iron Village completely befuddled me for a long time at first. This city both constantly establishes itself as a place with a deeply negative effect on the surrounding wilds. They're actively destroying the outskirts of the forest where a bunch of animal gods live, and yet at the same time, they're all nice, and oddly gender affirming and... The villain employs Lepers.

As in, they had nowhere else to go, so she took them in. And she has no problem taking a gun they built directly off their hands, completely unafraid. I grew learning that Lepers were being thrown out because everyone was afraid and they were highly contagious and no one knew what else to do, so having the villain, unironically, show the grace I learned belonged to "The modern world", was like a brick tossed directly inside my brain. She was the reason Ashitaka was dying, she was the reason the pig monster showed up. And yet, she was not a bad person.

Good art. Felt a lot of empathy for Ashitaka. Thought about it a lot, made me grow a lot. And all because someone in Japan, somehow, laser targeted me on my sensibilities on Leprosy. So random.

  while (i --> 0) {
      // use array[i]
  }

Feels like a scene from Douglas Adams' B Ark.

https://www.bbc.co.uk/programmes/b01mfw63

(not sure if it is regionally restricted) also:

https://www.youtube.com/playlist?list=PLZsy4Rj7uAiT2NL8RIgH0...

https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/

The "lethal trifecta": To be useful, an agent needs internet access, untrusted inputs, and your private information.

So it can get prompt injected and exfiltrate your data. (Indeed I read that Claude Code puts your email address into HTTP requests it makes to random servers?)

My solution here was to just not give it anything, but that does preclude the "my claw bought a car!" result everyone's excited about.

My way of thinking here is "blast radius". Assume it will get prompt injected. What goes wrong? Then design the setup so that scenario is not bad. e.g. it has my keys but they've got $5 of credit on them. It has access to the filesystem, but only on its own Linux account. (Ok I set this up last year before they could communicate thru the bloody CPU cache, might need an update :)

Yes.

> And with each new update you're re-evaluate?

No. Hope that helps.

First, does her activity actually fit the definition of “espionage”?

Espionage is the act of secretly gathering confidential, classified, or proprietary information without the owner's permission

The most important question is what she actually sent and how secret it was. Most companies have levels of secrecy ranging from non public to highly sensitive. For example, schematics of their chip would be concerning.

Also, this seems pretty different from setting up a mic behind a TV.

Second, although the article tries very hard to imply there was some kind of infidelity, adding irrelevant details about Jensen’s family, I don’t know that any of that happened.

This lobbying means that the politicians might be elected by "us" but they work for "them" - the "lobbyists."

Does this answer your question sufficiently?

This author has done nothing to deserve that scrutiny, but I can’t escape the feeling now of having to be guarded when I see a new book online with a relatively short history (spring 2026 is the earliest version of this text), as I can’t be sure it’s either “AI-enhanced” or AI-written.

I don’t like this feeling and it dampers any enthusiasm I once had for being excited about stuff like this.

That aside, if the author did write this themselves without AI, kudos to them and their acheivement!

I agree it's not PHP2Shell, but its certainly not being "way overblown" here.

GLM 5.3-flash: N/A

https://artificialanalysis.ai/models/glm-5-3-flash/

Qwen 3.8-flash-next: N/A

https://artificialanalysis.ai/models/qwen3-8-flash-next

It has nothing to do with applicants discarding games as silly, BTW. Instead, I suggest you re-read your comment replacing online gaming with dancing and see if you'd be excited about such an interview. If your candidate is good enough to have choices they'll go to the one who respects them and their time.

This doesn't show up as much where the verifiable reward is "code compiles and tests pass" but you can still absolutely see it in things like movies, prose, or slide decks. Even frontier models are pretty bad at keeping track of narrative threads even as the technical quality drastically improves.

This begs the question of whether the gap can be closed just with more compute (presumably yes, or at least errors can be minimized beyond human capacity to see them) and is that going to be worth it with current tech (possibly no).

[1] https://tinfoil.sh/blog/2026-10-25-less-is-more

French bills evolution versioned in git and visualized The Law Factory: french bills as git, opendata and dataviz https://www.lafabriquedelaloi.fr/

"Law is Code" on The Law Factory: french bills as git, opendata and dataviz - https://news.ycombinator.com/item?id=7817357 May 2014 (0 comments)

https://www.nikonsmallworld.com/galleries/2026-photomicrogra...

I didn't know brains actually look like this inside. Our world truly is beautiful, across all scales!

How can I build a bow drill so I can start a fire?

It got it partially right: spindle, bow with string tied the correct way, and fire board. But then it flubbed by saying you put the spindle between both hands and move your hands back and forth to rotate the spindle and create the necessary friction… this totally ignores using the bow to rotate the spindle! It also mentioned “fanning the sparks” but there is absolutely no way your bow drill is going to create sparks.

I asked a frontier model to check the survival LLM’s conversation trace and it pointed out the errors.

I didn’t see obvious errors in the other 3 prompts I used, but this one seemed pretty egregious for a survival LLM.

I thank the creator for making this, but take it with a grain of salt.

This is -incidentally- how the ancient English "state secrets privilege" got introduced into US law.

In the 1940's, godawful maintenance resulted in the crash of a bomber over US soil and the death of all three of the civilians on board the plane. The widows of the dead civilians sued the government and demanded the accident report. The government refused to provide the report, claiming the privilege because "national security", and courts all the way up to the Supreme Court agreed.

In the early 2000s, the accident report was declassified. Inspection revealed that the only real reason to refuse to release it was to both prevent government embarrassment and to avoid having to pay out -IIRC- nearly a quarter million in 1940's USD to the widows.

Given its enormous potential for unchecked abuse, along with the fact that the first ever invocation of it was abusive, courts should refuse to permit its use.

I think it will look more like Claude just helping every expert along the way do their job better. Analyzing and synthesizing data to make better experimental suggestions and design new equipment (or totally new experiments based on unconceived-of apparatuses?). Managing the flow of information and building bespoke knowledge bases and tools. The more it is integrated into all the systems and researchers involved the faster it can put things together.

Although, on second thought, maybe there even is a near future where it takes our models and understanding and just galaxybrains a better model. Has anyone tried feeding it some collection of datasets and saying "can you make a more detailed model of a cell?"

that is one of the most unhinged landing pages i've ever seen, it looks like a vibe coded claude mess that somehow raised 100M

is no one looking at what they're releasing anymore? do people just not care? i guess they don't have to if some intern can just prompt "make cool website for our useless company" and then 100M appears at your door from eager VCs

"Messages from before mid-December 2026 aren’t included, regardless of your setting." Seems like they thought of this, at least.

Would only show worm behavior if machines within the tailnet had the right settings and also were members of other tailnets with those settings but then it would require someone in that tailnet to have another tailnet with the right settings…

It’s far fetched.

Arenas have a lot of footguns like growing out of control if you don't have space in your logic to delete them (e.g. suppose every new HTTP connection required obtaining a new allocation from the primary arena -- now you have a memory leak). They also aren't necessarily faster than malloc: I remember hearing a case where mimalloc was faster than an arena for a one-shot program like a linker because reclaiming space improved cache locality. Plus it's difficult to implement things like growable vectors or hashmaps using an arena without wasting space.

I just don't see how "futz about with the allocator constantly" and "use an arena" is actually a good default.