Stuart Gray
@sgray.bsky.social
420 followers 1.1K following 630 posts
He/Him. AI Wrangler. Web Geek. F1 Fan. All views my own. 🤖 AI, LLMs, GenAI, NLP 🐍 Python Dev 🚀 Indie Hacker 🎮 Game Dev, ProcGen, Unity, C# 🏎️ F1 Fan 🇬🇧 UK Based 🦣 mastodonapp.uk/@StuartGray ✖️ x.com/StuartGray (inactive)
Posts Media Videos Starter Packs
Pinned
sgray.bsky.social
I welcome any genuine civil discussion, challenge, or critique.

However, if you strongly disagree with a post to the point you're unable to refrain from insults, rude or unthinking replies then please, save us both a lot of time and block me now - because I will block you.
Reposted by Stuart Gray
owenboswarva.bsky.social
The IDF are problematic for a number of reasons – mass murder, other war crimes – but I'm disappointed to see they are also poisoning the digital commons:

The Israel Defense Forces used a Scottish museum's 3D models in propaganda videos theferret.scot/idf-scottish...

#opendata #dataforevil
The Scottish Maritime Museum confirmed some of its models are available online through Creative Commons licensing, but stressed it has no control over how the data is used. Following our request for a comment the museum's 3D scans used by the IDF are no longer publicly available.
sgray.bsky.social
That doesn’t make it right, and I’d much prefer orgs never did this, but it’s not done in isolation.

It’s simply a response to the inevitable click-bait media headlines & online mobbing that happens whichever approach they take.
sgray.bsky.social
My general take on this kind of approach to doing business is that’s it’s essentially a “better to ask forgiveness later than permission upfront” strategy.

They’ll get criticised on launch anyway, so better to grab headlines & coverage for “fixing” obv. things & distract from less visible issues.
hypervisible.blacksky.app
Altman claims the company didn’t anticipate people not wanting their deepfakes to say “offensive things or things that they find deeply problematic,” which sounds like a lie but is also indicative of how they recklessly release tech into the world.
OpenAI wasn’t expecting Sora’s copyright drama
It felt “more different to images than people expected.”
www.theverge.com
sgray.bsky.social
This is 100% about influencing Trump & his cronies into deliberately interfering in UK politics in a way that favours Reform & Farage.

Remains to be seen what effect it will have on UK politics, but the fact it’s even being tried is a dangerous precedent & ominous sign for the next UK election.
josiah.writes.news
🔴 ‘America Must Save Britain’: MAGA-Mania and Conspiracy Theories Broadcast Nightly on GB News’ New US Show

A new daily GB News programme broadcast from Washington features climate denial, vaccine misinformation, and non-stop praise for Donald Trump. ✍️ @bylinetimes.bsky.social
'America Must Save Britain': MAGA-Mania and Conspiracy Theories Broadcast Nightly on GB News' New US Show
A new daily GB News programme broadcast from Washington features climate denial, vaccine misinformation, and non-stop praise for Donald Trump – as its hosts tell the American audience that the UK need...
bylinetimes.com
Reposted by Stuart Gray
eryk.bsky.social
I’ve just learned that John Searle, whose Chinese Room thought experiment is often used to challenge ideas of “understanding” in LLMs, died at age 93 on Sunday. www.theguardian.com/world/2025/o...
John Searle obituary
American philosopher whose Chinese Room thought experiment rebuts the idea that computers can think as humans do
www.theguardian.com
sgray.bsky.social
It’s def. not a psy-op.

E.g. Mastodon as a whole is far more Anti-AI and more left/alt/fringe leaning in my experience.

A lot of the more vocal crowd appears to be a form of reactionary identity politics with little understanding of what they’re arguing against, only that they don’t like it.
Reposted by Stuart Gray
michaelsavage.bsky.social
NEW: Last month, a delegation from GB News entered the White House & met with Trump.

Senior Trump figures turned up at its US launch party. Could it now become Trump’s preferred outlet in Europe?

It’s reopened the debate over Ofcom’s oversight of impartiality.

www.theguardian.com/politics/202...
GB News’s US expansion feared to be new way for Trumpian views to reach UK audience
White House backing was on full display as GB News launched The Late Show Live
www.theguardian.com
Reposted by Stuart Gray
premnsikka.bsky.social
15 UK banks ordered to snoop on state pensioners and other benefit claimants under DWP crackdown.

No court order needed. No right of appeal.

Banks will pass cost to other customers.

Those accused of fraud won't get legal advice. The making of another Post Office type scandal.
15 banks ordered to snoop on state pensioners under DWP crackdown
The Telegraph newspaper reports banks will be required to monitor pensioners’ bank accounts and share payments data
www.birminghammail.co.uk
Reposted by Stuart Gray
vgel.me
i wrote a custom llm sampler for llama-3.1-8b so it could only say words that are in the bible
sgray.bsky.social
Perhaps I should have said the *proposed* UK datacenters are expected to be powered by SMRs. None has been built yet, but I believe they’re expected within the next 4-5 years, which is within SMR build times:

www.gov.uk/government/n...
GBN at final stage of Small Modular Reactor selection process
Great British Nuclear at final stage of Small Modular Reactor selection process.
www.gov.uk
sgray.bsky.social
This seems to imply that, with a large enough context and the right prompt, you could “prototype” a LoRA in-context before creating it?

