I'm going to be a bit contrarian here. In general, I 100% agree, just send me the prompt; however sometimes I feel an LLM can help structure thoughts. It can be a good sparring partner. If you proofread your LLM content, re-prompt until it sounds right and no longer like AI content, and feel that it actually conveys the information incredibly well, then I'm happy.
One of the instructions I've hammered into Claude is "Write like a human. I don't want this to sound like AI content. Your standard style of writing would fail miserably if it was reviewed by an English teacher. I want you to write prose that is nice to read. For example, write full sentences instead of bullet points."
It works wonders. Suddenly, my plan documents are something I actually understand, and something I'd be happy to share as an RFC. They still need iterating every time before they're ready to share, but I no longer have "the honest truth is" and other BS in my output. It's so refreshing for my brain to be able to actually focus on the content.
I've noticed this with my partner, too. She used Claude to draft a strategy document and felt completely overwhelmed. A classic moment of "AI did all the thinking for me, but now I don't know what I'm presenting". Once I helped her prompt Claude into writing the strategy in proper English, she understood what the AI was proposing, rejected large parts of it, iterated several times and ended up with a draft document where she edited the finishing touches herself and felt that it was truly hers. The AI was still incredibly useful: it helped her with the blank slate problem, and tremendously sped up her workflow.
So yeah, if you proofread and iterate on your AI's output until you feel you'd be proud if you had written it yourself, I'm happy to read it, too.
I'll be even more contrarian: I don't care if someone didn't write it as long as it's good (that's the major problem). I don't mind people using quotes when writing either. It's just that typically people (used to) curate their words before presenting them as writing. Although I'd prefer it if they also credited/explained what they used to write it in the case of heavy AI generation.
This, really. If I am entertained and learning then I do not care where the content came from.
Of course Claudeish or GPTish ("unusually") will easily push me away. Just like any repetitive or obnoxious tendencies that might appear in human writing.
> One of the instructions I've hammered into Claude is "Write like a human. …" It works wonders.
If this actually works, it's absurd. The implication is that Anthropic could trivially make Claude sound less obnoxious, but chooses not to. I don't see a way this could be justified as a safety feature or anything, so.
I tried a lot to get rid of Claude vibes and I failed. I made blacklists, I used prompt evolution to test against good human writing samples, even created a dynamic policy for dialogue, a cli tool used to steer agents dynamically by well chosen questions. The poor writing style made even models like Fable look worse than Sonnet 3.7 to me.
> So yeah, if you proofread and iterate on your AI's output until you feel you'd be proud if you had written it yourself, I'm happy to read it, too.
I agree, using LLMs is disproportionately frowned upon, but what matters is if you invested your own attention in the process. I use LLMs a lot for sparing, usually ask it to assume some opposing persona or use web search, not relying on its defaults.
It turned into a luxury place because that's where the highest margins in construction are. Every developer wants to build high margin houses, nobody wants to build low margin houses.
City development must cater to all its inhabitants, and if you have a prime estate like the Battersea Power Station, it's a failure, from the city's POV, if you burn that only on high margin development. From a negotiation standpoint, the city negotiated badly here. The city has to accommodate the needs of less wealthy inhabitants in other means now, but without the negotiation leverage that a high margin property can provide.
If you view it like this, the complaints "only luxury flats" are pretty rational. The city negotiated badly.
There actually is some "affordable" housing in the wider project but only because there was a very long battle to keep the developers to their legally-bound promise to build some affordable property that they — of course — tried to wriggle out of.
Imagine you're the mayor. You want to hire companies to build prime housing and affordable housing. Every company only wants to build prime housing, because that's where the money is. If you only give out contracts like that, basically nobody will apply to build affordable housing.
