| Here’s a look at today's AI briefing: - Apple to power AI tasks using own server chips.
- Mistral AI raising funds at $6B valuation.
- OpenAI explores allowing ChatGPT NSFW content.
- TikTok to auto-label more AI content.
- Survey: Majority opposes freely training AI models on public data.
Beth
p/beth-duckett | |
| 1 | | Apple plans to power some of its AI features using its own chips in data centers, according to Bloomberg. Apple will initially use its own M2 Ultra chip for processing AI tasks in the cloud. More: - The M2 Ultra, used in Mac Pro and Mac Studio computers, will be Apple's first AI server chip.
- The high-end chips will process advanced AI tasks — such as those coming to iPhones, iPads, and Macs — in cloud-computing servers.
- These advanced tasks could include image generation, summarizing articles, and crafting long emails.
- Less advanced tasks, like summarizing texts or web pages, will be handled by chips directly in the devices.
Zoom out: - Apple fast-tracked its cloud AI plan, which it started in 2021, due to the AI surge driven by technologies like OpenAI’s ChatGPT, according to Bloomberg.
- The project, code-named Project ACDC, or Apple Chips in Data Centers, is part of Apple’s AI-focused iOS 18 launch this fall.
- Apple has long favored on-device processing to protect privacy, but insiders suggest that existing processor components can also ensure user privacy in the cloud.
|     | |
| 2 | | French AI startup Mistral AI is raising funds at a post-money valuation of $6B, up from $2B in December. Mistral is close to securing around $600M from existing investors General Catalyst and Lightspeed Venture Partners, as well as other potential investors. More: - Paris-based Mistral, founded a year ago, last closed a $415B round at a $2B valuation in December.
- The new round, planned since March or earlier, has received major inbound interest.
- Another potential investor is DST, which has backed Facebook, Twitter, Snapchat, and other tech giants.
Zoom out: - Mistral develops open-source and proprietary large language models including Mistral Large, a competitor to GPT-4.
- The startup, described as Europe's OpenAI rival, released its first paid products in February.
- That same month, Microsoft signed a $16.3M partnership with Mistral to commercialize its models on its Azure AI platform and scale up its AI development.
| | | |
| 3 | | OpenAI is exploring if it can "responsibly" allow ChatGPT to generate NSFW content. The company outlined the possibility in its new draft framework, which will dictate how its AI technology behaves in the future. More: - The framework, called Model Spec, specifies how OpenAI wants its models to function in its API and ChatGPT.
- OpenAI models currently prevent users from generating explicit content.
- However, OpenAI says in the Model Spec it is “exploring whether we can responsibly provide the ability to generate NSFW content in age-appropriate contexts through the API and ChatGPT.”
- The company says it anticipates getting a better understanding of "user and societal expectations of model behavior" with NSFW content.
Zoom out: - The company invites the public and others to give feedback on the framework through May 22.
- Model Spec won't impact existing OpenAI models such as GPT-4, which still follows its current usage policies.
| | | |
| 4 | | TikTok says it will label more AI-generated content in its app. While TikTok already tags in-app AI-generated content, the new policy will extend to content created with outside AI tools, such as OpenAI. More: - TikTok claims to be the first social media app to use Content Credentials, a digital watermark to identify images, videos, and audio created with generative AI.
- Content Credentials attach metadata to track the origin and editing history of AI-generated content.
- The “nutrition label for digital content" can be viewed by users.
- OpenAI's Dall-E, for example, adds the watermark to AI images, which can be automatically detected and labeled when uploaded to TikTok.
Zoom out: - Content Credentials are led by the Coalition for Content Provenance and Authenticity, a group co-founded by Adobe.
- The technology is integrated into Adobe tools like Photoshop and Firefly. Google, Microsoft, OpenAI, and Meta said they will support the standard.
| | | |
| 5 | | A majority of respondents in an AIPI Survey believes AI companies should not be allowed to freely train their AI models on public data. The pro-regulation AI Policy Institute surveyed 1,039 online respondents from April 12-13. What the numbers show: 60% of respondents don't believe AI firms should be allowed to freely use public data to train their AI models. An ever higher percentage -- 75% -- believes companies should compensate creators for any data they do use. Nearly 80% of respondents support government regulations on the matter. What it means: Public sentiment leans negative regarding the use of publicly available data for AI training. The topic has gained attention due to recent legal action, such as lawsuits against OpenAI and Microsoft by major publishers and authors for using content without permission. Last week, eight major U.S. newspapers, including the Chicago Tribune and New York Daily News, accused both companies of copyright infringement for training their models on their content. The New York Times sued OpenAI and Microsoft last year over the same argument. In response, OpenAI says training AI models on publicly available internet data "is fair use" and backed by "long-standing and widely accepted precedents.” 
| | | |
| |
| 7 | | QUICK HITS: - Chinese tech giant Alibaba launched Qwen2.5, the latest version of its Tongyi Qianwen model, with upgrades in reasoning, code comprehension, and textual understanding. Qwen2.5 outperforms GPT-4 in language and creation but lags in knowledge, reasoning, and math, per an OpenCompass analysis.
- Microsoft will soon update its paid Copilot service to auto-complete AI prompts in Office apps.
- VC firm Good Capital plans to invest $25M in up to 20 early-stage AI startups.
- A Bloomberg report speculates on possible successors to Apple CEO Tim Cook. While Cook, 63, has made no indications of leaving, sources suggest COO Jeff Williams as a near-term successor, with John Ternus, senior VP of hardware engineering, as the most likely long-term replacement.
| | | |
 | AI and technology writer | | Beth is a contributing editor and writer of Inside's AI and Tech newsletters. She has written for publications including USA Today, the Arizona Business Gazette, and The Arizona Republic, where she received recognition with a Pulitzer Prize nomination and a First Amendment Award for collaborative reporting on state pension cost increases. You can reach her at Beth.Duckett@yahoo.com. | | This newsletter was edited by Beth Duckett | |
| |
| |
RE:
RE ALL:
FWD:
