iiinova — where your digital self keeps evolving

Thomas House @tah-sci.com · 15/09/2026

I think that's when it's most relevant - something that's not got any substance to it will seldom have any impact. It's when a substantial point is combined with humour, polemical writing and the general human desire to pile on people that you get an uncomfortable dynamic.

010

Thomas House @tah-sci.com · 14/09/2026

Basically there is a good reason why academic writing tends to avoid being too definitive - the Internet pushes in the other way to everyone's detriment.

220

Thomas House @tah-sci.com · 14/09/2026

Despite the expletive, it makes a clear argument - but the issue with a lot of "dunks" on all sorts of analysis methods is they're highly polemical, not constructive. Lots of systems do have unobserved categorical variables - do we try to work out what to do about that or post the relevant XKCD?

110

Thomas House @tah-sci.com · 14/09/2026

So, if the approaches to finding these aren't delivering quite as we'd like, or often end up "low medium and high", it's not entirely clear where the problem lies. Just my two pen'orth.

110

Thomas House @tah-sci.com · 14/09/2026

So I conclude that when criticism of a particular method gathers up a head of steam, it seldom leads to overall improvements in methodology, but instead it's mainly a destructive process. We know that there's often unobserved categorical variables and that matters an awful lot.

120

Thomas House @tah-sci.com · 14/09/2026

Basically, people love a dunk. I've followed a few very negative scientists who spend their whole time criticising others, in different fields over the years and not a single one of them could accept even the mildest criticism of their own (often flawed because human) work gracefully.

120

Thomas House @tah-sci.com · 14/09/2026

So ... Sometimes methods are over-used or over-hyped or ... But the cure can be worse than the disease. How many positive examples of logistic regression are there? The true connoisseur can destroy almost every published one somehow. Compared to that benchmark, how does cluster analysis do?

120

Reposted by Thomas House

Paul Mainwood @paulmainwood.bsky.social · 13/09/2026

He said we keep launching moral startups to beat capitalists to AGI, but progress just accelerates due to competition. So I asked where they got founders and he said the AI safety community and I said it sounds like you’re feeding AI safety researchers to capitalism and his daughter started crying.

031

Thomas House @tah-sci.com · 13/09/2026

> heads of AI companies are relatively publicly sympathetic to this worldview. If you think it's more like "one of these numpties might make a chatbot that can break something important", then more openness and standard regulation, which the companies largely don't want, is clearly preferable.

050

Thomas House @tah-sci.com · 13/09/2026

This thread brilliantly explains the real security issues with AI. To develop it further, if you believe it's a "Terminator" type situation, you will see the "rogue" AIs and conclude that more openness - creating extra dangerous entities - is actually a bad thing, and that's why the >

2112

Reposted by Thomas House

Jess Miers 🦝 @jmiers230.bsky.social · 13/09/2026

I did a somewhat deep dive into the recent OpenAI and Anthropic "hacks" for an upcoming interview. IMO, the AI companies are better off running with the "rogue AI" story because the real story is actually pretty embarrassing. Like making your Wi-Fi password "wifi" embarrassing. 🧵

341835794

Thomas House @tah-sci.com · 12/09/2026

Great thread - I lived for some time in Coventry and while there was a big planning mistake made in designing it for the car, two tone was a famous example of a quite special civic culture.

010

Thomas House @tah-sci.com · 12/09/2026

If that's right, then she would be the only one that Kirk would have believed went to heaven - most of the US right believes the doctrine that no matter how wicked your life is (and presumably they believe suffragism was wicked) salvation is by faith alone.

010

Thomas House @tah-sci.com · 12/09/2026

I thought after Bin Laden it's Nietzsche, Freud and a suffragette, possibly Emmeline Pankhurst.

220

Reposted by Thomas House

Ben Recht @beenwrekt.bsky.social · 12/09/2026

The right way to pace the frontier is to commit to a robust, open-source AI ecosystem, thereby diluting the power of these badly behaving frontier labs.