You're not dumb, of course, so your next step is to say, "if I give you the contract to build prime housing, you also must build affordable housing". That's not ideal for the developer, because the time spent on building affordable housing is time they don't spend on stuff where they make proper money, but at least they can make some money. They agree to take the deal. You think, problem solved, but turns out the developer outsmarts you: they structure the company in a way that they can chicken out of building the affordable housing. In real estate, it's very common to have every building project be its own limited liability company so that one failed project doesn't influence the others. In this case, you put the affordable and the luxury housing together in one Ltd, build the luxury housing, extract the profits by moving them to the parent company and then declare your Ltd bankrupt, meaning they unfortunately can no longer build the affordable housing.
As a mayor, you observe how you're being cheated, you get real angry but the walk through the courts is too slow, and your voters demand action now. Realistically, you do the court route in the background, but you learned your lesson and now you tell the developer "before I give you a high margin luxury contract, you must have completed affordable housing first".
Alternatively, you mix the buildout into the same building so that you cannot open the luxury flats/office space without having completed the affordable part. Either way, affordable first, luxury second.
That is basically the story of the London property market.
Another issue is how do you even build affordable housing on land worth millions? The only way to make that work is if the government owns the land forever and commits to never selling it. Which is kind of what the government was going for after WWII with council houses, but Right to Buy ruined that.
It was only dragged into zone 1 in order to charge higher fares to recoup some of the cost of building the Northern Line extension. If it weren't for that, it'd be zone 2 - and one of the less well-connected bits of zone 2, at that.
I don't have that much of a problem with it, tbh - before the development it was a weird light industrial bubble, disconnected from the rest of the city. Now it's a weird bland commercial bubble, disconnected from the rest of the city.
And we've lost Marco Polo House, the home of QVC, and gained Apple. Seems about a wash in terms of cultural value as well.
Thanks so much for sharing this. As someone who mostly watches from the sideline, can you share what you can see in this dashboard that someone like me can't see? Is it the metrics themselves that they measure (the metrics tab is absurdly detailed), something in the notices, or something else I missed?
I might turn this into a blogpost if folks are interested, but my god there is so much clever info in that dashboard.
Here is one really neat bit:
A cutting edge training idea (for agents, it's been used elsewhere for ages) is on-policy RL, basically, it's not enough to say "here is an end to end agentic sequence (including tool calls etc.) that is perfect" you want to say "here is a sequence you might actually have generated that turns out to be correct".
Basically, it's more training efficient to improve models with small tweaks to do more of the right thing they are already doing sometimes than from some perfect oracular "this is the way" answer.
(if you've ever tried to teach humans new skills, you’ve probably noticed this too!)
When you do that, you care about how far the model you are updating (improving) has deviated from the one being used to generate rollouts (agentic rollouts for hard problems can take hours with lots of tool calls, so you can't keep redeploying every slight improvement).
Lo and behold, the dashboard literally has:
partial/avg_staleness (likely the measure of how many micro iterations the "generate answers" model is behind the "improving based on the occasional right answer" model)
train_infer_diff/new_infer/kl (a more direct KL divergence based way of measuring how differently the two models generate tokens)
How cool is that?!
And don't get me started on the clever ideas hiding behind dynsam/avg@n ...
Hold on, isn't that just standard practice for post-training LLMs for agentic use? Give task, generate n rollouts, grade rollouts (either at termination or after each tool call)? Or is the difference that the rollouts are generated ahead of time and then graded? (Of course, then it's not really on-policy.)
The best thing they did is being open about all the setbacks they had to deal with. They logged every restart with a reason, talked about dropping a cyber dataset after it degraded coding benchmarks. Also published real time training loss, benchmark scores after every checkpoint and running cost estimates.
Really the only thing missing was dataset descriptions, the dashboard only had random IDs like "dataset-zrso". I guess it's their lawyers fault.
It'd be great to see a description of even just a subset of training datasets. It feels very much under-reported how much expense is worth investing in preparing and selecting training datasets versus just using masses of random quality unprepared training data. This dashboard appears to be good though in showing the limits quickly reached when throwing parameters and compute at the problem.
For example, if they were to train on Wikipedia dumps, do they consider every article to be the same quality across each language, or have they done more work beyond Wikipedia's own article quality ratings to make training decisions such as "Ignore cebwiki it's machine-generated spam" and "Treat dewiki articles with coordinates within Germany as being higher quality (weight it higher) than their equivalent enwiki articles".
