AI Roundup December 2023: Gemini Arrives, Mixtral Drops a Torrent, and the EU AI Act Lands

TL;DR

December closed out 2023 with a bang. Google finally answered GPT-4 with Gemini, Mistral dropped Mixtral 8x7B as a bare torrent link and handed builders a genuinely competitive open mixture-of-experts model, and the EU reached political agreement on the AI Act, the first comprehensive AI law of its kind. Microsoft also slipped out Phi-2 to prove small models still have teeth, and Midjourney shipped its v6 alpha just before the holidays.

Google Ships Gemini, and the Demo Video Gets Side-Eye

On December 6, Google unveiled Gemini, its long-promised GPT-4 competitor, in three sizes: Ultra, Pro, and Nano. Pro went live inside Bard the same day, Nano was aimed at on-device work on the Pixel 8 Pro, and Ultra was held back for early 2024. Google leaned hard on the claim that Gemini Ultra beat GPT-4 on a stack of benchmarks and was the first model to top human experts on MMLU.

Then the internet did what the internet does. The slick hands-on demo video turned out to be edited: responses were sped up and the model was actually working from still images and text prompts, not the smooth real-time voice-and-video interaction the footage implied. Google admitted as much. The takeaway for builders: Gemini Pro was real and available, but treat launch-day benchmark theater with the skepticism it earned.

Mistral Drops Mixtral 8x7B as a Torrent, No Blog Post Required

On December 8, Mistral did the most Mistral thing possible: it tweeted a magnet link with no announcement, no marketing, just weights. That torrent was Mixtral 8x7B, a sparse mixture-of-experts model under Apache 2.0. It routes each token through 2 of 8 expert blocks, so it carries roughly 47B total parameters but only activates around 13B per token, which means it punches like a big model while running closer to the speed of a small one.

The numbers backed up the swagger. Mixtral matched or beat Llama 2 70B on most benchmarks with far faster inference, and traded blows with GPT-3.5. For the local-AI and homelab crowd, this was the headline release of the month: a permissively licensed, quantizable model you could actually run yourself that didn't feel like a compromise. The MoE-as-default era started here.

The EU AI Act Gets Its Political Deal

After a three-day negotiating marathon, EU lawmakers reached provisional political agreement on the AI Act on December 8 and 9, the world's first comprehensive horizontal AI regulation. The deal introduced a risk-tiered framework, banned certain uses outright, carved out narrow law-enforcement biometric exceptions, and, crucially for anyone training large models, added obligations for general-purpose and high-impact AI systems that could pose systemic risk.

This was a political agreement, not the final enforceable text, and the real timelines stretched well into the following years. But it set the global tone. If you build with or deploy foundation models touching EU users, this is the regime you were going to live under, and December is when the shape of it became concrete.

Microsoft Phi-2 Makes the Case for Small

Microsoft Research released Phi-2, a 2.7B parameter model, arguing that quality data beats raw scale. Despite its size, Phi-2 posted reasoning and language-understanding numbers that embarrassed models many times larger, leaning on the textbooks-are-all-you-need data-curation thesis from Microsoft's earlier Phi work.

For builders, Phi-2 was a useful counterweight to the bigger-is-better narrative. A model that fits comfortably on modest hardware and still reasons well is exactly what you want for edge deployment, cheap fine-tuning experiments, and anything where you cannot ship a 70B model.

Midjourney v6 Alpha Lands Before the Holidays

Around December 21, Midjourney opened the alpha of v6, its biggest image-model jump in a while. The standout upgrades: noticeably better prompt adherence, more coherent fine detail, and, finally, the ability to render legible text inside images, long a weak spot for diffusion models.

It was an alpha, so expect rough edges, but the quality leap was obvious enough that it reset expectations for what a single prompt could produce. If you generate marketing or product visuals, v6 was the moment Midjourney stopped fumbling on typography.

Key Takeaways

• Open weights closed the gap. Mixtral 8x7B proved a permissively licensed MoE model could trade blows with GPT-3.5, and you could run it on your own boxes.
• Benchmark theater is back. Gemini was a real GPT-4 competitor, but the staged demo is a reminder to verify launch claims before you architect around them.
• Regulation got real. The EU AI Act political deal means foundation-model obligations are coming, so factor compliance into anything serving EU users.
• Small models still matter. Phi-2 showed that careful data curation can beat scale, which is great news for edge and cost-sensitive deployments.
• Image models learned to spell. Midjourney v6 alpha cracked in-image text and prompt adherence, raising the floor for AI-generated visuals.