AI Roundup March 2023: GPT-4 Lands, Claude Arrives, and 30,000 People Sign a Pause Letter

TL;DR

March 2023 was the month the LLM race went fully public. OpenAI dropped GPT-4 with image input and ridiculous benchmark scores, Anthropic launched Claude, Google opened Bard to the public, and Microsoft wired Copilot straight into Office. Then, a week after GPT-4, more than 30,000 people including Yoshua Bengio and Elon Musk signed an open letter asking labs to pause for six months. The hype and the backlash arrived in the same four weeks.

GPT-4 Ships: Multimodal, Benchmark-Crushing, and Closed as Ever

On March 14, OpenAI released GPT-4. It accepts image and text inputs (text output only at launch), and the headline numbers were the kind that make people stop scrolling: it cleared a simulated bar exam in roughly the top 10 percent of test takers, where GPT-3.5 sat near the bottom 10 percent. OpenAI also claimed it was 82 percent less likely to produce disallowed content and meaningfully more factual than its predecessor.

For builders, the immediate reality was less glamorous. GPT-4 landed behind ChatGPT Plus with a usage cap, and API access was a waitlist. Image input was demoed but not generally available. And true to form, OpenAI published almost nothing about architecture, parameter count, or training data, citing competition and safety. The technical report read more like a capabilities brochure than a paper.

The takeaway: the best general-purpose model on the planet got dramatically better and dramatically more opaque on the same day. If you build on closed APIs, your ceiling just went up. If you care about open weights, this was a reminder of who holds the cards.

Anthropic Launches Claude, and the Frontier Becomes a Two-Horse Race

Also on March 14, Anthropic took Claude public. After months as a private beta talked about mostly in safety circles, Claude shipped in two flavors: the full model and a faster, cheaper Claude Instant, both available through an API and early partners like Quora's Poe and Notion.

What mattered here was not just a second strong model. It was Anthropic's pitch: a model trained with Constitutional AI, an approach where the system critiques and revises its own outputs against a written set of principles instead of relying entirely on human labelers for every harmful case. Whether or not you buy the safety framing, having a credible second frontier lab with a real product changes the negotiating power of everyone building on top.

Microsoft Puts Copilot in Office, Google Opens Bard to the Public

The distribution war ran in parallel to the model war. On March 16, Microsoft announced Microsoft 365 Copilot, embedding GPT-4-class generation directly into Word, Excel, PowerPoint, Outlook, and Teams. Drafting documents, summarizing threads, and generating slide decks from a prompt went from demo to roadmap for hundreds of millions of seats.

Five days later, on March 21, Google opened Bard to the public via a waitlist in the US and UK. Bard initially ran on a lightweight version of LaMDA, and the reception was lukewarm next to GPT-4. But the signal was unmistakable: the company that invented the transformer was now shipping a consumer chatbot defensively, on someone else's timeline.

Why this combo matters

• Models are becoming features. The same week GPT-4 launched, it was already being stuffed into the most-used productivity software on earth.
• Defaults beat demos. Microsoft's advantage is not a better model, it is a billion existing users.
• Google blinked. Shipping Bard in a hurry told you how seriously they were taking the threat.

Stanford Alpaca Shows You Can Clone Instruction-Following on a Budget

While the giants traded press releases, Stanford researchers quietly published Alpaca on March 13. They took Meta's LLaMA 7B and fine-tuned it on 52,000 instruction-following examples generated cheaply using OpenAI's own models, for a compute cost of a few hundred dollars. The result behaved a lot like early instruction-tuned GPT-3.5 on many tasks.

For the local-AI crowd, Alpaca was a lightning bolt. It made the recipe legible: take a decent open base model, generate synthetic instruction data, fine-tune cheaply, and get something genuinely useful that runs on hardware you own. It also raised the licensing and terms-of-service questions that would dominate the open-model conversation for the rest of the year. The web demo came down fast, but the method was already out.

30,000 Signatures Later: The Open Letter to Pause Giant AI Experiments

The backlash crystallized fast. On March 22, the Future of Life Institute published an open letter calling on all AI labs to immediately pause for at least six months the training of any system more powerful than GPT-4. It went on to gather more than 30,000 signatures, including Yoshua Bengio, Stuart Russell, Elon Musk, and Steve Wozniak.

Reactions were predictably split. Critics argued a voluntary pause was both unenforceable and a gift to whoever ignored it. Sam Altman pushed back that the letter was missing technical nuance, and noted OpenAI was not training GPT-5 and would not for some time. The letter did not pause anything. What it did do was drag AI-risk arguments out of niche forums and onto front pages, setting up the regulatory and Senate-hearing season that followed.

Key Takeaways

• The frontier is now a race, not a monopoly. GPT-4 and Claude going public the same day means you can architect around more than one vendor, and you probably should.
• Distribution is the real moat. Microsoft embedding Copilot into Office showed that owning the surface matters as much as owning the model.
• Open and cheap is catching up to closed and expensive. Alpaca proved a few hundred dollars can buy useful instruction-following on hardware you control.
• Closed got more closed. GPT-4 shipped with almost no technical detail, a trend builders depending on transparency should plan around.
• The safety debate went mainstream. A 30,000-signature pause letter changed zero training runs but reshaped the policy conversation for the rest of the year.