And let's say one of the datasets is all the source code of packages in the Gentoo package repository. Not every software package is a good example of how to write code. You perhaps wouldn't want to train your LLM on 1990s era PHP web application source code as an example of how to write code in 2026. Instead, you'd possibly want to use such PHP web application source code as a negative training example of what _not_ to write. But when training an LLM to detect software bugs, maybe outdated PHP source code is good for training.
Similarly for translation, perhaps UN treaty documents translated into 4+ languages are good translation examples because of high accuracy needed, professional translators being used, and bigger budgets. However this training data would perhaps be a negative training example towards translating chat messages, movie subtitles, etc because it doesn't use everyday slang and could result in output of nonsense such as "Pending Your Excellency's response, please accept, Your Excellency, my sincere greetings." for a prompt asking to write a birthday card for a child.
Preparing training data and deciding how to best use it for training I assume would be the largest expense (cost of labour -- mostly expert labour too) and also the greatest opportunity in the future for LLMs to improve. It seems to me somewhat irrelevant if the dashboard indicates a compute expense of $1m or $5m if good training datasets (prepared by experts in their fields) cost $10m/y to maintain. For example, hiring expert software developers to tag 1000's of open source software packages according to their quality, on different metrics, such as human readability, performance optimisation with choice of algorithms, reasonable trade-off between coherence and coupling in the software architecture, currency with state of the art programming trends/preferred dependencies/operating system APIs, etc. And keeping that metadata continually updated rather than a rapidly obsolete once off tagging project completed in 2005.
Look at the kid in a costume. The shoulders of ChatGPT’s output are just wrong. Sure, the kid looks prettier, stronger with a wider upper body, closer to society’s ideal of a child. But it’s not him. If you put that child in a suit, he’d still have a smaller upper body and the arms wouldn’t fill out the suit as nicely as the child in the AI image does.
It’s just not the same child.
I know we’re gonna get there eventually, but still, seeing as this is their first demo image on the page I just can’t help but scream inside “HOW CAN YOU NOT SEE THIS?”
It really makes me question whether the people building, or at least the people marketing this, actually understand their product. I love image generation for all kinds of use cases, but I find that particular example creepy.
This was true right when Weimar Germany fell to the Nazis also. There's a degree of inertia to politics. We're also talking about a region of Germany that has been economically depressed for decades, with an unemployment rate twice the national average.
That's because the PM's party brokered that half-way deal with EU against the will of the people, after they had voted no to EU for the second time[1].
Not saying some of our export companies aren't enjoying the access to EU, but the people was against it.
> PM's party brokered that half-way deal with EU against the will of the people, after they had voted no to EU for the second time
what polls have to do with it? if the will of the people was against the eu deal, they should've voted against, they had 30 years to do it. norway is a democracy, right?
One of the instructions I've hammered into Claude is "Write like a human. I don't want this to sound like AI content. Your standard style of writing would fail miserably if it was reviewed by an English teacher. I want you to write prose that is nice to read. For example, write full sentences instead of bullet points."
It works wonders. Suddenly, my plan documents are something I actually understand, and something I'd be happy to share as an RFC. They still need iterating every time before they're ready to share, but I no longer have "the honest truth is" and other BS in my output. It's so refreshing for my brain to be able to actually focus on the content.
I've noticed this with my partner, too. She used Claude to draft a strategy document and felt completely overwhelmed. A classic moment of "AI did all the thinking for me, but now I don't know what I'm presenting". Once I helped her prompt Claude into writing the strategy in proper English, she understood what the AI was proposing, rejected large parts of it, iterated several times and ended up with a draft document where she edited the finishing touches herself and felt that it was truly hers. The AI was still incredibly useful: it helped her with the blank slate problem, and tremendously sped up her workflow.
So yeah, if you proofread and iterate on your AI's output until you feel you'd be proud if you had written it yourself, I'm happy to read it, too.
reply