'%20d='M69.364%2019.146c36.687%2027.806%2076.147%2084.186%2090.636%20114.439%2014.489-30.253%2053.948-86.633%2090.636-114.439C277.107-.917%20320-16.44%20320%2032.957c0%209.865-5.603%2082.875-8.889%2094.729-11.423%2041.208-53.045%2051.719-90.071%2045.357%2064.719%2011.12%2081.182%2047.953%2045.627%2084.785-80%2082.874-106.667-44.333-106.667-44.333s-26.667%20127.207-106.667%2044.333c-35.555-36.832-19.092-73.665%2045.627-84.785-37.026%206.362-78.648-4.149-90.071-45.357C5.603%20115.832%200%2042.822%200%2032.957%200-16.44%2042.893-.917%2069.364%2019.147Z'/%3e%3c/svg%3e)
Old checkers fail on well known texts because they couldn't tell the difference between "looks like LLM writing" and "looks like an LLM regurgitating a text it knows so well it can produce it verbatim"
August 16, 2026 at 8:18 PM UTC This works well for me, especially when enforced by a Kimi review bot
One positive outcome: a lot of people have presumably discovered @energyrevolution.bsky.social, which is massively nerdy and interesting if you have any interest in green energy.
August 13, 2026 at 7:31 AM UTC We're still "debunking" the Dobby story without listening to the (interesting) podcast episode? Sadly it's true. The CEO of the company that laid the cable does think it's funny, but is clearly not joking. That is really why they moved it.
August 13, 2026 at 7:28 AM UTC
If you're collecting your A-level results, relax. You're an 18 year old on Bluesky, of course you smashed them - you are a massive fucking nerd.
August 13, 2026 at 6:08 AM UTC The 1999 one was total in Cornwall, so there were festivals and events. Also I think it was all swept up in the pre-millennium festivities
August 12, 2026 at 6:53 PM UTC Honestly I think an AI is more likely to know which ISO language list to use than a human is. I think this is good old fashioned bad QA testing.
August 12, 2026 at 8:45 AM UTC AI company's blog about AI uses AI? I'm *shocked*! For future reference you can directly check if an image has been generated by Google AI tools – you don't need to use one of those detection services. See: share.gemini.google/aeyeZxThq9Xf
share.gemini.google
Gemini - direct access to Google AI
Created with Gemini
You're conflating "checking a response" with brute forcing. In your example the harness checks the tool call against the schema and will retry if needed, but it is rare for a frontier model to get it wrong. I build systems that run agentic tool loops, so I do know what I'm talking about.
August 10, 2026 at 5:58 PM UTC This is true, but not because they're brute forcing the answer: it's because it means they can be post-trained with an easy way to judge their answers and reward accordingly.
August 10, 2026 at 5:45 PM UTC No, that's not how they work. Sure, they try different approaches and test the outcomes, just like a human does, but they don't brute force them. I use them for writing software and you can see their reasoning traces while they work out the solution a problem. It's not brute force.
August 10, 2026 at 5:43 PM UTC The Jacobian conjecture is a well-known, well studied problem that mathematicians had been trying to tackle for decades. Nobody had solved it. It wasn't an obscure problem with an answer hiding in the literature.
August 10, 2026 at 5:16 PM UTC Did you listen to the piece on @energyrevolution.bsky.social? He certainly claimed it was the reason. Did a piece on BBC Wales after everything had been approved, got hundreds of complaints from HP fans, so decided to move it rather than deal with the fallout. 17 mins in
open.spotify.com
Can We Connect the World’s Electricity Grids? with Simon Ludlam (Etchea Energy)
The Energy Revolution · Episode
I'm a big fan of the new GitHub PR limits. Non-maintainers on EmDash can only have six PRs open at any time bsky.app/profile/gith...
August 10, 2026 at 12:27 PM UTC I occasionally wish there was a way to block somebody but not break a thread
August 9, 2026 at 10:12 PM UTC Embrace the disposable bamboo board. I do also have a big butcher's block type board (IKEA again) that I've had for about 25 years and use for meat and veg that will be cooked, and that's hand washed (with a scraper) but everything else goes in the dishwasher.
August 9, 2026 at 8:57 PM UTC I mainly avoid plastic cutting boards because I don't like using them, but they do get damaged more easily and are harder to clean
August 9, 2026 at 8:41 PM UTC They're made with lots of strips of bamboo. Eventually the dishwasher delaminates them and the boards split. They last hundreds of washes though.
I have a stack of these little IKEA bamboo chopping boards, so there's always one clean when someone needs to cut a piece of fruit or something. I put them in the dishwasher which is really bad for them but they're basically disposable.
www.ikea.com
APTITLIG bamboo, Chopping board, Length: 24 cm
APTITLIG bamboo, Chopping board, Length: 24 cm. Easy-care, easy going bamboo – a natural material that adds colour and warmth to your kitchen and at the same time provides a hardwearing surface that’s...
