AI Roundup October 2023: Biden's AI Executive Order, DALL-E 3 in ChatGPT, and ChatGPT Goes Multimodal

TL;DR

October 2023 was the month policy caught up with capability. Biden signed the first broad US executive order on AI, complete with compute-threshold reporting rules that put a number on "frontier model." Meanwhile OpenAI shipped the trifecta: ChatGPT got eyes (GPT-4V), a voice, and DALL-E 3 baked right into the chat. Washington tightened the screws on AI chip exports to China, and Adept quietly open-sourced Fuyu-8B, a multimodal model built for actually driving software.

Biden Signs the First Sweeping US AI Executive Order

On October 30, the White House issued Executive Order 14110 on Safe, Secure, and Trustworthy AI. It is the most comprehensive piece of US AI governance to date, and it is long. The headline for builders: models trained above a compute threshold (the order pegged it around 10^26 floating-point operations) trigger mandatory reporting to the government, including red-team safety test results. There are also provisions on watermarking, biosecurity screening for synthesized DNA, immigration for AI talent, and a pile of agency directives.

Two things matter here. First, the compute threshold is the first time a government put an actual number on "frontier model," and no model in training at the time was anywhere near it, so this was aimed squarely at the next generation of labs. Second, an executive order is not a law. It directs federal agencies and can be reversed by a future administration, which is exactly what makes it a weaker instrument than the EU's incoming AI Act. If you are running 7B models in your homelab, none of this touches you. If you are planning a billion-dollar training run, you now have a paperwork dependency.

ChatGPT Gets Eyes, Ears, and a Voice

OpenAI rolled out multimodal ChatGPT this month, and it genuinely changed what the product is. GPT-4V (vision) lets you drop in an image and ask about it: debug a screenshot, read a chart, identify a plant, reverse-engineer a UI. On mobile, voice conversations went live using Whisper for speech-to-text and a new text-to-speech model with five synthetic voices.

For builders, GPT-4V is the real story. A capable vision-language model behind an API opens up document parsing, UI automation, and accessibility tooling that previously needed bespoke CV pipelines. The voice stack is impressive demo material, but vision is the feature you will actually wire into something. The catch, as always, is that this is closed and metered. The open-weight multimodal ecosystem (LLaVA and friends) was still well behind, which is part of why the next story matters.

DALL-E 3 Lands Inside ChatGPT

DALL-E 3 became available to ChatGPT Plus and Enterprise users on October 19. The big shift is not just image quality, though prompt adherence and text rendering took a real leap over DALL-E 2. It is the integration: ChatGPT writes and refines the image prompts for you. You describe a vague idea in plain language, the model expands it into a detailed prompt, generates options, and you iterate by just talking.

That conversational loop is the actual product innovation. It moves image generation from "learn prompt-craft" to "describe and refine," which is a different audience entirely. For folks running Stable Diffusion locally, DALL-E 3 is not a replacement (no control, no LoRAs, no inpainting workflow, no local generation), but the prompt-rewriting trick is worth stealing for your own pipelines.

The US Tightens AI Chip Exports to China

On October 17, the Commerce Department's Bureau of Industry and Security updated its export controls to close the loopholes from the 2022 rules. The previous controls keyed partly on interconnect speed, so Nvidia had designed the A800 and H800, slowed-down China-specific parts that stayed just under the line. The new rules dropped the interconnect parameter and re-drew the boundary around total performance and performance density, sweeping the A800 and H800 back under the ban.

This is the macro story sitting underneath everything else in AI: compute is geopolitics now. The controls reshaped Nvidia's China business and accelerated domestic Chinese chip efforts. If you care about GPU pricing and availability, supply chain decisions made in Washington are increasingly upstream of what lands in your rack.

Adept Open-Sources Fuyu-8B

On October 17, Adept open-sourced Fuyu-8B, an 8B-parameter multimodal model with a deliberately simple decoder-only architecture. Instead of bolting a separate vision encoder onto a language model, Fuyu feeds image patches straight into the transformer as tokens. That makes it handle arbitrary image resolutions and keeps the architecture clean enough to actually reason about.

The point of Fuyu is agents. It was built to answer questions about charts, diagrams, and user interfaces, and to do fine-grained localization on screenshots, the exact skills you need to drive software. At 8B it is small enough to run on consumer hardware, which makes it a useful open counterweight to GPT-4V in the same month. The license is non-commercial (CC-BY-NC), so it is a research and tinkering release, not a production-ready one.

Key Takeaways

• Regulation got specific. A compute threshold is now a real line in US policy, which means "how big is your training run" is becoming a compliance question, not just an engineering one.
• Multimodal went mainstream. Vision plus voice plus native image generation in one app reset expectations for what an AI assistant is supposed to do.
• Vision over voice for builders. GPT-4V is the API feature worth integrating now; the voice stack is mostly demo magic.
• Compute is geopolitics. Export controls shape GPU supply and pricing, so policy in Washington is upstream of your homelab.
• Open multimodal is catching up. Fuyu-8B shows the open ecosystem chasing GPT-4V, even if licenses and quality still trail the closed players.