5599

Thomas House @tah-sci.com · 12/09/2026

I mean, sadly they often do, but we all ought to be wise enough to the tricks they use to achieve that outcome to avoid the blatant positioning here.

010

Thomas House @tah-sci.com · 12/09/2026

(not for AI in my case but for things like pandemics)

000

Thomas House @tah-sci.com · 12/09/2026

Agreed, I have been involved in official risk / contingency assessment and this shouldn't involve "number theatre" when uncertainty is large. bsky.app/profile/tah-...

140

Thomas House @tah-sci.com · 12/09/2026

This (sometimes called p(doom)) seems like mainly arbitrary guesses - it's not how serious contingency planning is done and mainly seems to distract from very predictable harms e.g. to infrastructure through hacking or human activities like mathematics that are valuable.

020

Thomas House @tah-sci.com · 11/09/2026

We had a departmental meeting about it all, with a particular focus on PhD study but of course more general talk, today, and honestly no-one is really very confident what to do about it all.

021

Reposted by Thomas House

Brian Grubb @briancgrubb.bsky.social · 10/09/2026

THE PREMISE IS THAT THE MUPPETS ARE PUTTING ON A SHOW CALLED THE MUPPET SHOW

474793953

Thomas House @tah-sci.com · 10/09/2026

I don't think "Skynet becomes self aware" is going to happen, but it's kind of irrelevant how plausible one finds that because it's also covered by the kind of regulations that we clearly unequivocally need.

110

Thomas House @tah-sci.com · 10/09/2026

Very much this - tech billionaire weirdos develop software that breaks some key bits of critical infrastructure looks all too plausible.

120

Thomas House @tah-sci.com · 10/09/2026

I think there are issues but from what I know they are ones these Governments would be able to solve if cooperating - and if done like CERN etc. as an open project with proper research ethics I expect that would attract talent.

020

Thomas House @tah-sci.com · 10/09/2026

This is one of the modern world's major problems ... :(

010

Thomas House @tah-sci.com · 10/09/2026

It is public knowledge that the UKHSA has an internal model called "Janus" for example, and it doesn't seem to bankrupt them, they are a major Government agency but not super rich. www.gov.uk/government/p...

gov.uk

UKHSA Advisory Board: Artificial Intelligence

020

Thomas House @tah-sci.com · 10/09/2026

The feedback I got from a meeting I was in today is to have something about as useful as a standard non-frontier model requires the resources of a national Government department, so would need to be a cooperation between Universities at this stage, but we might expect that to reduce.

220

Thomas House @tah-sci.com · 10/09/2026

Consider computer algebra - a true Butlerian would never use it, but most of us are happy with the different software packages that will do things like solve a cubic or look up the right special functions for a differential equation provided these function like we expect software to function.

020

Thomas House @tah-sci.com · 10/09/2026

I'm increasingly thinking that there should be more thought given to bespoke training of open weight models - this is what various health researchers are doing to work on sensitive data. So many of the problems come from the oddness of the big tech multi-billionaires.

430

Reposted by Thomas House

Grant @syncingsinking.bsky.social · 10/09/2026

This is just to say I have solved the open questions that were in your paper and which you were probably saving for your phd students Forgive me they were delicious so incentive-laden and followed so readily once your postdoc's drafts mysteriously found their way into my training data

4356114

Thomas House @tah-sci.com · 10/09/2026

I'm calling it: Trump's dividend will be in Crypto, probably Trumpcoin.

030

Thomas House @tah-sci.com · 09/09/2026

In the sense that, we don't expect health problems to be like physics where we'll get a beautiful simple theory - the best we probably can hope for in many cases is solution like Stockfish "solves" chess, which probably requires some level of data-driven automation.

000

Thomas House @tah-sci.com · 09/09/2026

Maybe we need a different word for the levels of hype going around from Anthropic and OpenAI, because it feels like a qualitatively different phenomenon from the previous somewhat justified optimism about data in health research.

