AI Roundup January 2024: GPT Store Goes Live, OpenAI Drops Its Military Ban, and Altman Chases Chip Trillions

TL;DR

January 2024 was an OpenAI month with a few asterisks. The GPT Store finally shipped, OpenAI quietly stripped the "military and warfare" ban out of its usage policy, and Sam Altman kicked off a fundraising tour for a global network of chip fabs. Underneath the headlines, the legal fight with The New York Times got nastier, and Mistral's Mixtral paper gave the open-weights crowd a genuinely useful mixture-of-experts blueprint to chew on.

The GPT Store Finally Opens for Business

On January 10, OpenAI launched the GPT Store, the marketplace it teased at DevDay back in November and then sat on through the board meltdown. It lives in a tab inside ChatGPT and surfaces custom GPTs built with the no-code GPT Builder, sorted into buckets like writing, research, programming, and education. OpenAI said users had already spun up around three million GPTs by launch.

The honest builder take: a GPT is a system prompt with some tools and files bolted on, not an app. If you actually ship software, this is not your distribution channel, and the promised revenue sharing was vague at launch. But if you want a zero-code way to package a workflow for non-technical people, the discovery surface is real and it is attached to the biggest chatbot audience on the planet. Treat it as a marketing funnel, not a moat.

OpenAI Quietly Deletes Its "Military and Warfare" Ban

Around January 10, OpenAI rewrote its usage policies and the explicit prohibition on "military and warfare" applications vanished. The company framed it as a cleanup toward a smaller set of universal principles, but the practical effect was unmistakable: the hard line was gone.

Within days OpenAI confirmed it was already working with DARPA on cybersecurity tooling and with the Department of Defense, which it discussed in the same window as Davos. The policy still bars using the models to build weapons or harm people. Where that leaves logistics, intelligence analysis, and "defensive" cyber is exactly the gray zone everyone is now arguing about. If you build on these APIs, this is a reminder that the acceptable-use document is a living thing that can change under you without a press release.

Altman Starts a Money Hunt for Chip Fabs

On January 19, reporting surfaced that Sam Altman was talking to global investors, including SoftBank and Abu Dhabi's G42, about raising billions to build a network of semiconductor fabrication plants. The thesis is blunt: AI eats compute, compute means chips, and Altman does not believe there will be enough of them for years.

The eye-watering numbers that came later (the "trillions" framing) landed in February, so take January as the opening move. For anyone running models locally, the subtext matters. The people building the frontier are betting the binding constraint on AI is not algorithms or data, it is fab capacity and power. That tension between centralized mega-compute and the efficient, quantized models you can actually run at home is going to define the next couple of years.

The New York Times Fight Gets Personal

The Times had sued OpenAI and Microsoft in late December over training on its articles. On January 8, OpenAI fired back with a blog post calling the suit "without merit" and arguing the Times had cherry-picked and coaxed regurgitation through atypical prompting. OpenAI also revealed the two sides had been in partnership talks as recently as December.

This is the copyright case that actually has teeth, because the Times has the resources to litigate for years and a clean record of high-value text. Whatever the courts decide about fair use and training data will ripple straight down to anyone fine-tuning on scraped content. Watch this one: the outcome shapes what "legal" training data looks like for the rest of us.

Mixtral of Experts Gives Open Weights a Real MoE

On January 8, Mistral put the Mixtral of Experts paper on arXiv, documenting the Mixtral 8x7B model it had dropped via magnet link in December. It is a sparse mixture-of-experts: eight feedforward expert blocks per layer with a router picking two per token, so each token touches roughly 47B parameters of capacity while only activating about 13B at inference.

Apache 2.0, 32k context, and benchmarks matching or beating Llama 2 70B and GPT-3.5 on many tasks. For the homelab crowd, this is the headline that lasts longest. MoE means you get big-model quality at small-model inference cost, and the paper made the recipe legible instead of folklore. Quantized Mixtral became a staple of local rigs precisely because of this math.

Key Takeaways

• Distribution beats novelty. The GPT Store is not technically impressive, but it bolts a marketplace onto the largest chatbot audience alive, which is its own kind of leverage.
• Acceptable-use policies are not load-bearing. OpenAI deleting its military ban overnight is proof that what you are allowed to build can change without warning, so do not architect a business around a vendor's current TOS.
• Compute is the real bottleneck. Altman's fab fundraising signals the frontier labs think chips and power, not models, are the constraint for the rest of the decade.
• Training-data law is getting decided now. The NYT case is the one with the resources to set precedent on fair use, and it will define what counts as legal training data downstream.
• MoE is the open-weights cheat code. Mixtral showed builders how to get 70B-class quality at 13B-ish active cost, and the Apache license meant you could actually use it.