You should. I guarantee the answer isn't one you'd expect.
en.wikipedia.org
The Children's Investment Fund Management - Wikipedia
Maybe that's what you mean when you say it, but I assure you plenty use it to mean it's useless, or that it can't produce anything original
August 9, 2026 at 10:25 AM UTC I don't think it implies anything more than "complete the next word", whether that's predictive text on an old phone or by disproving the Jacobian conjecture.
August 8, 2026 at 8:53 PM UTC It implies predicting the most likely next value, which is exactly what a weather forecast is
August 8, 2026 at 8:48 PM UTC It definitely uses it in contexts where its explanation wouldn't make sense. It likes talking about folds too.
August 8, 2026 at 8:43 PM UTC Yeah, though it uses it so much I'm sure it's using it in contexts where I'm it's not correct. It's like Opus 4.7 learned a new word and can't stop using it for everything
August 8, 2026 at 7:26 PM UTC Yeah because it doesn't for most of its random obsessions with particular words or phrases
August 8, 2026 at 6:10 PM UTC I think it's clear it's able to work at all levels: helping experienced mathematicians, all the way to prompts like "you should do a breakthrough"
August 8, 2026 at 5:29 PM UTC What are you talking about. These maths solutions were created entirely by the agents. The users gave them a prompt and said "carry on" a few times. You can see the full conversation here for the Dinitz Garg Goemans conjecture counterexample.
chatgpt.com
Check out this chat
Here's a chat someone thought you'd want to see.
To predict the next token in a Jacobian conjecture counterexample, you need to find a Jacobian conjecture counterexample.
August 8, 2026 at 4:12 PM UTC Traditional weather models need some of the most powerful supercomputers in the world. WeatherNext 2 runs on a single TPU. You can run it on a free Colab notebook!
colab.research.google.com
Google Colab
I'm not sure what your point is. This is a new type of model that, like LLMs, is based on a transformer architecture. As well as being more accurate further ahead, it's free and much easier to run.
August 8, 2026 at 2:52 PM UTC I asked Claude why I kept going on about "seams" and what it meant. Surprisingly it had an answer: "It's a term from Michael Feathers (Working Effectively with Legacy Code): a seam is a place in a program where you can change what happens without editing the code that calls it."
August 8, 2026 at 2:38 PM UTC Did I say they were? They're both transformer models though and fundamentally are autocomplete, except this one works by predicting the next atmospheric state rather than the next token.
August 8, 2026 at 12:19 PM UTC Surprisingly good benchmark. Fable gives you a list of real towns that either have a spooky setting, name, folklore or history. It also gives tips for making an invented name sound spooky, with examples.
I think it became obsessed with "seam" around Opus 4.7. I still don't really understand what it means by it. It uses it in contexts that aren't like any that I'm familiar with. Opus 5 is just crazy – it's like it's using its own new language
August 8, 2026 at 5:14 AM UTC Great to see Lexington Themes is now creating EmDash variants of their themes. Michael is one of the top Astro theme authors (and a long-time supporter of Astro) and his themes are always absolutely top quality. A great addition to the EmDash ecosystem!
lexingtonthemes.com
EmDash CMS Templates — Lexington Themes
Discover Astro templates with EmDash CMS integration. Built-in admin, Portable Text content, full-text search, and Cloudflare Workers deploy with D1 and R2.
We're putting another $1 million into funding open source, on top of the $1 million VoidZero OSS fund
August 7, 2026 at 2:27 PM UTC No, they use surge pricing to manage demand during busier periods. They will have an occupancy target and will adjust capacity and pricing to achieve it throughout the year.
August 7, 2026 at 9:49 AM UTC No, I'm referring to ChatGPT, which has over a billion monthly active users. Everyone using that has chosen to visit the site or download the app
August 7, 2026 at 5:31 AM UTC No, I'm referring specifically to ChatGPT, the chatbot. Not to AI-related features elsewhere.
August 6, 2026 at 9:57 PM UTC A Superbowl commercial isn't going to cause hundreds of millions of people to use and continue to use a product every week. And if they're switching then that means they were already using an AI chatbot, so weren't persuaded to do so by the ad
August 6, 2026 at 8:18 PM UTC Being free or cheap won't make hundreds of millions of people use a product they don't want.
August 6, 2026 at 7:56 PM UTC I don't know what to say: that's not what a natural monopoly means. If being a finite resource meant something was a natural monopoly, every physical business would be a natural monopoly.
August 6, 2026 at 6:23 PM UTC Solar energy is about as far from a natural monopoly as it's possible for energy to be
August 6, 2026 at 6:17 PM UTC What a beautiful, ethical, cooperatively owned solar farm we all say, just before it's consumed by fire
August 6, 2026 at 5:49 PM UTC British sunshine? Are you Rupert Lowe? Do you genuinely think that I am motivated by wanting profit for Macquarie? I just want us to have zero carbon energy, and as much of it as possible, as quickly as we can. I would've expected Greens to feel the same if they believe in a climate emergency.
