They are going to lose the next election, maybe to Reform, if they don't turn things around. Pandering to Reform voters is pointless, they are not going to vote Labour unless they go full fascist. Meanwhile they are alienating everyone else.

LibreOffice doesn't have cloud sharing features that allow multiple users to access a shared file with different permissions.

LibreOffice Calc does allow multiple users to edit a spreadsheet on a network drive, but doesn't have a user permission system or integration with a single login somewhere. The other apps like Writer don't support collaboration at all.

No, that is not correct.
It is close though.
1) "The internet" is not persisted on disk.
2) It's a gross simplification, but the latter half is correct.

And I'm glad you phrased it like that, since that equally describes a functional description of what your brain does.

Yes. These are simple tasks. Try it. Quit looking for contrived gotchas. You are not acting in good faith. Try it.
These things are not fooled by orientations. They're not OCR. The picture is treated as embeddings- multi-dimensional vectors of features.

Handwriting and diagrams are old-hat.
Agentic loops can write code, run a test, evaluate a screenshot of that test, and correct problems visually, with no user intervention required.

For shits and giggles, I just fed a dental surgery referral through Qwen3 VL. Not only can it read my dentist's handwriting, it correctly identified which teeth had been circled, what boxes were checked and which were not, and even noted that there were 2 distinct handwritings, which it also correctly guessed was the dentist, and their secretary.

That seems unlikely. The worst case I could find for high speed rail was 52.7g/km of CO2 emitted, with a capacity of around 1,300 passengers. That includes the emissions from the stations and so forth, and equates to about 0.04g/passenger/km.

For a typical A350 you are looking at 0.18g/passenger/km in economy class, and that is just the fuel, not the airport or the aircraft or the transport to get to and from airports at either end etc.

Incorrect.

Your SQL is broken down into tokens that are operands and query values, which then scans a database to satisfy the provided grammar.
You are trying to gloss over everything into a compressed understanding space that works in your head. It's simple, and that probably works for you- but it's wrong.

The hotel contains handwriting (my name, and room number), and the texts are not a mixture of printed, and graphically designed cursive, and logos with text in circular shapes, etc.

It's a perfect test, because it's the kind of thing OCR will fail at, and this will succeed at.
Handwriting is easy in comparison.

That's what they are to you.

Behind the scenes, they're appended to a buffer called a context window, which serves essentially as the model's "short term memory".
To the LLM you're talking to, they're just a series of very large 1000-dimensional vectors that were embedded via the self-attention layers from the context, within the network itself, they're not words, or tokens, but semantic concepts, relationships, inferred meanings, anything it could learn from them and similar configurations of them in its training- called embeddings.

It has no built-in concept of a query. It is not a search engine searching a database.
It's quite literally nothing but an obscenely large nonlinear function with a literally astronomical configuration space.
At the end of all that, a decoder layer turns that into token probabilities, and the application you're using turns that back into actual text via a sampling method.

The fact that it is treated as a query is precisely because the model is so fucking good. It has a concept of questions. It knows how they're typically arranged in a sentence. It has also been trained to be helpful to you- to answer questions you might have.

Odometers aren't all that reliable in general. A few percent at best. They don't account for things like putting different size wheels on the car either, and can be wrong if the wheel size is set wrong at the factory.

Doesn't excuse Tesla illegally voiding warranties.

Any multi-model LLM can read your notes, and cleanly interpret and describe diagrams you have drawn on them.
Qwen3 VL is my current favorite open-weight multimodal model.
That said, I'm quite certain the SOTA models (ChatGPT, Gemini, et al.) all can as well.

Not only do these models outperform traditional OCR at reading, but they can also describe other details like layouts in any kind of format you like, including re-creating the image they're looking at in HTML or things like that.

I will, as an example, ask Qwen3 VL (32B, FP16- the dense model, not the sparse) what it sees while holding open a recent hotel access sleeve and card from a conference in Las Vegas.
Its response.
As you can see, they can read quite fine. And this is an open-weights model, running entirely locally on my laptop. And it's old as hell (at least in LLM terms).

I'm unsure why it's my job to tell you what you could have tried, with the knowledge you already have (That uhhh, ChatGPT... exists) to have seen this for yourself.

I'll grant you that I don't use the paid models very often (and certainly not for something local models can do fine), and that free-tier distilled SOTA models might even suck at this.
Paid models are going to be better than the LLMs you can run locally, though, and they can read notes with ease- even notes I don't find remotely legible- like the room number on that sleeve, which can only be described as chickenscratch.

I haven't been able to find a source for this 12,000 claim, but it seems likely that it's untrue.

My guess would be that they simply looked at every arrest where evidence included social media posts, e.g. if someone was assaulted and the attacker happened to have posted on social media about it, that counted.

I am no fan of the UK or the way it is going, but there were clear directions from the government a few years ago that social media posts should only be the basis of arrests in very specific and fairly extreme circumstances, e.g. where it is reasonably believed that there is an imminent threat of violence.

That was not a joke, that was part of a very long running campaign to incite violence against trans people.

Graham Linehan has just been convicted of smashing a child's phone when she confronted him about the months long harassment campaign he waged against her on social media, which included attempts to dox her.

He got off extremely lightly, all things considered. The only reason he beat the harassment charge was because the judge didn't think that the victim was sufficiently harassed, and the prosecution didn't prep her very well.

He's a bigot and a convicted violent criminal, and has gone straight back to harassing people.

The biggest problem for the government is that the welfare bill is expected to rise by around £90 billion/year by 2030 (OBR forecast), and half of that is pensions alone. The population is ageing and also turning against immigration of the young and healthy workers we need.

Mistakes were made going back to the 1960s and no government has the will to address them, only to kick the can further down the road.

The more expensive public charging is about on a par with a decently efficient fossil car. When it can get expensive (aside from rip-offs) is when you also have to pay for parking, which is common in London.

If you can charge at home then it is much cheaper, around 2p/mile, or free if you have solar.