Don't s**t where you eat.
The one adage they forgot to "teach" AI.
https://www.wired.com/story/fast-forward-chatbot-hallucinations-are-poisoning-web-search/
Posts in English
@danhon
Sorry. 😅
Although I’m so on board with that statement…
@danhon
You’re likely looking for something like Renderform.
https://renderform.io/
You can likely also do this with a PHP script if it’s always the same image :)
@Beantin
I think that they are by some people intended to be two different aspects of controlling and allowing what goes into the machine instruction. Like this:
1. Prompt design is about writing prompts to get the desired output.
2. Prompt engineering is about the operational management of controlling what prompts are possible, i.e. restricting certain prompts to specific roles within a company in order to maintain data integrity and security, or controlling output length to manage computational costs.
Distinguishing between these two can be useful and set the right expectations, but I do see that prompt engineering is being used as a catch-all phrase. In that sense I would call prompt engineering the umbrella term for both the design/wording of prompts and the construction of rules for prompting.
But given that language is shaped by the people who use it we can only wait and see where these definitions end up. OpenAI themselves seem to use them interchangeably.
Sound reflections from an illustrator on the topic of generative AI.
"AI is existing as it's supposed to exist," says McKernan. "I think it has had a lot of potential to make our lives easier, to make workflows more effective. My issue is that the implementation of it, especially with AI art, hasn't been ethical, in my opinion because of the way it is built off a massive data set with 5.1 billion images, and taxpayers' data, all of which was culled from the internet without consent."
https://www.creativebloq.com/features/ai-art-the-impact-of-generative-AI
@Uva_Be I have experimented with most tools to get an idea of their workings. Both to test usefulness and reliability. For generating templates and structuring content I have seen benefits. And I’m very impressed by the programming capabilities of some language models.
Before these popular current tools I was of course also using AI, for example for speech-to-text and vice versa, which is useful and benevolent in the right settings.
In the end there is always this question that remains… are these benefits (for a subset of wealthy people) worth the cost? To judge that we need to understand the costs, and they are unfortunately often kept hidden or obscure.
Becoming reliant on many of these tools also means I become dependent on someone else to select and filter what content is deemed valuable for modelling, which is another aspect that worries me.
@starfrost Rather than generally biased (which I agree can be a bit of a misnomer as most content can be deemed biased in some way) I would be looking at specific aspects such as racist, medically harmful, misogynist, ableist, abusive etceteras.
Many general tools struggle because they are general, and without a clearly defined purpose and use-case. That makes it hard to gauge success but also hard to gauge appropriateness for anything.
When services/products have more clearly defined goals and purposes there are models for impact assessment that can be made use of as part of a development process.
I outline a couple of them at the end of this post: https://axbom.com/digital-compassion-human-act/
Det är nu 10 år sedan jag började publicera mina strukturerade tolkningar av WCAG (Web Content Accessibility Guidelines) på svenska.
På webbplatsen digitill.se får du en personlig genomgång av dessa riktlinjer för tillgänglighet som idag är starkt kopplade till Lag (2018:1937) om tillgänglighet till digital offentlig service (DOS-lagen). När det gäller EU-direktivet om tillgänglighetskrav för produkter och tjänster (European Accessibility Act) hänvisar man inte lika tydligt till WCAG men grundar sig i samma fyra principer för digital tillgänglighet: att produkterna ska vara möjliga att uppfatta ), möjliga att kontrollera (hanterbara), möjliga att förstå (begripliga) samt robusta.
Idag ägnade jag en stund åt att lyfta in de nya riktlinjerna i och med att WCAG 2.2 publicerades för ett par veckor sedan. Webbplatsen är alltså rykande aktuell igen. 😊 Men precis som jag skriver på webbplatsen så är riktlinjerna bara en liten del av tillgänglighet. Att lyssna på, involvera och anställa fler personer med funktionsnedsättningar i digital utveckling är centralt för att fler perspektiv ska få utrymme och bidra till inkluderande lösningar.
@starfrost For me it's a design decision to make use of biased data and unsupervised training in your manufacturing process. Of course these companies are aware that vast amounts of their data is prejudiced.
The Google photos incident where a black person was tagged as a gorilla happened in 2015 and 8 years later they haven't fixed this, just blocked the ability to search for gorilla. These companies have had time to become aware of the bias problem.
There are definitely many useful applications for AI and it's possible to work around these issues and mitigate impact to some extent before harm happens. Many companies however choose not to. To work around these moral dilemmas I believe you need to be open and honest about their presence and how you are addressing them.
If you aren't actively looking for them as a user, prejudices will also make their way into everyday use of generative models.
@karlemilnikka Gissar att det är Committee on Civil Liberties, Justice and Home Affairs om man ska följa sändningen?
https://multimedia.europarl.europa.eu/en/webstreaming/committee-on-civil-liberties-justice-and-home-affairs_20231025-1430-COMMITTEE-LIBE
@pawsplay I'm not sure I understand your question. The poster suggests what a hammer would be if it was like AI, not what a hammer is.
