On why GPT-3 is the language generator we (unfortunately) deserve: ”At its core, GPT-3 is an artificial bullshit engine—and a surprisingly good one at that ... like a human bullshitter, it also has no intrinsic concern for truth or falsity.”
I’m not sure it’s necessarily recognised among people who publish web content - but language generating tools need open access content to learn from, which means the kind of content that is open access is a massively important part of the machine-learning ecosystem.
Common Crawl commoncrawl.org and Wikipedia are two of the many sources that GPT-3 learns from. Gender bias on Wikipedia is such a well-known phenomenon that there is even, wait for it, a Wikipedia page on gender bias on Wikipedia en.wikipedia.org/wiki/Gender_bi…
There is a massively complex bit of the social contract to be negotiated here - what is a representative body of content, who enforces use of that representative content, who supports the creation of it, how is profit earned and distributed? It goes on and on.
And we’re headed to more automated content moderation, and not just at Facebook - the @DCMS Data Strategy refers to creating "a competitive commercial market in tools able to detect online harms such as cyberbullying, harassment or suicide ideation” gov.uk/government/new…
Anyway, sorry, I’m not sure where I’m headed with all this woe, except that, idk, this is why good global governance is simultaneously so important and so impossible.
Sigh.
(Mild lol)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
So, Kids and Mobile Phones: The Moral Panic seems to be building to an exciting fever pitch with the publication of Haidt's book.
I have some pragmatic, middle-of-the-road opinions about this, which can be roughly summed up as "Just enough Smartphone".
My position is roughly: some things about technology are great, but excessive datafication and corporate capture mean we've ended up in an extractive and exploitative place, in which most of us are making a small number of businesses a great deal of money.
In almost 30 years of working on the Internet (including a stint running an online community for teens and many years in online safety) it's repeatedly struck me that the personal nature of our digital experiences can be hard to communicate.
I think I'd go a bit further than Simon's post though, because it seems to me that using our human instincts for what may or may not be trustworthy is an essential line of defence. If the link looks bad, don't click it; if the alleged news story looks like BS, check the source
I don't think that trustworthiness can necessarily be improved by transparency alone though - I'll defer to Onora O'Neill who says that we need "actual communication" rather than mere transparency and "honesty, competence, and reliability" thebritishacademy.ac.uk/documents/2563…
Quick thread on the state of digital policy in the UK. Interested to know if this reflected in other areas.
In what is, presumably, the last year of a Conservative govt we find ourselves in an odd place that I think is almost peak Theatre of Consultation.
Unless I was asleep under a giant rock and missed it, there was no consultation about the formation of the AI Safety Institute, or about the methods of societal impacts that have been selected, which make no reference to human rights and wch appear technocratic at best.
Instead, we had AI-pa-looza at Bletchley Park. While reams and teams has been written about this, there has been no consultation and the PM appears to be making off-the-cuff policy decisions. ft.com/content/ecef26…
I'm doing a panel this morning on digital inclusion and AI. This is what I'm going to say:
- the paradigm for AI governance the UK govt is working towards deepens social exclusion
- so we need to do two things: challenge the paradigm while also mitigating it
Mitigations for structural power imbalances can have the unfortunate outcome of entrenching existing power imbalances so it's important to do both. Being included in an oppressive system can still be oppressive. I wrote about that here medium.com/careful-indust…
Meanwhile, technologists are always trying to write new social contracts based on what AI can do. But no amount of polling or public deliberation will displace the Declaration of Human Rights and the SDGs in the short term. They may not be perfect but they are what we have.
Quick thread on Responsible Capability Scaling, one of the safety measures outlined in a @SciTechgovuk paper published yesterday - and why it is both welcome and insufficient. assets.publishing.service.gov.uk/media/653aabbd…
Parts of Responsible Capability Scaling have a lot in common with Consequence Scanning, a tool we developed at @doteveryone in 2018/9, in collaboration with many SMEs, which is freely available and widely used by businesses and research teams doteveryone.org.uk/project/conseq…
@doteveryone What Consequence Scanning tries to do is help teams start to apply an external lens on what they are doing, beyond internal OKRs/KPIs, and help teams envisage how their product will grow and change in the world, beyond their business goals
Well I guess this is my daily "read the news and complain about today's ridiculous AI story" tweet. Buckle up, I have a thread theguardian.com/technology/202…
Firstly, let's look at the headline. Sure, Stuart Russell is an expert, but he's not an expert in either education or child welfare, he's an expert in AI. You know the saying, "When you have a hammer, everything looks like a nail" - well, that applies here. The idea that teaching… twitter.com/i/web/status/1…
The idea that it might be desirable for teaching to become redundant assumes, I think, that children need to learn in the same way as neural nets. But, vitally, school also teaches kids about relationships and people and communication.