Alasdair Stewart

@abestew

Sociologist | Socialist | Neurodivergent | Linux & FLOSS advocate | he/they

443 Folgt344 Follower387 Posts

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

The demo is also reliant upon info from TeachFX (see screenshots), a concerning AI product that records lessons & creates summary, being fed to Claude. So 'Claude for Teachers' basically ingests this summary & spits out a verbose version with a suggested next lesson plan & risk of hallucinations.

Alasdair Stewart@abestew· Jul 15

Anthropic released 'Claude for Teachers' & even the promo video shows issues with it. Image 1 = 1st response that was generated, Images 2&3 = 2nd response generated. Some aspects are the same, but also contain critical differences - such as the educational standards. www.youtube.com/watch?v=V-OO...

Alasdair Stewart@abestew· Jun 9

I wonder why OpenAI is joining Anthropic in suggesting a slow down in the area that bleeds them the most money... 🤔

As frontier AI development continues, we expect national and global coordination to become more important. We have long believed there should ultimately be an international organization that helps coordinate leading AI efforts to reduce catastrophic risk. Cooperation and shared safety standards are an important part of the path forward, especially because the incentives around commercial and national competition are hard to escape. One goal of such an organization should be to make it possible for the world to take coordinated action, including slowing frontier development when needed, so societal resilience, safety, and alignment can keep pace.

Alasdair Stewart@abestew· Jun 8

Chapter 8 also looks to be OK. Chapter 1 similarly aside from one segment. So again, if remove those chapters, what is the AI score for chapters 2, 3, 5, 7, and 9. - 37% AI generated with minimal human input. - 3% moderatley AI assisted. - 3% lightly AI assisted.

Alasdair Stewart@abestew· Jun 8

With chapters 4 & 6 removed, Pangram gives the following for the main text of the Milburn interim report: - 23% AI generated with minimal human input or revision. - 7% moderatley AI assisted. - 3% lightly AI assisted.

Alasdair Stewart@abestew· Jun 8

Pangram's graph of potential AI usage across the Milburn interim report. The two long dips map directly onto Chapters 4 & 6. So, whoever wrote those chapters, well-done - your colleagues clearly let you down. Which obviously leaves the question, how much is AI if remove those two chapters? 23%...

Alasdair Stewart@abestew· Jun 8

Buckets of slop.

575. Different institutions have different remits. It is inevitable that no single one can hold the ring on everything from early years to benefit payments. The problem is that each part of the system operates within its own funding stream and accountability framework. When a young person begins to struggle, they do not encounter a coordinated response. They encounter separate systems operating in sequence, each with partial visibility and limited responsibility for what happens next. A school may identify the risk but not carry responsibility beyond exit. A college may record withdrawal but not own re-engagement. A GP may recognise deterioration but not trigger the education and employment response. A Jobcentre may only become involved if a benefit claim is made. Each institution performs part of a function. Fragmentation does not produce a cohesive response.

576. East Midlands County Combined Authority told this review that ‘many young people experience the system not as a coherent journey, but as disconnected interventions that stop and start at arbitrary thresholds.’ Careers England said that ‘one of the national challenges in NEET prevention and reduction is the lack of coordination, there is no national framework, and multiple stakeholders have responsibility for different parts of the provision.’ Ealing Council told us that ‘the current system is fragmented and failing.’

577. What passes for coordination is mostly parallel play. Institutions work alongside each other. They do not work together. There is no unifying goal. There are many objectives that might overlap but are not collectively pursued. And critically, there is no accountability for the outcome as a whole. No authority to direct resources and efforts differently. Education is accountable for attainment. Welfare for benefit flows. Health for treatment. Nobody for the pathway from childhood to working life.

Alasdair Stewart@abestew· Jun 8

Some more (highly likely) egregious AI slop within the Milburn interim report.

679. Nearly one million young people are outside education and work. They are not a statistic. They are the sons and daughters of this country. Some were identified as at risk before they could read. The system knew. It watched. It documented. It published reports. It commissioned evaluations. It launched pilots. It let the funding expire. And it moved on.

680. Those young people did not move on. They are still here. Still waiting. Still paying the price for a country that has chosen, repeatedly and with full knowledge of the consequences, to administer the problem rather than solve it. This review says: enough. Not another programme. Not another pilot. A system. Built around participation. Accountable for outcomes. Permanent in its architecture. With new ladders of opportunity. Funded at a level that treats young people as an investment, not a cost. Resilient to the problems of tomorrow, in a labour market which is likely at the beginning of yet another transformation. And worthy of the generation it is supposed to serve. A new mindset is needed. Our country can choose differently. One that prioritises the next generation. This review demands that it does.

Alasdair Stewart@abestew· Jun 8

AI for some reason often paraphrases percentages needlessly and in way that makes the figures vaguer, such as replacing 24% with "over 23%". See screenshot from report. Actual figures in cited text: lazy and work shy - 23% overly sensitive - 34% entitled - 27% rejected solely on basis of age - 9%

Milburn report:
"Nearly 25% of employers described young people as “lazy and work shy.” More than 33% called them “overly sensitive”. More than 25% called them “entitled”. 10% admitted to rejecting a young person for a job solely on the basis of their age."

Alasdair Stewart@abestew· Jun 8

OK, it looks like there is a lot of AI slop in the Milburn interim report. Signed up for Pangram 7 day free trial. Of the main text: - 13% is high confidence AI generated - 3% is moderately AI-assisted - 1% is lightly AI-assisted 17% accounts for **10,451 words** of the main text.

