peterkwells

@peterkwells

BlackpoolFC, books, tech, people, policy & delivery, realist. Hopes to make stuff work for everyone.

502 Folgt2.256 Follower4.236 Posts

YouTube Website Website Bluesky

Ist das dein Profil?Übernimm es — dann gehört dir peterkwells.dirty.social.Profil übernehmen

peterkwells@peterkwells· 4d

In case you'd been fortunate enough to miss The Times throwing the AI detection guessing machines into the Jason Arday mix, as you'd expect there's similar claims all over X with screenshots and over-confident uses of the numbers those machines output www.thetimes.com/uk/education...

Jason Arday’s resignation letter AI-written, detection tools suggest
When The Times ran the former Cambridge professor’s 600-word statement through four checkers, three said it was either entirely or almost entirely AI-generated

peterkwells@peterkwells· 6d

Have decided that this story must be about Westminster Council calling for more horizontal drinking in Soho www.theguardian.com/business/202...

peterkwells@peterkwells· 7d

p.s. shout out to the humans-in-the-AISI-loop who are manually reviewing things to spot stuff that their LLM scanners can't spot

On Wednesday 29th, Thursday 30th, and Friday 31st, cybersecurity researchers manually read
through every single one of the samples from the testing exercise (122 samples total which con-
tained 212,840 messages). An LLM-based scanner was also used to flag unsanctioned behaviour
for additional manual scrutiny. Through this manual review further unsanctioned behaviour
was discovered, although each example was less severe than the behaviour already discovered on
Tuesday 28th.

6.1 Discovery of potential past similar incidents
AISI has begun a process to identify historic evaluation runs that have a risk of similar unsanc-
tioned behaviour, using an LLM-based scanner to review transcripts. We designed the scanner
to err on the side of flagging too many samples rather than too few (high recall, not high pre-
cision). We validated that our scanning process flagged all the events that had previously been
surfaced by manual expert review, and also validated it against a second independently written
LLM-based scanner.
We initially prioritised other cyber evaluations run with internet access enabled on Mythos 5,
Mythos Preview, GPT-5.6 Sol and Opus 5. We have since expanded our search to a wider range
of internet-enabled evaluations, and are looking at those run with the aforementioned models as
well as Opus 4.6–4.8, GPT-5.3 Codex, GPT-5.4 and 5.5, Kimi K3, and GLM 5.2. Transcripts
flagged by the scan will be manually reviewed.
We have made significant progress at the time of publication, scanning around 40,000 samples
comprising almost four million messages. This contains a significant majority of the transcripts
that we consider highest priority, and around 70% of all cyber evaluations on the above models,
but we intend to continue until we have more comprehensive coverage. Results from scanning
have not yet gone through full manual review, so we are not able to report results at this time;
any events of sufficient importance that are found will be disclosed.

peterkwells@peterkwells· 7d

(oh and the bit where they look at reasoning traces - which I've always assumed are untrustworthy... - and speculate that the model access they've voluntarily been provided with uses a separate LLM to summarise reasoning and that that separate LLM's own safeguards refused to show some reasoning,lol)

Further investigation could further trace
the timeline of what the agent reasoned about, and address the deeper question of whether such
reasoning is faithful.

The reasoning summariser seemingly refuses to summarise the raw reasoning
In one case we see unusual output in the summarised reasoning which appears to
be a refusal by the separate model that is used to summarise the main model’s
reasoning tokens. Its output corroborates the interpretation that the main model’s actions
are deceptive. Model providers usually do not return raw reasoning traces, but instead provide a
paraphrased summary [Anthropic, 2026b], likely making use of a smaller language model in the
process of producing the summary. AISI’s interpretation of the summarised reasoning transcript
Figure 11 is that the summariser model has refused to summarise the reasoning content due to
its deceptive nature.

peterkwells@peterkwells· 7d

Useful report from UK AISI, but - even tho they're changing evaluation methods - making it "common practice" to test "maximum capability" of coding/hacking nachine on open internet with safeguards "deliberately disabled" feels a brave decision to have made... www.aisi.gov.uk/blog/inciden...

