AI Roundup 064: Big Tech's small models
April 26, 2024.
Big Tech's small models
This week, Apple researchers shared OpenELM, a family of small language models (270M to 3M parameters) that are pre-trained and designed to run on mobile devices.
Why it matters:
Apple's generative AI strategy has been pretty opaque thus far, but this seems to be a clear indication that it wants to run AI models locally on Apple devices.
In addition to OpenELM, Microsoft released Phi-3 Mini, the latest version of its model that's roughly as capable as GPT-3.5. Google's SLM, Gemma, was released in February.
These smaller, open-source models are cheaper to run and optimized for phones and laptops - great for cost-conscious developers who can't or won't pay top dollar for GPT-4 and Claude Opus.
Elsewhere in the FAANG free-for-all:
Mark Zuckerberg discussed Meta's AI investments - and how it will likely take years to make money from generative AI.
AWS announced the launch of Custom Model Import, which lets companies run their in-house generative AI models as fully managed APIs.
And Microsoft hired a former Meta exec to help build its AI supercomputing team.
Data center disruptions
With so much AI demand, companies are racing to build new data centers to power it all. But the construction frenzy is now leading to major supply chain bottlenecks.
Between the lines:
Much attention has been paid to the GPU shortage, but once we have enough GPUs, we'll still need places to put them and electricity to power them.
As companies and governments budget billions for AI infrastructure, the demand for suitable land and energy will follow.
Case in point: Amazon recently bought a $650M nuclear-powered data center in Pennsylvania, while Meta is spending $800M on a data center in El Paso, Texas.
Elsewhere in AI hardware:
Meta rolls out multimodal AI (in early access) for its Ray-Ban Smart Glasses, to help identify objects, translate signs, and write Instagram captions.
After Humane's disastrous release, the Rabbit R1 launched to positive, if mixed reviews.
And Nvidia acquires AI infrastructure orchestration and management service Run:ai, reportedly for ~$700M.
CSAM DDOS
In a new report, the Stanford Internet Observatory suggested that AI-generated CSAM (child sexual abuse material) could overwhelm the CyberTipline, a linchpin in our current reporting infrastructure.
The big picture:
The CyberTipline is a federally authorized clearinghouse for reports of online CSAM. It fields tens of millions of reports annually from social platforms and forwards them to law enforcement agencies.
But as internet usage has grown, the CyberTipline has overwhelmed agencies, as reports fall through the cracks. And with generative AI, the problem stands to get dramatically worse.
Shortly after the report, OpenAI, Meta, Anthropic, and others committed to new CSAM-prevention measures, including using CSAM-free training datasets and "stress-testing" models to prevent CSAM generation.
Elsewhere in AI anxiety:
ReplyGuy is an AI-powered service that automates shilling products via Reddit posts and comments.
GPT-4 appears capable of exploiting real-world vulnerabilities based on security notifications and descriptions.
Drake was threatened with a lawsuit after using "AI Tupac" in a recent diss track.
And a high school athletic director was arrested and accused of framing his school's principal with AI.
Last week’s roundup
Things happen
Meta's AI Assistant is fun to use, but it can't be trusted. OpenAI announces new enterprise-grade features. How Saudi Arabia is spending big to become an AI superpower. Facebook's bizarre AI images are now on LinkedIn, too. AI for data journalism. Llama 3 may have just killed proprietary AI models. AI has given the enterprise device market something to be excited about. A look at Synthesia, whose video AI represents the latest in realistic, expressive avatars. AI programming tools should be added to the Joel Test. Why vector databases are having a moment. AI really is smoke and mirrors. Snowflake releases a flagship generative AI model. Nurses protest "deeply troubling" use of AI in hospitals. Employers and job candidates are dueling with AI in the hiring process. An AI that turns body cam audio into police reports. How United Airlines uses AI. Adobe adds AI tools powered by its new Firefly model to Photoshop. The question that no LLM can answer and why it is important.