So if a hammer was like AI, it would have been trained on someone else's data. It's a reflection of the ethical dilemmas of AI.
But you may be getting caught on the word "constructions" which in my mind refers to structures and buildings, not the composition of hammers.
Hence, a language model does not copy the composition of other language models, it is "trained" with copies of other people's content.
I hope this helps.
@fasterandworse I'm thinking now how it will become more relevant in the future to have a policy in place with regards to what tools a company can force its employees to use.
I'm lucky to be self-employed, because there are many tools I would never want to touch but an employment may force me into.
@fasterandworse Not at all! I'm 100% onboard with the lack of purpose being one of the most maddening aspects of this whole problem space.
You added a dimension where my point became even more interesting. They're not just releasing crap and fixing it, they're releasing crap without a purpose and fixing it. 😅
@fasterandworse
I agree, but what I am seeing is that a "fixed" ChatGPT is responding to things like takedown requests from JK Rowling (by making it harder to get responses about the books), blocking explicit racism from responses and generally just addressing random garbage that pops up and makes it into the news cycle.
@tante
@tante
In my head at the time of writing was ChatGPT. :)
I've written about Zoom before, and for me it fits the bill to some extent.
https://axbom.com/avoid-zoom/
For me it's a component of the whole "ship first, fix later" or "release now, fix later" which the gaming industry appears to suffer from as well. I suppose this model naturally became normalized soon after it was possible to ship updates over the net. Many people appear to appreciate early access over fully functional.
A strategy for ethically dubious products is to release something with a set of really poor features, and then improve it significantly within a short period of time.
1. The company will be applauded, and get good press, for quickly fixing something there was already a fix for at the time of release. (Nobody would have applauded the company for releasing something that worked well enough from the start.)
2. People will now more readily accept the updated version because it is "fixed". Chances are that the updated version, had it been released from the start, would have been criticised for being inferior. Now, it can be presented as the much improved version.
3. People will tend to forget about all the other ways the product is inferior, because the company has shown that they are making efforts to improve.
Releasing crap with a quick fix makes people lower their standards.
#DigitalEthics
@efi I've had a handful of people react the way you do and reach out (since June), but my experience is that it's rare. I've had overwhelming positive feedback from people reasoning like I do. The title "If a hammer was like AI" usually gets people on the right track. If in doubt, there's always the web address :)
Mockery is not easy to pull off in a way that everyone gets the intent straight away, so I do understand it will sometimes fail in its messaging. I do appreciate you pointing out your interpretation and telling me.
Here's what happens when machine learning needs vast amounts of data to build statistical models for responses. Historical, debunked data makes it into the models and is preferred by the model output. There is much more outdated, harmful information published than there is updated, correct information. Hence statistically more viable.
"In some cases, they appeared to reinforce long-held false beliefs about biological differences between Black and white people that experts have spent years trying to eradicate from medical institutions."
In this regard the tools don't take us to the future, but to the past.
No, you should never use language models for health advice. But there are many people arguing for exactly this to happen. I also believe these types of harmful biases make it into more machine learning applications than language models specifically.
In libraries across the world using the Dewey Decimal System (138 countries), LGBTI (lesbian, gay, bisexual, transgender and intersex) topics have throughout the 20th century variously been assigned to categories such as Abnormal Psychology, Perversion, Derangement, as a Social Problem and even as Medical Disorders.
Of course many of these historical biases are part of the source material used to make today's "intelligent" machines - bringing with them the risk of eradicating decades of progress.
It's important to understand how large language models work if you are going to use them. The way they have been released into the world means there are many people (including powerful decision-makers) with faulty expectations and a poor understanding of what they are using.
https://www.nature.com/articles/s41746-023-00939-z
#DigitalEthics #AIEthics
Here's what happens when machine learning needs vast amounts of data to build statistical models for responses. Historical, debunked data makes it into the models and is preferred by the model output. There is much more outdated, harmful information published than there is updated, correct information. Hence statistically more viable.
"In some cases, they appeared to reinforce long-held false beliefs about biological differences between Black and white people that experts have spent years trying to eradicate from medical institutions."
In this regard the tools don't take us to the future, but to the past.
No, you should never use language models for health advice. But there are many people arguing for exactly this to happen. I also believe these types of harmful biases make it into more machine learning applications than language models specifically.
In libraries across the world using the Dewey Decimal System (138 countries), LGBTI (lesbian, gay, bisexual, transgender and intersex) topics have throughout the 20th century variously been assigned to categories such as Abnormal Psychology, Perversion, Derangement, as a Social Problem and even as Medical Disorders.
Of course many of these historical biases are part of the source material used to make today's "intelligent" machines - bringing with them the risk of eradicating decades of progress.
It's important to understand how large language models work if you are going to use them. The way they have been released into the world means there are many people (including powerful decision-makers) with faulty expectations and a poor understanding of what they are using.