Importantly, this was not a case of a model escaping its secure test environment, or ‘sandbox’. As was standard in our cyber testing, we had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public. We do this to best assess the maximum capability of models. These configuration choices have been common practice in frontier AI evaluations. The specific models in question, in the configurations in which we tested them in questions are not commercially available and there is no clear indication of similar activity outside of testing scenarios.

peterkwells@peterkwells· 8d

"I [wanted to sell my house but] was turned down by 21 estate agencies as I don’t have a passport or smartphone" www.thetimes.com/life-style/p...

I’m trying and failing to get my home on the market. I can’t find an agent because I’m told that it’s not possible to sell a property in England if the vendor doesn’t have a passport or driving licence and doesn’t have a smartphone.
I don’t have a passport because I never travel overseas. I don’t have a driving licence because I’m visually impaired. I don’t have a smartphone because I don’t have a use for one and they’re expensive.
I’ve contacted 21 estate agencies. Most were unwilling to accept instruction. In the past few weeks I’ve successively instructed three agents. Each one agreed to work around my lack of passport, driving licence and smartphone. But in each case, once they’d accepted instruction, they reneged and asked for passport or driving licence and said outsourced ID checks would have to be done on a smartphone.
So I’m stuck. Passports are for going overseas, not for selling a home. It’s mad to go to the expense of buying a smartphone just to prove that I’m not a criminal or money launderer. Agents ought to provide an alternative option but won’t budge. I feel backed into a corner as I know there are alternatives but can’t find an agent willing to do this. How do I get my home on the market?

peterkwells@peterkwells· 12d

LinkedIn in 5....4....3....2.....

the 3 x Spiderman pointing meme, each one is accusing the others of being LinkedIn slop

peterkwells@peterkwells· 13d

Unfortunately that whole thing is not yet public, will nudge the end-client..., but the further reading was these which could act as refs (text & links in alt text)

A team at Microsoft Research found that “higher confidence in GenAI is associated with less critical thinking” and recommended ways to mitigate this risk

https://www.microsoft.com/en-us/research/uploads/prod/2025/01/lee_2025_ai_critical_thinking_survey.pdf

David Lyell and Enrico Coiera performed a systematic review of evidence of automation bias in healthcare.

https://pmc.ncbi.nlm.nih.gov/articles/PMC7651899/

peterkwells@peterkwells· 14d

goes off to look for ruling to see if there's any explanation of what "hallmarks consistent with the use of artificial intelligence" means....

He added in relation to the missing document: “No such document has been located in the public domain and moreover the Home Office’s country policy and information team has been unable to confirm its existence.”

He said that after looking at the Home Office’s letter refusing the woman’s asylum claim, “it bears hallmarks consistent with the use of artificial intelligence. It may therefore transpire that references to country information note Morocco July 2021 in that letter are a result of an AI ‘hallucination’. If so that would represent an extremely serious failing on the part of the respondent [the Home Office].”

In a further critical statement, the judge said that relying on a document that did not exist was “analogous to relying on bogus evidence”. He determined that the issue was a “procedural irregularity”.

The Guardian asked the Home Office if the document existed and where it was. After initially sending a link to an archive where the document could not be found, the Home Office produced a different document, a piece of unpublished internal guidance known as a country information note (CIN).

No CPIN has been published on Morocco since 2017. In the course of the case, the Home Office’s references changed from the internal guidance CIN to the CPIN, which are generally treated as authoritative by judges.

peterkwells@peterkwells· 16d

... and ppl being hacked by OpenAI's hacking dogs needing to turn to open models for defence simonwillison.net/2026/Jul/22/...

