Draft EU AI Act regulations could have a chilling effect on open-source software
- Reference: 1662894010
- News link: https://www.theregister.co.uk/2022/09/11/in_brief_ai/
- Source link:
The [1]proposed EU AI Act, yet to be signed into law, states that open source developers have to ensure their AI software is accurate, secure, and be transparent about risk and data use in clear technical documentation.
Brookings [2]argues that if a private company were to deploy the public model or use it in a product, and it somehow gets in trouble due to some unforeseen or uncontrollable effects from the model, the company would then probably try to blame the open source developers and sue them.
[3]
It might force the open source community to think twice about releasing their code, and would, unfortunately, mean the development of AI will be driven by private companies. Proprietary code is difficult to analyse and build upon, meaning innovation will be hampered.
[4]
[5]
Oren Etzioni, the outgoing CEO of the Allen Institute of AI, reckons open source developers should not be subject to the same stringent rules as software engineers at private companies.
"Open source developers should not be subject to the same burden as those developing commercial software. It should always be the case that free software can be provided 'as is' — consider the case of a single student developing an AI capability; they cannot afford to comply with EU regulations and may be forced not to distribute their software, thereby having a chilling effect on academic progress and on reproducibility of scientific results," he [6]told TechCrunch.
New MLPerf results for inference are out
The results for the annual MLPerf inference test, which benchmarks the performance of AI chips from different vendors across numerous tasks in various configurations has been [7]published this week.
Nearly a whopping 5,300 performance results and 2,400 power measures were reported this year for inference in the datacenter and in edge devices. The tests look at how fast a hardware system is able to run a particular machine learning model. The rate of data crunching is reported in spreadsheets.
[8]
It's no surprise that Nvidia tops the rankings again this year. "In their debut on the MLPerf industry-standard AI benchmarks, Nvidia H100 Tensor Core GPUs set world records in inference on all workloads, delivering up to 4.5x more performance than previous-generation GPUs," Nvidia gushed in a blog post. "The results demonstrate that Hopper is the premium choice for users who demand utmost performance on advanced AI models."
Although an increasing number of vendors are taking part in the MLPerf challenge, it can be difficult to get a good idea of the competition. There are no reported results for Google's TPU chips in the datacenter track this year, for example. Google did, however, seem to [9]ace MLPerf's training competition earlier this year.
AI artists discovers horrific face lurking behind image
A viral Twitter thread posted by a digital artist reveals just how strange text-to-image models can be beneath the surface.
Many netizens have found joy and despair in experimenting with these systems to generate images by typing in text prompts. There are sorts of hacks to adjust the models outputs; one of them, known as a "negative prompt, allows users to find the opposite image to the one described in the prompt.
When an artist, who goes by the name of Supercomposite on Twitter, found the negative prompt for what described an innocent-looking picture of a fake logo they found something truly horrifying: The face of what looks like a haunted woman. Supercomposite has named this AI-generated woman "Loab" and when they crossed images of her with other ones, they always ended up looking like a scene from a horror film.
🧵: I discovered this woman, who I call Loab, in April. The AI reproduced her more easily than most celebrities. Her presence is persistent, and she haunts every image she touches. CW: Take a seat. This is a true horror story, and veers sharply macabre. [10]pic.twitter.com/gmUlf6mZtk — Supercomposite (@supercomposite) [11]September 6, 2022
Supercomposite told El Reg random images of AI-generated people can often show up in negative prompts. The weird behavior is yet [12]another example of some of the weird properties these models can have that people are only beginning to probe.
No sentient chatbots here at Google says CEO Sundar Pichai
Sundar Pichai contradicted claims made by former engineer, Blake Lemoine, that Google had built a sentient chatbot during his talk at the Code conference this week.
Lemoine made headlines in July when he [13]announced he thought Google's LaMDA chatbot was conscious and might have a soul. He was later [14]fired for reportedly violating the company's privacy policies after he hired a lawyer to chat to LaMDA and assess its legal rights, claiming the machine had told him to do so.
[15]
Most people - including Google's CEO - disagree with Lemoine. "I still think there's a long way to go. I feel like I get into philosophical or metaphysical talks often about what is sentience and what is consciousness," Pichai said, [16]according to Fortune. "We are far from it, and we may never get there," he added.
