AI Roundup 055: Race conditions
February 23, 2023.
Race conditions
This week Google took a PR hit after users accused Gemini of refusing to depict images of white people. After a combination of backlash and ridicule, Google first apologized and then paused the ability to generate images of people entirely.
Why it matters:
On the one hand, there's an argument that image generation doesn't actually matter that much. On the other, it seems reasonable to want LLM providers to be transparent in their system prompting and filtering.
I don't want to blindly add to the Google pile-on, but this does seem like a major unforced error. Both the heavy-handed system prompt as well as the initial corporate response.
Of course, Gemini will continue to get better and more ubiquitous. It's coming both to Chrome as well as advertising tools shortly.
Elsewhere in model updates:
Despite the hullabaloo, Google also released Gemma - a family of open models based on the same research as Gemini, albeit at smaller sizes.
Stability AI previewed Stable Diffusion 3, which, unlike previous versions, now uses diffusion transformers in its architecture.
And ChatGPT had a wild night of spouting gibberish before being fixed; OpenAI noted that the issue was with converting between text and tokens.
Reddit sells out
Ahead of its IPO plans, Reddit announced a deal with Google to license its content for AI training purposes - the deal is reportedly worth $60 million per year.
Between the lines:
Reddit has been moving in this direction for months - last year, the company began charging for API access, killing many third-party clients.
The deal echoes OpenAI's agreement with Axel Springer late last year, and I suspect we'll see many more in the coming months.
These publishing deals are necessary because web crawlers are now controversial - a new study shows that 48% of top news sites have blocked OpenAI's crawlers.
Elsewhere in the FAANG free-for-all:
Microsoft will reportedly partner with Intel to manufacture an internally-designed chip.
DeepMind announced AI Safety and Alignment, a new organization focused on AI safety.
And Nvidia reports a crazy 265% YoY Q4 revenue growth, adding $277B in market value in a single day.
People are worried about AI deepfakes
Last Friday, major tech companies signed a voluntary accord to prevent deepfakes from impacting worldwide elections. On Wednesday, hundreds of signatories signed an open letter calling for more regulation around the creation of deepfakes.
Why it matters:
While many are concerned about deepfakes in general, the most pressing area of concern is specifically political deepfakes.
Companies are pledging to fight AI-based election misinformation, Anthropic being the latest, but users continue to create problematic content via loopholes in ChatGPT and Gemini.
And a new study found that AI-generated propaganda was about as effective as the real thing.
Elsewhere in AI anxiety:
AI is already making critical health care decisions despite being unregulated.
AI hiring tools may be filtering out the best applications.
And a new report found that 60% of ChatGPT's outputs contained some form of plagiarism.
Things happen
Princeton professor John Mayor is the DOJ's first Chief AI Officer. Magic startup may have made an AI reasoning breakthrough. Insiders say China lags behind the US in generative AI by at least a year. Scale AI signs a Pentagon contract to test military decision-making LLMs. Demos from AI chipmaker Groq go viral with lightning-fast completion speeds. AI can determine your sex via brain scans. Why the NYT might win its OpenAI copyright lawsuit. Adobe launches AI assistant in Acrobat to chat with your documents. How Anthropic raised $7.3B last year, including via unusual deal structures. Microsoft President says the company will invest $2.1B in Spain over the next two years. Instacart's AI recipes look literally impossible. VCs and founders are flocking back to SF, after leaving during the pandemic. A look at Nvidia's AI startup investments. UK, US, and other allies to provide Ukraine with AI-enabled drones. OpenAI completes its tender offer at an $80+B valuation. 'Better than real men': Young Chinese women turn to AI boyfriends. Masayoshi Son seeks up to $100B for new AI chip venture. Khan Academy's founder wants to change your mind about AI. Artificial investment. Retell AI: real-time conversational speech for your LLM. Things I don't know about AI. LlamaIndex's new managed services for document parsing and retrieval. The NYT plans new AI ad tools for later this year. LLM agents can autonomously hack websites.