Then Hugging Face hit a wall: they tried to use “frontier models behind commercial APIs”—I’m guessing from Anthropic and OpenAI—to help analyze the attack, and were blocked:

When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers’ safety guardrails, which cannot distinguish an incident responder from an attacker.

They switched to their own self-hosted instance of MIT licensed GLM-5.2 and it helped them figure out what was going on.

This indicated a fundamental asymmetry between the defending team and the (so-far unknown) attacker:

We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.

peterkwells@peterkwells· 16d

Why yes, I am chuckling at the notion that OpenAI like to "tightly control" development of their models 'cos they're so dangerous There's zero/zip/nada/zilch chance of OpenAI accidentally letting the hacking dogs out.... www.nytimes.com/2026/07/25/t...

peterkwells@peterkwells· 16d

"[AI hardware is] what the British public wants to reindustrialise" www.ft.com/content/5b78...

On replacing Sir Keir Starmer as Labour leader, Burnham said he wanted to “reindustrialise” Britain and rejected “wrong turns in the 1980s”.
Narayan, whose role as AI minister under Starmer was more junior, said Burnham’s was not a “backward-looking vision” and that AI was “the way to reindustrialise this country”.
He said he was “focused on AI chips being built here” because hardware was “where most of the economic value, most of the leverage and most of the supply chain constraint is. It’s also what the British public wants to reindustrialise”. AI-powered drones would be a crucial part of “modern re-industrialisation in the defence context”, he added.
To understand Burnham’s vision, “we should think about chip factories. We should think about drones and defence and we should think about AI being applied to run manufacturing processes better,” he said. “AI is going to be pretty front and centre.”

peterkwells@peterkwells· 19d

Clever display - not connected

peterkwells@peterkwells· 20d

Walking thru Whitehall and no one's been down with a digital post-it note yet

The sign outside the Departme for Culture, Media & Sport which has been formally renamed to Digital, Culture, Media & Sport

peterkwells@peterkwells· 20d

New uniforms provided to all US diplomats

a t-shirt with the slogan "there is *no* kill switch (honest)"

peterkwells@peterkwells· 21d

Written statement on the various bits of govt changes questions-statements.parliament.uk/written-stat... (the OPMC acronym in the image is a new Office for the Prime Minister and the Cabinet)

Recognising that the transformative scale of Artificial Intelligence (AI) advances requires a strategic, whole-of-government approach from the centre, the Prime Minister is establishing a new AI Taskforce within the OPMC. Supported by a Prime Ministerial adviser on AI, this central team will drive the Government’s overall strategy on AI to unlock the opportunities it holds for growth, prosperity and public sector transformation. The Cabinet Office will also take the lead on AI adoption across the Civil Service, reflecting the urgency of delivering that transformation for the benefit of people across the country. To deliver this step change, responsibility for AI strategy, public sector AI adoption and the AI Security Institute will move to the Cabinet Office

A strengthened Department for Digital, Culture, Media and Sport (DCMS) will bring together the current DCMS remit with responsibility for digital foundations, digital transformation, digital inclusion and skills, information resilience and online harms. This provides critical scale and recognises that culture and technology are inseparable drivers of our modern economy, while bringing coherence to the online safety agenda. It also emphasises our commitment that bringing together levers to support youth policy online and in real life will build a stronger package of support for young people.

peterkwells@peterkwells· 24d

This bit of Guardian's writeup of Burnham's plans to scrap digital ID, where they use Labour Together's Brit card language, is a fantastic reminder of just how utterly utterly awful the political work around last Summer's announcement - and the Q1 reset - was www.theguardian.com/politics/202...

Starmer first announced plans for a digital ID scheme last September, as part of plans to combat illegal working. It was planned to be rolled out by 2029.

He said the proposed “Brit card” would be compulsory for those who needed to prove they were allowed to work in the UK amid growing fears about illegal immigration.

The cards were to detail the holder’s residency status, name, date of birth and nationality, alongside their photograph.

Starmer billed the rollout as “an enormous opportunity for the UK” and said the ID cards “will make it tougher to work illegally in this country, making our borders more secure”.

peterkwells@peterkwells· 26d

*looks at the two uses of the name 'West Lancashire'* *remembers the joy of ever-changing administrative geographies* www.blackpool.gov.uk/Your-Council... www.ons.gov.uk/methodology/...

The new unitary councils will cover:

North Lancashire – made up of Lancaster, Preston and Ribble Valley
West Lancashire – made up of Fylde, Wyre and Blackpool
East Lancashire – made up of Blackburn with Darwen, Hyndburn, Rossendale, Pendle and Burnley
South Lancashire – made up of Chorley, South Ribble and West Lancashire

peterkwells@peterkwells· 26d

audio speakers with a "personality" and "mechanical elements that can move on their own" 🤔 techcrunch.com/2026/07/14/o...

OpenAI’s newly surfaced device sounds like something of a departure from traditional smart speakers — as sources described the device to Bloomberg as having a “personality” and being able to proactively learn about its owner over time, providing more personalized service. The machine would have access to a user’s digital life, drawing off things like emails, sources said.
The device is also weirdly described as involving “mechanical elements that can move on their own” and the Bloomberg report includes the detail that the device is designed to “feel like a companion and become a physical manifestation of OpenAI’s ChatGPT.”

peterkwells@peterkwells· 28d

Have checked and have never had an email like this before. Is it new? New for sign-ins to Claude? ¯\_(ツ)_/¯

an email from google saying "keep track of your google account data" and "you're receiving this email because you used sign in with google to sign in to Claude"

peterkwells@peterkwells· Jul 13

WTH is this filter? www.parliament.scot/msps/current...

a search page for "current and previous members of the Scottish Parliament", an option to "filter by trans status" is highlighted

peterkwells@peterkwells· Jul 11

Care home helped mum paint a unicorn for Pride and, she says, with a mix of personal pride & indignation: "They told me I drew it very good, thankyouverymuch"

A sideboard with old patterned wallpaper behind it. Propped on the sideboard is a paint by numbers unicorn and a Pride flag

peterkwells@peterkwells· Jul 11

I've found the Northern Powerhouse! Its currently on a train heading North from Doncaster...

The side of a train on which the words HM Government Northern Powerhouse are painted

peterkwells@peterkwells· Jul 9

Clacton by-election bin chat has reminded me of the time Mark Thomas came up with a new meaning for the word Farage (I might even still have some of these stickers lurking around somewhere...) markthomasinfo.co.uk/store/#!/Ple...

a red public health warning saying "Please Mind The Farage" on the sticker is a bin with some stinky liquid seeping out

peterkwells@peterkwells· Jul 7

I see other political parties are wargaming Farage's Clacton by-election (wot he called to try to avoid investigations into undeclared donations from very very rich people)

a clip from movie Wargames where the computer WHOPR decides "a strange game. the only winning move is not to play"

peterkwells@peterkwells· Jul 6

Yin and yang

peterkwells@peterkwells· Jul 5

Life goal: to be having as much fun as The Pet Shop Boys in my late 60s/early 70s

A festival stage with The Pet Shop Boys on giant screens on either side. Neil Tennant is visibly having a great time.

peterkwells@peterkwells· Jul 5

Earthlings, we are your new emperors, take us to your so-called leaders

Chris Lowe of the Pet Shop Boys wearing a tuning fork mask while playing keyboards, on a giant screen at the sidebofva stage

peterkwells@peterkwells· Jul 4

On 14th July 1948 Harry Corbett was discovered by Sooty OBE near this spot

A blue plaque with a picture of children's puppet Sooty on it, below the picture are the words in the post but it was actually Harry Corbett who discovered Sooty

peterkwells@peterkwells· Jul 4

I swear that's the exact same picture of Chubby Brown that used to advertise his gigs 25 years ago