To stress his point further, he admitted that Google's AI voice assistant sometimes fails to understand and respond appropriately to requests. "The good news is that anyone who talks to Google Assistant—while I think it is the best assistant out there for conversational AI — you still see how broken it is in certain cases," he said. ®
Get our [17]Tech Resources
[1] https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206
[2] https://www.brookings.edu/blog/techtank/2022/08/24/the-eus-attempt-to-regulate-open-source-ai-is-counterproductive/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Yx4GFX1didhn56Vudx3BewAAAMU&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Yx4GFX1didhn56Vudx3BewAAAMU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Yx4GFX1didhn56Vudx3BewAAAMU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://techcrunch.com/2022/09/06/the-eus-ai-act-could-have-a-chilling-effect-on-open-source-efforts-experts-warn/
[7] https://mlcommons.org/en/news/mlperf-inference-v21/
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Yx4GFX1didhn56Vudx3BewAAAMU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://cloud.google.com/blog/products/ai-machine-learning/cloud-tpu-v4-mlperf-2-0-results
[10] https://t.co/gmUlf6mZtk
[11] https://twitter.com/supercomposite/status/1567162288087470081?ref_src=twsrc%5Etfw
[12] https://www.theregister.com/2022/06/07/in_brief_ai/
[13] https://www.theregister.com/2022/06/13/google_lamda_sentient_claims/
[14] https://www.theregister.com/2022/07/23/in_brief_ai/
[15] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Yx4GFX1didhn56Vudx3BewAAAMU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[16] https://fortune.com/2022/09/07/alphabet-ceo-sundar-pichai-says-broken-google-voice-assistant-proves-that-a-i-isnt-sentient/
[17] https://whitepapers.theregister.com/
Indeed
And anyone deploying a system where functional safety is a consideration is required to ensure that the safety level of any component that is used satisfies the project's safety objectives (SIL / ASIL, ...).
The due diligence around that would include ensuring that documentation is appropriate, test coverage is complete, and that test artefacts are available to show that the claimed requirements are satisfied and that testing has been completed. This is unlikely to exist for an open source project, so this work is going to have to be done in house.
To me, this act is really just saying that using open source doesn't mean all of the above can be ignored (which, unfortunately, is quite often the case when open source is used within safety-related systems). The only way there could be any comeback on the open source developers would be if they are making a claim that the software "is safe", but that would be tricky, as who made that claim?
I think the scenario they are talking about is a company who uses an open sourced AI component and subsequently falls foul of the regulations; they might decide to try to pass the legal buck on to the open source developer(s).
Pretty sure that there is a way to solve this by adding appropriate clauses to the license. Firstly you can put a condition in the license that states that use of the open source component in any commercial application is at the company's own risk, and that the authors of the component do not accept any responsibility for legal problems that this might cause the company. You could re-enforce that with a clause that says that the AI component cannot be used for any commercial purposes within EU legal jurisdiction, and that the legal costs of the open source developers must be paid by the company regardless of the outcome of any legal action.
Note: I am not a lawyer, this is simply my take on the subject.
TLDR
To me, from the summary of the legislation, it looks quite reasonable and basically says that a company can't use open source as an excuse for not meeting other requirements. I fail to see a direct impact on the open source community.
Job opportunities!
> Supercomposite told El Reg random images of AI-generated people can often show up in negative prompts
And that, mark my words, was the birth of the profession of "AI shrink"!
"Let me explore the subconscious of your AI model: Does your AI dream of electric sheep, or are there lost, repressed childhood memories buried deep within, which could be triggered unexpectedly and cause it to misbehave, resulting in lawsuits, lost profit and all that?"
Future Building via the Advanced IntelAIgent Virtual Machine Root Route
There's a hell of a lot of ifs and ors, somehows and unforeseen or uncontrollable probabilities invented in that Brookings tale, Katyanna, and none here would expect them capable of spouting anything else for that is surely undeniably their usual bog standard modus operandi/vivendi.
It would be more than a tad odd though for them to think AIdDevelopment be listening to take heed of their concerns rather than having extensive and expansive plans already rolling along to eliminate them and in so doing have IT and media paint a whole new book of novel panorama pictures for human realisation and virtual machine colonisation.
As far as I can see from a very quick look, it says nothing of the sort about open source. Just that AI applied to anything has to still meet the rules on safety and discrimination that any other technical or business method would have to.
If you use anything in your company you ought to do due diligence on it and/or have supervisory capacity to control it. AI is just another tool.