choeger

That's a very old idea that never really worked. There are two main objections against that concept:

* Any tree structure would be highly language-specific. Your VCS wouldn't support experimental or new or propietary languages. Even worse, for most real languages, operations like refactoring require not a concrete or simple abstract syntax tree but some transformed or enriched (e.g., with type information) variant that is highly specific to the tool that performs the operation (think Eclipse internal tree structures). Language servers might help here, but they're far from the only implementation.

* In practice, if I have a textual diff and a tool that can generate the required tree, I can always apply that tool to the textual representation, just paying a little bit of compute. Compute is cheap.

This demonstrates how any algorithm, software or not, works: Start with something desirable or necessary and add complications until you end up with a mess that can be used to yield arbitrary results, drive outsiders into madness, and summon some lesser demons if read in a chanting manner.

Now what could be done about this? The obvious answer is refactoring. That's what politicians promise in election campaigns and junior software developers demand when encountering a new code base. It's rarely ever done in practice because no one can disambiguate unnecessary from desired features and complexities. The second answer would be a clean implementation in some more formal framework (a statically typed language, a theorem prover, ...). That second solution often fails when attempted because the system is already unsound. Fortunately, no one cares about those "weird corner cases" and everything is left as-is until someone runs away with a million dollars or so.

Huh? Who's been playing with fire? Fedora ships 15k packages. They don't develop them. They're trying to modernize hundreds of packages on a volunteer basis.

Germany today (massively rainy and windy day) had a peak of over 50GW. Surprisingly we still had 2GW solar peak even today with the worst possible weather and over 9GW just two days ago.

Certainly a cool achievement and something to be proud of. Maybe even a very important step towards real usable fusion power plants.

But:

> The facility’s laser system is enormously inefficient, and more than 99% of the energy that goes into a single ignition attempt is lost before it can reach the target.

So it's still two orders of magnitude off hitting break even. And that's without a suitable device to actually capture the released energy and it does not account for pellet production. For all we know, it might very well be three orders of magnitude away from a real powerplant.

That being said, making more efficient lasers and working on capturing energy might be more fruitful than trying to improve a Tokamak design. So it's good to have options.

> Have you ever been to Germany? one of the safest countries in the world, never had any issue with safety there.

Have you ever been around German 20 story apartment complexes? I have and I can tell you it's not pretty.

> Usually you find variety of size, between 1 and 4 bedrooms the most common.

3 bedroom apartments in German high-rise complexes are a rarity. I never heard of 4 bedroom apartments except for the very rare penthouse or designer building. The standard here is definitely 2 bedrooms with 3 being a luxury.

Interestingly, you don't give a source for Germany, which according to this source has much lower emissions per kWh these last days:

https://www.agora-energiewende.de/daten-tools/agorameter/cha...

And even this source is a bit dubious, as there's no clear correlation between the amount of coal and gas and the emissions. Presumably they also count imported energy.

Looks like a very interesting paper for everyone trying to test any kind of tool that works on a (functional) language.

Love the Acknowledgement section.

Maybe I am missing something, but when you project every content into a fixed-size space, you have to lose information at some point, don't you?

Did somebody ever run an image of a scene vs some parts of the scene in e.g., CLIP?

I sometimes use the term for children in a joking manner.

If you're a large male, especially if you sport a beard, you will probably experience that smaller children react to you with shyness or fear quite often. I have seen children crying when a young men entered their car. The small kid hiding behind their mother's legs is a stereotype born from real-life experiences.

But the behavior often completely changes when that male person is accompanied by another child that isn't shy around them (e.g., their daughter). Apparently, there's a deep-rooted instinct in children that tells them "this one is a father and thus cannot be a bad person".

I assume these are mostly procedures, not functions, (that is, they have side effects) as applying a function 1.3M times per minute immediately raises the question of caching which doesn't seem to be important for them.

But if you execute that many procedure calls, how do you guard against them influencing each other due to their side effects? Memory leaks come to mind, or other weird bugs. Also, how do you manage the credentials to effectively issue side effects in a (hopefully) zero-trust environment?

The initial landing would have been successful, assuming German air superiority and an actual will to execute it, of course. There's simply not enough time for the Royal Navy to interrupt the landing unless they're already assembled so close to the beachhead that they'd risk getting taken out before D-Day by the Luftwaffe. Also, the landing would have involved plenty of paratrooper and airborne infantry as a first wave.

The big question would have been resupply. Germany didn't have the mulberry harbors and thus would have been forced to capture a port and ferry supply and reinforcements through it. That's when the Royal Navy could have run interference and I have no idea how that would have turned out.

So in the end, I think that the Royal Airforce saved the island.

Please, please, please, make sure that it uses a context-free grammar. For the live of Knuth, correct that one stupid mistake he made and allow other people to parse your syntax. Good things will follow!

Resistance against antibiotics cannot come free for bacteria. If we stop using some, the resistance should vanish as quickly as it turned up, shouldn't it?

