AI Roundup 096: Happy Shipmas
December 6, 2024.
Genie 2
This week DeepMind unveiled Genie 2, an advanced AI model capable of generating interactive 3D worlds from simple text prompts and images,
Why it matters:
The model can create consistent worlds with multiple perspectives, interactive elements, and complex features like physics and NPC behavior, though only for short durations.
If you've been paying attention, though, this isn't out of left field: World Labs released a very similar (though less sophisticated) model right before DeepMind, and Decart's Oasis was a Minecraft-specific version from way back in October.
These "world models" are the next frontier in AI development - AI labs are trying to build systems that don't just understand language but understand causality and physics like humans do.
And while current applications focus on generative video and AI agent simulations, the ultimate vision is far more ambitious - from interactive 3D worlds to robots with real-world awareness.
Elsewhere in frontier models:
On top of Genie 2, Google also released several additional new models:
PaliGemma 2 models that can analyze images and generate captions describing actions, emotions, and scene narratives.
GenCast, an AI weather model claiming to outperform traditional methods for up to 15-day forecasts and deadly storm predictions.
Veo, an AI video model that generates high-quality 1080p videos in various styles, beating OpenAI's Sora to market.
Meta's internal coding tool, Metamate, now incorporates OpenAI's GPT-4 alongside Llama to enhance its usefulness for developers and other staff.
And AWS unveiled Nova models for Bedrock, including text, image, and video generation capabilities (Simon Willison has some early notes on Nova so far).
Elsewhere in the FAANG free-for-all:
Microsoft rolled out Copilot Vision in preview to analyze text and images on webpages and answer queries for some Copilot Pro users in the US.
Meta plans to invest $10B to build an AI data center in Louisiana, expected to be its largest in the world, with work starting in December.
AWS launched Automated Reasoning checks to combat AI hallucinations and Model Distillation to transfer capabilities from large models to smaller ones.
And AWS added new generative AI features to Amazon Connect, including letting Lex-powered assistants use Amazon Q to help companies run their contact centers.
Happy Shipmas
Yesterday, OpenAI kicked off its "12 days of Shipmas," with new releases and announcements being livestreamed daily.
What's new:
Thursday's stream included the full version of o1, and ChatGPT Pro, a $200/month subscription with unlimited usage and the most powerful models.
ChatGPT Pro is already causing sticker shock, but I'd argue that's because o1 continues to be a misunderstood model. Besides, I'm guessing that the subscription tier will have more perks by the end of Shipmas.
For example: rumor has it that Sora, the highly anticipated video model, will be one of the upcoming releases. Given how computationally intensive it is to run, it wouldn't surprise me if it started off as a ChatGPT Pro model.
Elsewhere in OpenAI:
OpenAI hired its first CMO, Kate Rouch, who previously served as marketing chief at Coinbase and spent over 11 years at Meta.
OpenAI CFO Sarah Friar says the company is considering an ads model while hiring talent from Meta and Google's ad divisions.
OpenAI plans to build data centers in parts of the US Midwest and Southwest, with sources indicating the company is spending over $5 billion annually.
And OpenAI partnered with defense company Anduril to deploy advanced AI systems for national security missions, focusing on improving counter-unmanned aircraft systems.
Make way for Musk
As Elon Musk settles into President-Elect Trump's orbit, some business rivals are worried about how he might settle personal scores and reshape industries in his favor.
Between the lines:
Given his newfound status, Musk's history of targeting business rivals has created anxiety among competitors like OpenAI, Meta, and Amazon about potential regulatory favoritism.
Altman, in particular, has said it would be "profoundly un-American" for Musk to wield political power against his competitors. And yet - there appears to be a rush to hire lobbying firms with ties to Trump.
And it's not just Musk who stands to remake the AI landscape: David Sacks, the investor and All-In podcast host, has been named Trump's "AI and crypto czar."
Elsewhere in AI anxiety:
A CISAC study warns that generative AI could cost music artists ~20% of their revenues by 2028, with ABBA's Björn Ulvaeus cautioning about AI's potential for damage.
Job interviews have become more grueling due to AI's rise, with an increase in technical tests and assignments as recruiters struggle to find suitable candidates.
A new Death Clock app claims to predict a user's likely date of death using AI trained on over 1,200 life expectancy studies.
And ByteDance executives fear that launching their own AI models or apps in the US could give lawmakers another reason to attack TikTok.
Things happen
Q&A with Anthropic CEO Dario Amodei. It turns out AI election misinformation was less catastrophic than feared in 2024. A deep dive into Amazon's Trainium2. Key members of the NotebookLM team leave to launch "user-first AI product" startup. European startups bet on AI-powered weapons as defense budgets increase. Apple and Baidu clash over AI features for Chinese iPhones. The generative AI bridge to the future. An artist turns to Replika's AI bot Lila for companionship. Indigenous engineers use AI to preserve Native American languages. Commonwealth Bank of Australia says AI tools lowered fraud by 30%. US adds new HBM chip transfer restrictions for China. ChatGPT weirdly refuses to state some names. Hugging Face CEO "a little worried" about Chinese open-source AI models. The abject weirdness of AI ads. Spotify Wrapped adds AI podcast feature. AI-enabled dashcams detect driver fatigue and deliver real-time alerts. Tim Cook on Apple's biggest AI contribution being in health. AWS announces Project Rainier, a new Trainium2 supercomputer. AI helps find lost oil and gas wells in old maps. Canadian news organizations sue OpenAI for copyright infringement. Why LLMs make things up and how to fix it. AI poetry indistinguishable from human poetry (and rated more favorably). I've just cucked myself using AI. TSMC in talks with Nvidia to make Blackwell AI chips in Arizona. White House OSTP Director on AI risks and CHIPS Act. Vanguard economist warns of AI stock "correction". UK faces "widening gap" in cyber threat defense. Why investors don't mind that AI is a money pit.
