AI Roundup 040: GPTs
November 10, 2023.
If you’re new, welcome to another AI Roundup - our Friday issue covering the week’s top AI stories in a few minutes.
GPTs
DevDay, OpenAI's first developer conference, took place on Monday. The company announced a slew of models, products, and pricing changes - you can read the full recap here.
The highlights:
GPT-4 Turbo, a new model with 128K-token context, an April 2023 knowledge cutoff, and a 2-3X price drop from GPT-4.
GPTs, custom versions of ChatGPT that can be built without any code and will be available in a GPT Store.
The Assistant API, a native way to manage ChatGPT threads and conversation state.
API access to DALL-E 3, GPT-Vision, and text-to-speech models.
And custom models, a service to train completely custom models from scratch, starting at $2-3M.
Elsewhere in OpenAI:
The post-conference rollout of GPTs has been a bit bumpy, and the company had a major outage Thursday, reportedly due to a DDOS attack.
It also announced Data Partnerships, a program to work with organizations directly to create public and private AI training datasets.
And Snap unveiled ChatGPT-powered filters at its AR event this week.
Elsewhere in developer conferences:
GitHub Universe, GitHub's annual conference, was Wednesday and Thursday this week. The conference is mostly focused on software development, but there were a few AI-related announcements.
Copilot Chat will be available on web and mobile in December, and Copilot Enterprise, which offers models fine-tuned on an entire codebase, will be available next year.
And Copilot will be adding to its capabilities, via an ecosystem of third-party plugins (Copilot Partner Program) and an end-to-end AI-powered workflow (Copilot Workspace).
Humane
After quite a lot of hype, Humane finally launched its Ai Pin, a new LLM-powered wearable.
How it works:
The pin is meant to be a smartphone without a screen - which means you interact through voice, gestures, projections, and directional speakers.
You use it primarily via voice and use "Ai Experiences" rather than apps. Examples include translation, web browsing, messaging, and shopping.
It will be available for $699 (plus a $24/month subscription) on November 16th.
Between the lines:
It's very cool to see hardware innovation, especially hardware that integrates generative AI. The engineering alone is deeply impressive.
That said, I don't fully understand the Ai Pin's fanfare yet. I'm not sure we're ready to dispense with smartphones entirely - as addictive as they are, they also provide incredibly rich I/O.
Humane is pioneering new ways of using LLMs, and someone will find their killer piece of hardware. But we don't yet know whether it’s pins, glasses, earbuds, or something else entirely.
People are worried about AI politics
On Tuesday, Meta barred political campaigns from using its generative AI advertising tools, while Microsoft offered tools and services to defend elections from AI-powered misinformation.
Why it matters:
As much as technologists don't like to engage in politics, AI is inherently political. And the threat of tampered elections is perhaps the fastest way to get AI regulated.
The last two weeks saw plenty of government actions on AI, and more are coming - the FTC is now weighing in on AI copyright rules from the US Copyright Office, while the SEC is facing pushback on its recent AI-related proposals.
And beyond elections, we're still seeing AI bias and misinformation permeate news about the Israel-Hamas conflict.
Elsewhere in the FAANG free-for-all:
Amazon is reportedly working on Olympus, a new LLM to challenge GPT-4.
Google's SGE adds 120+ new countries, while YouTube tests new generative features for viewers.
Microsoft and Oracle sign a multiyear deal giving Microsoft access to A100s and H100s via Oracle's cloud.
And of course, Elon Musk's xAI releases Grok, an AI with "an incurable case of boomer cringe."
Things happen
Game developers criticize Xbox for its AI game dev tools. Samsung announces Galaxy AI and Samsung Gauss. Project Olympus, a new LLM inside of Amazon. Biden's elusive AI whisperer finally goes on the record. IBM's $500M enterprise AI fund. China's 8-month old, billion-dollar AI startup. Adept Experiments. AI is about to completely change how you use computers. Man uses AI to apply to 5000 jobs, gets 20 interviews. Q&A with Deepmind's Head of Robotics. How AI is impacting the Lego community. Fakespot Chat, Mozilla's first LLM. Bulletpapers: an AI ArXiv paper summarizer. The AI app graveyard. AI companies argue against paying for copyrighted content. Om Malik talks to Imran Chaudhri.
Last week’s roundup
AI Roundup 039: The governance issue
On Monday, President Biden signed a sweeping Executive Order on AI and launched AI.gov, a new site promoting policies and AI recruitment. Key themes include limiting large model training, adding more AI talent, and directing government agencies to think about AI.
Humane might not crack it at all, or this go... but there will be a brief window (few years at least) where AI wearables are going to unlock some cool experiences / capabilities for users.
I think not long after... we'll be in the "cognification of everything" environment and it won't be required to carry your own edge compute device w/sensors and whatnot..
We'll just talk / BCI-WIFI and get picked up by new devices as we walk in range, and continue the conversation that is saved/generated/served from the Cloud to the smart devices we're locally networked with on our walk.