I've yet to see a convincing new business model for LLMs. They are impressive, don't get me wrong, but right now they only seem to augment many already existing businesses or workflows in one way or another.

Unless that changes, "A.I." will be a commodity that you purchase like you already do with storage. Hardly exciting, tbh.

At least in theory, HDCP keys can be revoced.

Manufacturing non-compliant devices could become extremely risky.

That's not true, AFAIK. The DRM module (e.g., widevine) has to negotiate keys with the host and will decrypt the stream for decoding. The HDCP keys are static and independent from the streaming service. That decryption can happen with hardware support but it can also happen in software. I think most streaming services fall back from 4k to FHD or lower if their "security requirements" aren't fully met.

Also, if the DRM module supports hardware decoding, the GPU will need to see a decrypted stream. The GPU memory is in principle accessible to the driver. So if you can manipulate the driver, or setup another kernel module, you might be able to grab that content.

Realistically, no. You don't have access to the device that contains the decrypted content.

Of course, you could try to open up or bypass the DRM module (e.g., via the GPU driver during hardware accelerated decoding), but that won't work with a TPM module and signed kernel space.

If you look at the security technologies introduced into consumer decides over the last decades, it all serves the dual purpose of securing your hardware from ... you.

There's a huge fraudulent scene when it comes to scientific papers, IMO largely driven by the "schoolification" of universities. Many authors seem to consider writing and publishing a research paper as just another form of an exam and thus cheating as the logical approach.

First of all: I always have and always will be a proponent of statically strongly typed languages (although I would argue that "statically strongly" is redundant here, there are only typed and untyped languages, IMO).

But there's one cost that the author seems to miss and that I didn't fully appreciate for some time. That's the cost of composition. If you look at python you can pretty much compose all possible packages on pypi in your project. All their means of abstraction are usually fully compatible (except for concurrency). Composition often only requires a little bit of glue code.

In a statically typed language's ecosystem you will probably a handful of fundamentally different means of abstraction like functors vs. classes that simply don't compose well together. Hence the fragmentation increases massively.

Could that be a consequence of TikTok's raise in popularity? AFAIK, it's much harder to share a link there compared to FB/X.

Did anyone ever attempt a recursive architecture?

So you take the first window or logical separatation (chapter, paragraph) and let the model summarize it into one or two sentences. Then you repeat that with the next window (and that derived sentence as context) and create a new logical separatation out out of a fixed number of sentences. Rinse and repeat until the result fits into your window.

I have a hunch that this is somewhat how the brain works when reading.

"exceed the mass of tailpipe emissions" - that implies all tailpipe emissions.

They compare "particulate emissions" to "tailpipe emissions". That's at the least very sloppy writing, at the worst it's intentionally misleading.

> Indeed, the scale of these emissions is significant. Particulate emissions from tires and brakes, particularly in the PM2.5 and PM10 size ranges, are believed to exceed the mass of tailpipe emissions from modern vehicle fleets, as per a study published in Science of the Total Environment this year.

This simply cannot be true. Tailpipe emissions have more mass than the gasoline that gets burned. A car uses several liters od gasoline per 100km. I am pretty sure no car loses several kilograms of tire or brake matter over 100km.

That approach works fine in gerrit. I always preferred it.

What ~everyone~ most people miss about self-improving AI:

1. It would require to control hardware development, manufacture, and operation.

2. It would require to control Software development and operation.

3. It has no purpose other than the one given to it by its creators/operators.

Right now, every so-called "AI" is completely dependending on hardware and software being "fed to it" by humans.

There's no "AI" creating software, there's no "AI" creating hardware, let alone operate it, and there's no "AI" operating or controlling other "AIs", nor is there any "AI" that acts in any way or form selfish.

It's certainly fun to think about how a self-motivated language model that tries to take control over the humans that operate its server farm, but without an army of (humanoid) robots, it certainly poses no danger, as it cannot replace humans or even threaten them.

The big challenge with "AI" is that we will probably be unable to distinguish its output from the one of a human being. This will have severe impact on how we can trust any form of remote communication.

That sounds like a slam-dunk GDPR violation case and a hefty fine.

Personal LLMs make a lot of sense. Until they don't.

Right now, having a OCR combined with an LLM would open the way to a personalized assistant (rather a well working search machine and letter/email generator). No need to manually manage or catalogue documents, photos, videos, contacts, it would all be available by grepping your input stream smartly.

But once everyone uses such a tool, why bother generating human-readable text in the first place? There's so much efficiency to be gained: Send plaintext instead of html or pdf. Replace plaintext with its embedding. The sending AI could learn what the receiver understands. Finally, don't send stuff at all. Just point to the relevant vector DB entry.

In the end, the LLM itself will be obsolete, except as an appendix to communicate with humans. For all other tasks it will be superseded by a evolvednit engineered, form of machine-to-machine language.

Git clearance, Clarence.