The AI IPCC
New reports indicate the UK plans to announce an international advisory group on AI, modeled on the UN's IPCC (Intergovernmental Panel on Climate Change).
Between the lines:
Several leading tech figures are backing this move, including Eric Schmidt (the former CEO of Google and Mustafa Suleyman (cofounder of DeepMind).
Others, less so - Yann LeCun, one of the "godfathers of AI," has said that regulating R&D for AI is "incredibly counterproductive".
And while the UK is working on its summit, the EU has its own plans - it's reportedly considering a three-tiered approach to regulating generative AI under this year's AI Act.
Elsewhere in AI proclamations:
Marc Andreessen of a16z published a 5000-word manifesto on techno-optimism and AI. It's... a lot, and the internet has had a lot of thoughts.
Yoshua Bengio, another "AI godfather", has called for a "humanity defense organization". Bengio and LeCun have argued online over AI safety.
And best-selling author William D. Cohan wants writers and publishers to band together in the fight against unauthorized (and unpaid) collection of content for model training.
Official fakes
YouTube is working on an AI tool to let creators make songs with the voices of famous musicians. But it needs record labels' permission to do so - or risks yet another copyright lawsuit.
Why it matters:
The cat's out of the bag with AI-generated vocals, and YouTube wants to (legally) enable creators in this regard.
The music industry sees AI as both a big opportunity and big threat. Several publishers are suing Anthropic over copyright violation, claiming the company used song lyrics to train its AI models.
But YouTube is finding ways to incorporate more and more AI, with its recent tool to help brands target ads on videos referencing specific cultural moments (e.g. Halloweeen or the Oscars).
Elsewhere in AI anxiety:
A different copyright lawsuit launched this week, this time from authors including former Arkansas Governor Mike Huckabee.
Some AI image detecting tools are labeling real photos of the Israel-Hamas war as fake, creating a "second level of disinformation."
Gary Gensler, the head of the SEC, warns that over-reliance on AI will make a financial crisis "nearly unavoidable" without regulation.
OpenAI has a lot going on
Despite no major launches, OpenAI made quite a few appearances in the news this week.
What to watch:
The company's public-facing core values were changed, putting a greater emphasis on building AGI.
They also reportedly scrapped a new model, codenamed “Arrakis,” that was meant to produce GPT-4 level output more efficiently.
OpenAI is debating whether or not to release an AI image detector, which is currently "99% reliable" at detecting images specifically made by DALL-E 3.
Speaking of DALL-E 3, the image-generating functionality is now available to all ChatGPT Plus subscribers.
Sam Altman says OpenAI isn't exactly building a phone (at least right now).
PWC is partnering with OpenAI to offer clients advice generated by AI.
Nvidia unveiled Eureka, a new agent powered by GPT-4 that can teach robots complex skills.
And Sam Altman (plus other tech leaders) reiterated at the WSJ's Tech Live conference how huge an impact AI will have on people's jobs.
Things happen
Small amounts of fine-tuning can undo model safety guardrails. A 21-year-old solves a 2000-year-old historical mystery (with AI). Foxconn and Nvidia plan to build "AI factories." NYC Mayor deepfakes himself to make robocalls in multiple languages. Open questions for AI engineering. A new antibiotic identified by AI. Stanford's Foundation Model Transparency Index. Japanese tea commercial features an AI-generated actress. The AI-generated porn industry. Nvidia's H800 AI chip for China blocked by new export rules. Stack Overflow lays off 28% of staff, as AI-coding tools keep growing. Meet two open-source challengers to GPT-V. Anti-ChatGPT app Superfy uses AI to match people for live chats and answers. Reddit is considering walling up its content if AI companies don't pay up. The case against AI everything, everywhere, all at once.