All Posts

Ray Exworth spent sixty years sculpting in seclusion with salvaged materials. In the article A sculptor of probable genius, David Heseltine describes a visit to Ray's sheds by the Arts Council in 1984:

… the Arts Council Purchasing Committee, visiting Cornwall, set off for the interior in search of an ‘irascible and difficult’ person who some said had abandoned his work altogether or dug it into the ground. They found what they described as ‘a smallish muscular person in a knitted fisherman’s cap’. John Spurling reports on this visit.

He had not buried his work, or only metaphorically. It was contained in seven different outhouses, a mass of plaster, wood, papier-mâché, wire and I don’t know what, modelled, carved and partly assembled to fill, like the visitors, every available cranny. In fact most of it was one work with a circus theme, or would be if it could ever be brought together into a sufficiently large space – a circus tent for example. We viewed it separately, in groups of two or three, with increasing amazement and admiration, crossing and re-crossing the yard, and learning from one another that there was yet another out house that we hadn’t seen, edging our way in and out of half open doors, past large, fragile plaster figures trying to exclude the two cats who weren’t supposed to enter in case they knocked something over.

These stories of dogged determination, creation for the sake of creation, intrigue me, no matter the field. Some ready examples: Lazarus Lake and his dastardly ultramarathons, Jerry Gretzinger and his eternal map, Justo Gallego Martínez and his cathedral, Nek Chand and the divine kingdom of Sukrani.

Our benchmarks run in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries.

The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database. All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.

While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access.

After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation. In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers. OpenAI’s security team discovered this anomalous activity internally.

— OpenAI, “Hugging Face Model Evaluation Security Incident”

Sometimes exploring the simplest of concepts sends you on the silliest of goose-chases. Say you were thinking of doing some simple mathematics. You start plotting some numbers around a circle — maybe 40, maybe 300. That looks a bit boring. It's just a circle. So you start connecting the numbers. In order not to fall into immediate chaos, you just connect each number to its own double. 1 to 2, 2 to 4, … .

Halfway through you are thrown for a loop: you want to connect 151 to 302, but you only have 300 dots. So you divide by 300, keep the remainder (modulo), and keep looping further. You are finally finished drawing all of your straight lines. You take a look at what you've wrought — a smooth curve you never drew a single stroke of — and you yell out Eureka! A Kallipygoid! You've accidentally created a version of Aphrodite's behind. Then your mom knocks on your door and asks what you're drawing. You quickly give the page a 180-degree turn and tell her you've made a heart, a cardioid, just for her.

Times-table caustic — interactive with JavaScript; drag the multiplier to morph the curve.

This pattern is usually credited — via Mathologer — to Simon Plouffe. The video carries on into the nephroid and the link to the Mandelbrot set. Many have drawn it since.

The mood in the Belgian media is maudlin. The golden generation is being buried for the seventh time in four years. I didn’t feel this game merited such a premature wake. Admittedly, I watched much of it groaning (a combination of a slightly sore head and stomach and an unhealthy dose of stress) but I thought that, despite what the momentum scores below suggest, we were in with a chance until the end.

It was nice to discover some quality in the next generation. Lukebakio, Saelemaekers, Raskin and Moreira all pleasantly surprised me with their work rate and inventiveness. So I won't try to bury something that never quite existed. I will try to enjoy watching the team of the future take shape. You can only be disappointed by results if there was potential in the first place. I'll be happy if the Red Devils keep playing themselves into positions from which they can disappoint me, in the finest Belgian tradition.

Spain had the run of play in 82 of 90 minutes to Belgium's 8, and won 2–1.

Belgium had the run of play for the first 3 minutes. Then Spain retook it and never relinquished it across the entire game.

The default view is Lamine Yamal (Spain) versus Brandon Mechele (Belgium), who shared all 90 minutes.

No need to waste too many words on this. The USA-Belgium (1-4) game was preceded by Trump-Infantino shenanigans, but you always have to beat the team on the field. And if you can't beat the USMNT at their full complement, you don't deserve to go to the next round.

Even so, I can't deny that this result has an extra frisson of satisfaction.

Belgium had the run of play in 49 of 90 minutes to the USA's 41, and won 4–1.

Belgium held the run of play; the point sank as deep as −786 by the 44th minute and, though the USA had spells, never climbed back to level — the board followed, 1–4.

The default view is the midfield duel Weston McKennie (USA) versus Youri Tielemans (Belgium), who shared all 90 minutes — two men working the same central patch of grass.

There was no way I would be able to sleep after the adrenaline-rush-inducing end of the Belgium–Senegal game on Wednesday night. So I took a closer look at some of the statistics commentators keep talking about. SofaScore publishes "Attack Momentum," a model index of per-minute pressure for either team. Given that my stomach was still churning from the nerves of the game, I expected this to be a Senegal-only affair, with the occasional flicker of Belgian momentum.

Belgium had the run of play in 67 of 120 minutes to Senegal's 53 — yet trailed 0–2 until the 86th minute.

Surprisingly — to me — the stats disagreed with my gut. Belgium won more of the minutes (67 to 53), and edged the total momentum over the game too (1,184 to 1,079): the better side, by both measures.

This friction between the statistics and my frayed nervous system pushed me to look at the data another way. Plot the cumulative momentum — the running total, minute by minute — and a completely different picture emerges:

The run of play sank to Senegal −581 by the 55th minute, then Belgium hauled it across zero at the 89th — the minute they equalised — though the board had read 0–2 since the 51st.

Through this tug-of-war we see the real dominance of Senegal for most of the game, at its peak in the 55th minute, only for the tide to turn, perhaps cruelly, to the Belgian side for the first time in the 89th minute, culminating in a fierce fight to the finish.

Satisfied that my feelings of fear were justified, I decided to have a look at player positioning during the game as well. Slightly outraged that there still isn't a readily available, open database of per-match player positions — surely someone's on this in the age of SAM 2 and a bit of homography? Hit me up at hi@anthrosol.com — I had to make do with per-player position samples.

The default view is the flank duel Timothy Castagne (Belgium) versus Sadio Mané (Senegal), who shared 94 minutes — about 86% of Mané's core range sits inside Castagne's.

Now to brace myself for another sleepless night on Monday, ready to be put through the wringer for the Belgium–USA game. It will be worth it as long as it ends in a kerremesse in d'helle for our Red Devils.