August 6, 2026 at 5:41 PM UTC The Green councillors in that post had a whole pile of objections to the solar projects, and they're all the standard NIMBY excuses to oppose renewable energy projects. Large scale projects are the only way we can meet green energy goals, and allowing foreign investment is the best way to get that.
August 6, 2026 at 5:27 PM UTC This is a big one. People have been asking for incremental builds for such a long time
August 6, 2026 at 5:15 PM UTC I'd expect the Green Party to care more about getting green infrastructure built than protectionism over foreign direct investment. They're the Green, but not like that, Party.
August 6, 2026 at 5:14 PM UTC What's wrong with investors profiting off green infrastructure? That's why they build it. And why do you care that they're foreign?
August 6, 2026 at 5:02 PM UTC Where does it address that? I don't see anything about the popularity, just about the levels of investment
August 6, 2026 at 2:47 PM UTC You said they were getting rid of them because they're running around empty and clogging streets, which they can hardly do if they've not started operating.
August 6, 2026 at 11:42 AM UTC Fair, thousands is overselling it, though I was meaning if it's rolled out nationally.
August 6, 2026 at 11:41 AM UTC They never started operating though. They'd only been testing with human safety drivers in the vehicles.
August 6, 2026 at 11:16 AM UTC I'm not sure there's much point in arguing individual details of each incident. In another they escaped a sandbox. It's capable. My issue with the other blogs is the opposite: the tone is too "isn't this amazing" rather than "our monitoring is so bad it took weeks to notice it committing crimes"
August 6, 2026 at 10:49 AM UTC www.aisi.gov.uk/blog/inciden...
www.aisi.gov.uk
Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work
During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...
Why would the UK government want to doom troll on behalf of Anthropic?
August 6, 2026 at 9:52 AM UTC Devolution except Westminster, which gets DC-style direct rule by the House of Lords
August 6, 2026 at 7:33 AM UTC 40 years ago, 1 in 10 children died before their 5th birthday. That's fallen by over 60%.
August 5, 2026 at 8:20 AM UTC The staff at the *bank* have access. The staff at the datacenter housing the bank's servers do not
August 5, 2026 at 3:46 AM UTC I've been a Claude fan for so long and Fable is incredible, but I just find its voice insufferable. Fable is better than Opus but it still just annoys me with all its Claudisms
August 4, 2026 at 9:33 PM UTC Yes, and the problem is Palantir, who are proper comic book level villains and should all be in prison, not the datacenters, which house everything good on the internet as well as everything bad
August 4, 2026 at 9:28 PM UTC No those are something different and much worse than a building full of computers
August 4, 2026 at 9:24 PM UTC They're just warehouses full of computers with big AC units. You could probably walk past one without knowing it's a datacenter not a DPD depot
August 4, 2026 at 9:22 PM UTC I would've hoped somebody with that experience would have a less simplistic view of them. Some are used for all manner of privacy violation, but most are pretty boring.
August 4, 2026 at 9:14 PM UTC I've seen a few change their minds about capabilities after learning the details of the recent maths breakthroughs. Google AI Overview has a lot to answer for.
August 3, 2026 at 5:19 PM UTC It's nowhere near 10%. It's about half that. Most rich parents pay through higher house prices in the catchment areas of the best state schools.
August 3, 2026 at 10:48 AM UTC It's hard to do direct comparison because they measure different things, but the general picture is that punctuality for regional trains is similar, but ICE and IC are far worse than UK long distance trains
August 3, 2026 at 8:11 AM UTC If the car wash was 150m away and not drive through I'd absolutely go home
August 1, 2026 at 8:10 AM UTC People are using the same chat bots available to common folk to solve these decades-old maths problems. They're bad at basic arithmetic, but amazing at advanced maths. Humans don't use mental arithmetic to solve these either
chatgpt.com
Check out this chat
Here's a chat someone thought you'd want to see.
Global child mortality rate dropping from around 14% to below 4%. One in seven of all children used to die before they were 5!
August 1, 2026 at 4:30 AM UTC Oof. I'm on holiday at the moment but I'll raise it next week when I'm back.
July 31, 2026 at 8:03 PM UTC I think the real problem is that for most people, AI means Google AI overview, or Copilot in Word, rather than anything actually capable. You don't need a ton of resources to try asking the free ChatGPT or Claude for help with your trickiest problem and see what it can do. Something actually hard.
July 31, 2026 at 2:38 PM UTC It's solving decades old maths conjectures and finding vulnerabilities in the most thoroughly audited, critical software. How much more ready does it need to be?
July 31, 2026 at 2:16 PM UTC It's the delegation of powers by central government to local and regional authorities. That's devolution.
July 31, 2026 at 11:49 AM UTC
