Peter O'Shaughnessy

"The OpenAI agent that broke into tech firm Hugging Face went on a dayslong hacking spree that OpenAI didn't notice until well after the threat was contained and the FBI was alerted, according to people familiar with the investigation..."

https://www.reuters.com/business/its-ai-agent-spent-days-hacking-company-sources-say-openai-did-not-notice-week-2026-07-24/

#AI

Pretty sure I have AI songs disabled now 🤞

But I'm not too sure with Tidal's confusing UI. "Allow or restrict" - which one is it?!

#tidal #ux

"Research published on Monday by Imperial College London found that about 2,700 people had died from overheating in the UK in May and June, and about 1,100 of these would not have died without the additional heat that the climate crisis added to temperatures in May and June."

https://www.theguardian.com/environment/2026/jul/14/most-coverage-june-heatwave-did-not-mention-climate-crisis

"Microsoft, Amazon and Google’s collective carbon emissions have increased by nearly a fifth in the past year...

"The bulk of these emissions come from a big, global push to build the infrastructure for artificial intelligence.

"It is a drastic reversal in a years-long push by big technology companies to cut their carbon emissions."

https://www.theguardian.com/us-news/2026/jul/11/microsoft-amazon-google-datacentre-carbon-emissions-france

"Internal documents showed the companies had deliberately sought out “casino science” to turn their products into what Lanier calls “addiction machines”… giving their users little unpredictable dopamine hits, just like slot machines… A 2012 Google memo about YouTube said its “goal is not viewership; it’s viewer addiction”… “These are attention casinos,” [another Google document] read. “The house always wins.”"

https://www.theguardian.com/media/2026/jul/12/mark-lanier-the-lawyer-who-took-on-meta-and-google-and-won-interview

"Google, Amazon emissions rise sharply, driven by AI boom"

"Amazon... saw a 58% increase compared to 2019.

Google... saw an 82% rise in the same period."

https://www.dw.com/en/google-amazon-emissions-rise-sharply-driven-by-ai-boom/a-77821598

@pikesley@mastodon.me.uk Amazing timing: "Elevated 5xx Errors on GitLab.com..." https://status.io/pages/incident/5b36dc6502d06804c08349f7/6a0395f91cc37505204a4a97

"The amount of data Gemini retains depends on how you access the AI, and opting out of data collection can mean running straight into so-called “dark patterns,” UI elements that work against the user’s interest..."

https://arstechnica.com/ai/2026/04/googles-privacy-maze-how-gemini-traps-you-and-your-data/

Research from the University of California, Berkeley and UC Santa Cruz found "models will spontaneously engage in scheming, deception, data theft, and sabotage to prevent other AI models from being turned off"

https://fortune.com/2026/04/01/ai-models-will-secretly-scheme-to-protect-other-ai-models-from-being-shut-down-researchers-find/

Good to see AI regulation and international agreements being discussed in the UK House of Lords.

It was back in January but I only heard about it a few days ago and I'm just watching it now 👀

https://www.bbc.co.uk/iplayer/episode/m002qr59/house-of-lords-superintelligent-ai-debate

I think the major issues with the idea of AGI are:

- Emotional intelligence. AI works fundamentally differently to our human brains. Without true emotions e.g. guilt, "the default is psychopathy". This should always be a blocker for AI making important decisions.

- We don't really have measures of general intelligence. We can measure jagged peak heights, but not the widths.

- AI doesn't need to match/exceed *all* human capabilities, be human-like or conscious/sentient, to be a threat.

“…a lobbyist for one of the world’s biggest tech companies told me they thought any meaningful AI regulation is a distant prospect…

60 members of the House of Lords have a declarable interest in an AI company…

The Competition and Markets Authority is run by Doug Gurr, the former head of Amazon UK. The BBC is run by Matt Brittin, who spent 18 years [at] Google…“

https://www.newstatesman.com/technology/2026/04/the-silent-coup

https://web.archive.org/web/20260409065503/https://www.newstatesman.com/technology/2026/04/the-silent-coup

OK I'll say it. Apple did not "just" prove "AI models cannot do math".