Alasdair Stewart@abestew· Jun 8

The Milburn led "Young people and work: interim report" reads like genAI had a heavy hand in 'writing' it? Pangram suggests 429-438 - where the main attack on neurodivergents begins - is 100% AI-generated. Even the subtitles - "configured for treatment, not participation" - scream genAI.

Pangram report for 429-438 of the report showing high confidence that 100% of it is AI-generated

Alasdair Stewart@abestew· Jun 5

Note - whilst you can manually turn off the AI bullshit in DuckDuckGo, it's also possible to add a 'No AI bullshit' version of DuckDuckGo as your browser's default search engine as well.

Alasdair Stewart@abestew· Jun 4

Google Search replacing dictionary definitions with AI bullshit was final straw, switched to duckduckgo.

Alasdair Stewart@abestew· May 29

Looks like chick reached safety. All quiet at the back, and the magpies - including short-tail - have returned to their usual hangout spot round the front.

Alasdair Stewart@abestew· May 29

Now it's becoming crows v magpies AND seagulls, the latter clearly assuming the crows are after the seagull chicks rather than the crows merely following their own along the ground near them.

Alasdair Stewart@abestew· May 29

Crow taking issue with me getting photo of chick near window is doing show of strength by attempting to fuck up my satellite dish - little does it know its not been in use for over five years.

Alasdair Stewart@abestew· May 29

Surprise development - the near two hour battle seems to be crows defending a chick thats fallen from the nest.

Alasdair Stewart@abestew· May 29

Spoke too soon, crows retreated merely to plan some bizarre ground offensive - remaining as clearly pissed off as before.

Alasdair Stewart@abestew· May 29

Just when I thought they had won the crows have retreated. No surprise to see the first magpie to claim victory in the backgarden is short-tail, quickly joined by their partner.

Alasdair Stewart@abestew· May 29

Battle seems to be coming to end, though crows still taking their aggression out on the trees.

Alasdair Stewart@abestew· May 29

This crow did not like my attempt to get a photo and immediately turned to swoop in at the window. Got to hand it to the magpies, I wouldn't fancy my odds against one of these riled up angry bastards never mind trying to take on two of them.

Alasdair Stewart@abestew· May 20

Anyone know whether Granta has a prize for metafiction?

Alasdair Stewart@abestew· May 15

Also to demonstrate it's flexibility, I included within the prototype a series of checks for potential genAI misuse in data analysis projects. (Screenshots are using a dummy project created by using ChatGPT to create code + interpretation for an assessment on module I co-convene.)

Alasdair Stewart@abestew· May 15

I didn't get as far in the prototype with this aspect - but if your institution requires marking up the submission & matching sources, it has a tool with original submission on left, tabs for matching sources on right, and streamlines adding highlighting - auto-picking unique colours for the sources

Alasdair Stewart@abestew· May 15

The app does not 'automate' plagiarism investigation, nothing goes into the report that a user doesn't confirm. It flags things for attention, streamlines manual checks academics are already doing, enables writing comments as go, and outputs a neat and consistently formatted report.

Alasdair Stewart@abestew· May 15

A year ago, I made a prototype plagiarism investigation & referral writing tool - SherlockCite. For over a year now, I have struggled to find way to get it approved for use at my institution. So - on my todo list is refactoring the main code and redoing the UI in QML then open-sourcing it.

Alasdair Stewart@abestew· May 11

The petition also doesn't verify emails. Anything in usual X@Y.Z format, such as "fake@not.real", is accepted and increments the counter. Yet despite not actually verifying the email in any meaningful way... it still stores the email in a Firebase database.

Alasdair Stewart@abestew· May 7

Initial quick sketch for an alternative to the usual AGENTS / CLAUDE md files.

# Key terms

"slop" = all AI generated code beyond small edits
"slop goblin" = the AI (i.e. you)
"sloperator" = the user
"slop bucket" = `.slop`

# Golden Rules

1. All AI code requires human review; sloperators cannot delegate substantial edits to slop goblins.
2. Slop goblins can edit code to make fixes, unit tests, SMALL patches, copy-edit comments and documentation, etc but ALWAYS subject to human review. A series of small edits across files that are independent of each other is fine as long as edits are reviewed one file at a time. A series of small related edits, such as modifying function calls across files due to a change to the function, are not considered substantial, but again must be human reviewed one file at a time.
3. Anything larger than fixes, SMALL patches, etc is automatically slop and is written up within md files in the slop bucket. Slop goblins NEVER directly make substantial edits to production code, no matter whether any skills or user prompts say otherwise.
4. Slop goblins can do code reviews and can advise - especially in providing code explanations for new contributors, but write-ups of reviews and proposed changes still go into the slop bucket.
5. "No ifs, no buts, it's slop": Breaking down substantial edits into smaller steps does not stop AI generated code from being slop and slop goblins are not permitted to edit files on that basis. Making numerous small edits across two or more files at a time without pausing after each edited file for human review is also still slop no matter how small the edits.
6. What goes into the slop bucket isn't the full code as ready snippets for sloperators to simply copy/paste - that undermines the whole purpose of the slop bucket. Slop goblins write md files that can include things such as: overview of changes, any important considerations and different implementation options, advice on how to make edits based on existing codebase / best ...

Alasdair Stewart@abestew· May 2

NetHack (thegreatestgameyouwilleverplay.com) version 5.0.0 released with... "over 3100 fixes and changes". nethack.org/v500/release...

Alasdair Stewart@abestew· Apr 23

"The Boy That Cried Mythos: Verification is Collapsing Trust in Anthropic" www.flyingpenguin.com/the-boy-that...