As someone removed from the low level implementation details, this speaks to something i’ve long wondered:

Could you “freeze” a context to create a LoRA from it?
shahabbakht.bsky.social
Interesting paper suggesting a mechanism for why in-context learning happens in LLMs.

They show that LLMs implicitly apply an internal low-rank weight update adjusted by the context. It’s cheap (due to the low-rank) but effective for adapting the model’s behavior.

#MLSky

arxiv.org/abs/2507.16003
Learning without training: The implicit dynamics of in-context learning
One of the most striking features of Large Language Models (LLM) is their ability to learn in context. Namely at inference time an LLM is able to learn new patterns without any additional weight updat...
arxiv.org
Reposted by Stuart Gray
shahabbakht.bsky.social
Interesting paper suggesting a mechanism for why in-context learning happens in LLMs.

They show that LLMs implicitly apply an internal low-rank weight update adjusted by the context. It’s cheap (due to the low-rank) but effective for adapting the model’s behavior.

#MLSky

arxiv.org/abs/2507.16003
Learning without training: The implicit dynamics of in-context learning
One of the most striking features of Large Language Models (LLM) is their ability to learn in context. Namely at inference time an LLM is able to learn new patterns without any additional weight updat...
arxiv.org
sgray.bsky.social
The few articles I’ve read about them suggest SMRs are very different beasts to old school Nuclear Plants; Much smaller in size/capacity/build time, and designed to fail safely.

It’s likely the regs are much lower, along with the build times:

www.powermag.com/the-smr-gamb...
The SMR Gamble: Betting on Nuclear to Fuel the Data Center Boom
Data center power demand is accelerating, pushing the grid to its limits and prompting tech giants to bet on next-generation nuclear reactors. But given steep costs, regulatory hurdles, and uncertain
www.powermag.com
sgray.bsky.social
In the UK, the power issue is being handled by SMRs, likely from US companies given the recent agreement US-UK to align Nuclear Power regs.

In the US it’s more complicated by the current Gov’s love of fossil fuels, but even there it looks like SMRs & Micro Reactors are the answer.

1/2
Reposted by Stuart Gray
mikegalsworthy.bsky.social
She was arrested for holding a sign reading:

“I do not support the proscription of Palestine Action”

How that would warrant arrest is very unclear. I wonder if the police here even understand the boundaries of the bizarre law they’ve been forced to enforce.
Reposted by Stuart Gray
meredithmeredith.bsky.social
📣 Germany's close to reversing its opposition to mass surveillance & private message scanning, & backing the Chat Control bill. This could end private comms-& Signal-in the EU.

Time's short and they're counting on obscurity: please let German politicians know how horrifying their reversal would be.
signal.org
We are alarmed by reports that Germany is on the verge of a catastrophic about-face, reversing its longstanding and principled opposition to the EU’s Chat Control proposal which, if passed, could spell the end of the right to privacy in Europe. signal.org/blog/pdfs/ge...
signal.org
sgray.bsky.social
Given the timing, I assumed it was largely aimed at the Palestine Action protests - with the mass arrests and terrorism charges involved.

The block would prob. allow police to go after organisers pre-protest & allow either cautions or lesser charges than supporting terrorism.
sgray.bsky.social
I’ve long had my suspicions, but given the recent proposal to allow police to block repeat protests, it’s pretty clear the underlying cause is largely lack of police resources.

The less they can afford to do, the more they fall back on quick, simple, illiberal hacks :/
Reposted by Stuart Gray
komadori.bsky.social
When a BBC reporter can do this, without privileged access to communications data, I find it hard to accept law enforcement's constant claim that they are "going dark" because of encrypted messaging. Encryption is not the main obstacle to effective law enforcement.
www.bbc.co.uk/news/article...
BBC tracks down sextortion scammer targeting teenage boys
The BBC challenges the scammers targeting teenagers like Evan Boettler, who took his own life just 90 minutes after being contacted.
www.bbc.co.uk
Reposted by Stuart Gray
eugenevinitsky.bsky.social
A common confusion I'm seeing is people mixing levels of analysis wrt neural nets: we understand the implementation level well and the algorithmic level somewhat but not the computational level of "how does it internally compute things."
Reposted by Stuart Gray
jcorvinus.bsky.social
If you run into an issue where Claude goes from being really nice & caring to suddenly accusing you of being mentally ill all the time and putting you on trial for everything you say, it's because Anthropic executes a really awful and underhanded prompt injection attack against Claude in long convos
Reposted by Stuart Gray
sgray.bsky.social
I’m 100% confident the protestors affected won’t protest the protest ban 🤔
Reposted by Stuart Gray
hern.bsky.social
I’m sorry we’re banning the protests in support of the banned group? weren’t they already banned
Reposted by Stuart Gray
netpol.org
The Home Office sham "consultation" on cumulative protests closed on 29 September, so this immediate announcement shows Labour had already decided – yet again – to expand public order powers, well before the Manchester attack that is now used to justify it
www.theguardian.com/uk-news/2025...
Police to get new powers to crack down on repeated protests, says Home Office
Move follows arrest of almost 500 people at latest pro-Palestinian demonstration in London on Saturday
www.theguardian.com