110

Thomas House @tah-sci.com · 09/09/2026

Yes, the understanding is that AlphaFold didn't solve the problem in the way we'd hope a human would, with some brilliant explainable insight that simplified the physics - but it's a real achievement, not the breathless hype we're drowning in about imminent cures for cancer.

100

Thomas House @tah-sci.com · 09/09/2026

We know how the folding process happens though - it's physical laws that we have absolutely no hope of ever solving directly, and no good reason to think that some human comprehensible heuristic can represent accurately enough to predict the outcome of.

100

Thomas House @tah-sci.com · 09/09/2026

Example of different context: en.wikipedia.org/wiki/Hamilto...

en.wikipedia.org

Hamiltonian Monte Carlo - Wikipedia

010

Thomas House @tah-sci.com · 09/09/2026

I started half remembering this elsewhere in the thread - it's the symplectic properties of Hamilton's equations that make them useful in completely different contexts, not e.g. energy conservation, I think this property is under-appreciated.

120

Thomas House @tah-sci.com · 09/09/2026

I'm not sure it's so cheap because you can see things like the symplectic structure of Hamilton's equations, I can't remember the deep theory but people use the unique physically motivated properties of these equations in other contexts so I keep coming across them.

020

Thomas House @tah-sci.com · 09/09/2026

I know a lot of serious mathematicians who find that the better LLMs can help them identify relationship with other papers, potential errors in arguments etc., but with significant guidance from them - they all say it's easy to get nonsense but they're genuinely useful.

020

Thomas House @tah-sci.com · 09/09/2026

This observation seems relevant for discussions around AI safety.

000

Thomas House @tah-sci.com · 09/09/2026

In popular British TV show, "Robot Wars", participants used to build robots with all kinds of weapons to fight each other in an arena, meaning no humans got hurt. If they'd been irresponsible with how they deployed the robots, they could have hurt someone without the robots being intelligent.

120

Thomas House @tah-sci.com · 09/09/2026

It's always hard to know - I expect they'd all be fascinated by the story around general functional learning, training through backpropagation, encoder-decoder architecture etc. and a bit sad about the sociology.

040

Thomas House @tah-sci.com · 09/09/2026

No, the recent ones are not bits of maths with immediate applicability. Even Navier Stokes, which are very useful equations, is known not to be a good model of a real fluid when it "blows up".

120

Thomas House @tah-sci.com · 09/09/2026

OK so here is the horrible calculation that I really upset myself with. Connectivity with wild reservoirs changes significantly since 2000 or so, after which we have 2-3 coronavirus emergences depending on if you count MERS. One per decade, in contrast to influenza's three per century.

020

Thomas House @tah-sci.com · 09/09/2026

When Good and Turing were at Manchester under Newman I believe there was a lot of interest in use of computers for pure mathematics, potentially as a specialism for the University - it's hard not to wonder what they'd all make of recent developments.

110

Thomas House @tah-sci.com · 09/09/2026

I became aware of it from seeing signs around Romford but looking at the actual routes it seems to be quite central in North London and further out elsewhere. Quite why one would push services out to the M25 where by design no-one actually lives clearly requires big Reform brains to answer.

110

Thomas House @tah-sci.com · 09/09/2026

Isn't this basically what the "Superloop" bus does? (Which everyone seems to think is a great thing absent an orbital rail service.)

100

Reposted by Thomas House

Parmvir K Bahia @parmvir.com · 09/09/2026

S4 of #BiologyInNumbers begins with another set of minisodes recorded at the @smbmathbiology.bsky.social/ESMTB meeting. We got to visit the lovely city of Graz to talk about: women’s health, social determination of measles, and a precious baseball card collection... smb.org/news/13686145 #MathBio

032

Thomas House @tah-sci.com · 09/09/2026

"Large Language Model" - it's a more specific description of ChatGPT etc. because there are other "AI" systems that are completely different.

120

Thomas House @tah-sci.com · 09/09/2026

This chippy is next to where I grew up - an area with many Muslims mainly with a family background in Pakistan and a large orthodox community around Moortown. bsky.app/profile/tah-...

120