AI Roundup September 2024: OpenAI's o1 Thinks Out Loud, Llama 3.2 Gets Eyes, Newsom Kills SB 1047

TL;DR

September was the month inference-time compute went mainstream. OpenAI dropped o1, the first model trained to think before it answers, and reset everyone's mental model of what a frontier release looks like. Meta gave Llama eyes with the 3.2 vision and edge models, Alibaba's Qwen 2.5 quietly became the open-weights model to beat, and California Governor Gavin Newsom vetoed SB 1047, sparing the open-source crowd from the most aggressive AI liability bill in the country. Builders won on every front except regulation, where nobody got what they wanted.

OpenAI Ships o1: The Model That Thinks Before It Speaks

On September 12, OpenAI released o1-preview and o1-mini, the first models in a new line trained to spend compute reasoning through a problem before producing an answer. Instead of one fast forward pass, o1 generates a long internal chain of thought, and on hard physics, chemistry, biology, and competition math it lands roughly at PhD level. This is the release that made "inference-time compute" a phrase you could not escape for the rest of the year.

The catch for builders: the chain of thought is hidden, you pay for those reasoning tokens you cannot see, and it is slow and pricey compared to GPT-4o. o1-mini is the interesting one for most of us. It is about 80 percent cheaper than o1-preview and punches well above its weight on coding and math, which is exactly the workload where a few seconds of extra thinking actually pays for itself.

The honest take: o1 is not a GPT-4o replacement, it is a different tool. If your task is a quick chat or a RAG lookup, o1 is the wrong call. If your task is a multi-step proof, a gnarly refactor, or a debugging session where a wrong answer costs you an hour, paying for tokens you cannot read suddenly looks like a bargain.

Meta Llama 3.2 Gives Open Weights Vision, and a Model That Fits on Your Phone

At Meta Connect on September 25, Llama 3.2 arrived with two things the open ecosystem badly wanted. First, real vision: the 11B and 90B models are Meta's first open multimodal Llamas, handling charts, documents, captioning, and visual grounding, and positioned as drop-in replacements for their text-only equivalents. Second, and more interesting for the homelab crowd, genuinely tiny text models at 1B and 3B with a 128K context window, built to run locally on phones and edge hardware.

Those small models shipped with day-one support for Qualcomm and MediaTek silicon and Arm optimization, which is the part people slept on. A 3B model that summarizes, rewrites, and follows instructions entirely on-device, no API, no per-token bill, no data leaving the hardware, is the kind of thing that quietly rewires what you build. The vision models grab headlines, but the 1B and 3B are the ones that show up in your projects.

Qwen 2.5 Quietly Becomes the Open-Weights Model to Beat

On September 19, Alibaba released the Qwen 2.5 family, and it deserved more noise than it got. The lineup runs from 0.5B all the way to 72B, most of it under Apache 2.0, pretrained on 18 trillion tokens, with specialized Qwen 2.5-Coder and Qwen 2.5-Math variants alongside the base models. The 72B Instruct model traded blows with, and on several benchmarks beat, Llama 3.1 405B and Mistral Large 2 while being a fraction of the size.

For anyone running models locally, this is the headline that matters. A 72B you can actually serve outperforming a 405B you mostly cannot is a better deal than any closed API announcement, and the permissive license means you can ship it commercially without a lawyer in the loop. Qwen 2.5 is the moment the open-weights frontier stopped being a Llama monoculture.

Newsom Vetoes SB 1047, and the Open-Source Crowd Exhales

On September 29, Governor Gavin Newsom vetoed SB 1047, the most far-reaching AI safety bill in the United States, after the California legislature passed it in late August. The bill would have imposed safety protocols, third-party audits, and a mandatory "kill switch" on developers of the largest frontier models, and held companies liable for catastrophic harms.

The fight split the industry in a way worth remembering. Anthropic and the Center for AI Safety backed it; Meta, OpenAI, and a16z fought it. Newsom's stated reason for the veto was that targeting only the biggest models was the wrong cut, since a smaller, cheaper system could be just as capable of harm. Whatever you think of the bill, the veto is a reprieve for open-weights releases that a strict liability regime could have frozen. It is not the end of California AI regulation, just the end of this round.

Key Takeaways

• Reasoning is a new axis, not a new default. o1 proved that spending compute at inference time buys real capability, but it is a specialist tool, so route hard problems to it and keep cheap models for everything else.
• Local AI got genuinely usable. Llama 3.2's 1B and 3B models running on phone-class hardware mean on-device inference is now a real product decision, not a science project.
• Open weights caught up. Qwen 2.5-72B beating models six times its size means you can self-host frontier-adjacent quality under a permissive license, no API contract required.
• Smaller models can win on price-performance. o1-mini and Qwen's mid-size models both show the same lesson: the biggest model is rarely the right one for the job.
• AI regulation is unsettled. The SB 1047 veto buys the open ecosystem time, but the liability question did not go away, so build with the assumption that rules are coming.