AI Roundup March 2024: Claude 3 Lands, Grok Goes Open, and Nvidia's Blackwell

TL;DR

March was the month Anthropic stopped being the polite second-place lab: Claude 3 Opus shipped and beat GPT-4 on the leaderboards. xAI dumped a 314-billion-parameter Grok-1 onto the internet under Apache 2.0, Nvidia revealed the Blackwell architecture at GTC, Stability put out the Stable Diffusion 3 research paper, and the EU AI Act cleared the European Parliament. A loud, builder-relevant month.

Claude 3 Opus Beat GPT-4, and Suddenly There Were Two Frontier Labs

On March 4, Anthropic launched the Claude 3 family: Haiku, Sonnet, and Opus, in ascending order of brains and cost. The headline was that Opus topped GPT-4 on a stack of standard benchmarks (MMLU, GPQA, math, coding), the first time a non-OpenAI model could credibly claim the frontier crown. For the better part of a year GPT-4 had been the uncontested ceiling. That ceiling just got a roommate.

The part builders cared about: a 200K-token context window (with Anthropic teasing 1M for select customers), genuinely strong vision, and a tiered lineup so you are not paying Opus rates to do classification. Sonnet, the middle tier, was free on claude.ai and good enough to be the daily driver for most apps. Haiku, the cheap-and-fast model, landed generally available on March 13.

There was also a delightful bit of AI folklore: during testing, Opus appeared to notice a planted "needle" sentence about pizza toppings was out of place in a haystack of unrelated text and basically said so. People read way too much into it, but it made for great marketing.

xAI Open-Sourced Grok-1: 314 Billion Parameters, Apache 2.0, Good Luck Running It

On March 17, xAI released the base weights and architecture for Grok-1 under an Apache 2.0 license. It is a 314B-parameter Mixture-of-Experts model (about 25 percent of weights active per token), and it is a raw pretrained checkpoint: no instruction tuning, no chat alignment, just the base.

For the local-AI crowd this was more flex than gift. 314B parameters means you are not loading this on your 4090, or your eight 4090s. It needs serious multi-GPU hardware just to do inference. But the gesture mattered: the largest openly-released model to date at the time, dropped as a torrent, with a permissive license that actually lets you build commercially. Whatever you think of the timing or the motives, "here are the weights, do what you want" is the kind of move that keeps the open ecosystem from calcifying into three closed APIs.

Nvidia's Blackwell: the Trillion-Parameter Era Gets Its Silicon

At GTC on March 18, Jensen Huang unveiled Blackwell, the successor to Hopper. The B200 GPU and the GB200 Grace-Blackwell superchip are built explicitly for trillion-parameter training and inference, with Nvidia claiming large reductions in cost and energy per token versus the H100 generation, plus rack-scale systems like the NVL72 that wire 72 GPUs into one giant coherent box.

You are not buying one. But if you rent GPUs (and you do, even if it is through three layers of abstraction), Blackwell is the thing that determines what frontier training runs cost in 2025, and therefore what the labs can afford to attempt next. It also reinforced the uncomfortable truth of this whole era: the most important AI company of the month did not ship a model.

Stable Diffusion 3's Research Paper Made the Case for Diffusion Transformers

Early in the month, Stability AI published the Stable Diffusion 3 research paper, detailing a new architecture that pairs a multimodal diffusion transformer with flow matching. The models scale from 800M up to 8B parameters, and the paper put up comparisons against DALL-E 3, Midjourney v6, and SDXL, with a particular emphasis on the thing open image models have always been bad at: rendering legible text inside images.

Weights were not public yet (that came later in the spring), and the company's finances were visibly wobbling, but the paper mattered as a signal. The diffusion-transformer recipe it leaned on is the same family of ideas powering the video and image models everyone is chasing. For anyone running local image pipelines, SD3 was the roadmap for where open image generation was heading.

The EU AI Act Cleared Parliament: the World's First Big AI Law Is Real

On March 13, the European Parliament adopted the AI Act by an overwhelming margin (523 in favor, 46 against). It is the first comprehensive, horizontal AI law from a major jurisdiction, and it sorts AI systems into risk tiers, bans a handful of uses outright (social scoring, untargeted facial-image scraping, certain biometric categorization), and puts transparency and documentation obligations on general-purpose models.

This was the parliamentary vote, not the finish line: formal Council adoption and entry into force came later in 2024, with obligations phasing in over years. But if you ship anything that touches EU users, this is now the regulatory baseline you build against. The era of "move fast, figure out compliance never" has an expiration date, and it has a postcode.

Key Takeaways

• The frontier is a race again. Claude 3 Opus beating GPT-4 means you should be benchmarking your own workloads across labs, not defaulting to one API out of habit.
• Tiered model families are the new normal. Haiku, Sonnet, Opus is the right mental model: route cheap tasks to cheap models and reserve the expensive brain for the hard calls.
• Open weights keep getting bigger and harder to run. Grok-1 is a reminder that "open" and "runnable on your hardware" are drifting apart, and quantization plus smart routing matter more every month.
• Compute is the real frontier. Blackwell sets the cost floor for everything the labs do next, so watch Nvidia's roadmap as closely as you watch model launches.
• Regulation is now a build constraint. The EU AI Act passing Parliament means provenance, documentation, and risk classification are becoming engineering requirements, not legal afterthoughts.