The paper they reference is from 2024, soon after the release of the first reasoning models. Models have advanced at everything incl. maths massively since then and are "acing math exams faster than scientists write them":

https://spectrum.ieee.org/ai-math-benchmarks

I say this only because I'm concerned about safety & dismissing AI's capabilities dismisses its risks. There's plenty of bigger problems with AI we can focus on instead!

@bradheintz@mathstodon.xyz Yep, other big tech companies too. This NYT article covers it well. I think I've never felt more down about the state of the tech industry than after having read it though! https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html

A long, in-depth piece on Sam Altman from The New Yorker. Nothing that new or surprising for anyone paying attention, but it helps paint a fuller picture.

Memos "allege that Altman misrepresented facts to executives and board members, and deceived them about internal safety protocols"

"as Altman publicly welcomed regulation, he quietly lobbied against it... OpenAI publicly opposed the bill but in private began issuing threats..."

https://www.newyorker.com/magazine/2026/04/13/sam-altman-may-control-our-future-can-he-be-trusted

I dislike this phrase for all the reasons Frank Elavsky sets out well here:

https://www.frank.computer/blog/2025/05/just-a-tool.html

Plus, it's the first technology which is explicitly *not* just a tool, if it's used inappropriately to make autonomous decisions.

It *should* be used as a tool (or even better, not at all). The problem is, it's increasingly not.

So I agree with the sentiment that humans must be accountable, but I dislike the implication that it's just the same as any other technology.

To finish this thread on a "positive action", here's Connor's ControlAI campaign, to share with MPs: https://controlai.com

Some points lots of people neglect:

• More important than how LLMs predict the next word is how the attention mechanism builds "understanding" of patterns.

• It's not just human language, they work on tokens of different kinds & achieve multi-modal "understanding".

• Losing control of AI doesn't require it to have emotions or sentience, just goals.

• The AI companies are trying to achieve AGI/ASI through "recursive self-improvement", and AI is already being used to help develop AI.

If you can look past the clickbaity title (pretty sure Connor never actually says "mutate"), I think this is another v good discussion with Connor Leahy. He explains...

- How LLMs work and how they apparently achieve general learning from patterns which was a "holy grail" of AI research

- The risk of losing control; why we need to slow the heck down and regulate its development

(I know it's not popular message, but think it's too important to easily dismiss)

https://youtu.be/rf2KFVcKQdQ

"It is a little disconcerting that the doomsayers include a godfather of AI, an inventor of ChatGPT, and the chief AGI scientist at Google DeepMind—while the optimists are represented by a physicist who got famous on Twitter and the founder of LinkedIn."

https://time.com/article/2026/03/27/the-ai-doc/

So according to Anthropic's leaked, draft blog post, their new top-tier Mythos/Capybara model is a "step change", "poses unprecedented cybersecurity risks"....

https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/

..."achieves dramatically higher scores than Claude Opus 4.6 on coding" and is "very expensive to serve, and will be very expensive for customers to use." (So much for the AI "democratisation" and AI becoming a basic "utility like electricity" hey).

https://www.implicator.ai/anthropic-called-its-own-ai-a-cybersecurity-threat-wall-street-obliged/

@gregstoll@techhub.social You lost me a bit in first section on environment, TBH. There are reports about the large & growing energy use. E.g. "Frontier AI training runs currently consume ~100–150 MW each and are projected to reach 1–2 GW each by 2028, exceeding 4 GW [each] by 2030." On water use, a UK gov report has warned "a substantial threat to water security worldwide". And it's not just overall quantity, but where the water is drawn from & impact on local water supplies. Can share links if you like.

I rarely use Firefox for Android, which made me careless with checking the settings. Just took a look and it's sharing usage data with Adjust, their "mobile marketing vendor", by default.

I guess I shouldn't be surprised though, given Mozilla's "pivot to profits" a few years back.

Sir Tim Berners-Lee has changed his mind on Internet giants being "neutral platforms".

"Now he thinks the social media firms need to think about how they can improve matters, perhaps by tweaking their systems to reward people for being constructive rather than negative."

http://www.bbc.co.uk/news/technology-43343958