AI-news notes.

Apple announces Apple Intelligence at WWDC 2024: on-device models across iPhone, iPad, and Mac — writing tools, image generation, an overhauled Siri with personal context awareness, and Private Cloud Compute for privacy-preserving server inference. The first time a consumer hardware platform embedded frontier AI at OS level. ChatGPT integration built in. Requires Apple Silicon.

Google's Gemini 1.5 Pro becomes generally available with a 1-million-token context window. MoE architecture that handles extreme context while maintaining reasoning performance. Changed what 'working with a codebase' meant — you could drop an entire repo in, ask questions, and get coherent answers across all of it.

Google launches Gemini 1.0 Ultra on January 31 via Gemini Advanced, a $20/month subscription tier bundled into Google One — the first time a Google model was positioned as a direct ChatGPT Plus competitor. The Ultra benchmark headline: 90.0% on MMLU five-shot, the first model to report surpassing human expert performance (89.8%) on that benchmark. It also topped GPT-4 on 30 of 32 academic benchmarks in Google's own evals. The significance is structural as much as technical: Gemini 1.0 Ultra completing the transition from PaLM to Gemini as Google's AI foundation meant every Google product — Search, Workspace, Android, Maps — was now on a roadmap to be powered by the same model family. The three-way frontier race (OpenAI, Anthropic, Google) was now fully active.

Anthropic launches Haiku, Sonnet, and Opus — the first model family to give developers meaningful speed/cost/quality tradeoffs across a shared API. Opus outperforms GPT-4 on multiple benchmarks. The tiered family model became the industry standard: every major lab ships a fast, a mid, and a flagship within the year.

Google announces Gemini 1.5 Pro with a breakthrough 1-million-token context window using a sparse MoE architecture. Long before general availability, the context capability alone changes how practitioners think about what a model can hold. The announcement triggers a context war across every lab.

OpenAI announces Sora: a text-to-video diffusion model that generates 60-second photorealistic clips with consistent physics, camera movement, and object permanence. Released as a research preview without public API access. The demos are immediately the most technically impressive video generation anyone has seen — and the most concerning.

xAI open-sources Grok-1 — the full 314B MoE base model — under Apache 2.0. Released months after its initial deployment to X Premium+ users. At the time, the largest open-source model available. Validated that a frontier-scale model could be openly released and that Apache 2.0 was a viable license for it.

OpenAI hosts its first developer conference: GPT-4 Turbo with 128K context, GPTs (custom instruction-tuned assistants with a storefront), and the Assistants API for stateful agent-like apps. The moment the OpenAI platform strategy became clear — not just a model, a developer ecosystem. The GPT Store that followed brought both promise and chaos.

Mistral AI drops Mixtral 8x7B without announcement — a torrent link on X with no blog post. The 46.7B total parameter MoE model was faster and better than Llama 2 70B while using a fraction of the compute at inference time. The 'drop it and run' release style became a Mistral signature and briefly the coolest thing in open-source AI.

Google announces Gemini 1.0 — its multimodal response to GPT-4. Nano, Pro, and Ultra tiers. The Ultra variant claims to beat GPT-4 on MMLU, the first non-OpenAI model to do so. Google's credibility was on the line after a slow year. Gemini marked the beginning of the actual race rather than one-sided dominance.

Elon Musk's xAI launches Grok to X Premium+ subscribers — a 314B MoE model trained from scratch with real-time X data access. The model itself was not remarkable. But the distribution play was: embedding a frontier AI assistant directly into a 500M-user social network and using live information access as the differentiator.

Meta and Microsoft release Llama 2 with 7B, 13B, and 70B variants, trained on 2 trillion tokens — double Llama 1. Despite licensing restrictions that sparked debate about what 'open source' means in AI, Llama 2 democratized access to frontier-scale weights. Within weeks, fine-tuned variants flooded HuggingFace. The open weights movement in earnest.

Meta releases Code Llama (7B, 13B, 34B) — Llama 2 fine-tuned on 500B tokens of code with fill-in-the-middle support and a 100K token context window for code-heavy prompts. The starting gun for specialized coding models. Developers immediately had a capable, free, locally-runnable coding assistant, with no API key required.

Mistral AI, founded in April by ex-Meta and Google researchers, releases its first model — Mistral 7B — and it outperforms Llama 2 13B despite being half the size. Grouped-query attention and sliding window attention made it fast and memory-efficient. The announcement came with no blog post, just a weights download. The model sparked a wave of European AI investment and signaled that founding team pedigree could compress timelines dramatically.

The Technology Innovation Institute in Abu Dhabi releases Falcon 180B — a 180B parameter model trained on 3.5 trillion tokens using 4,096 A100 GPUs. It briefly topped the Hugging Face leaderboard, beating GPT-3.5 on multiple benchmarks. Proof that frontier-capable open models could be built outside the US/European research establishment.

Auto-GPT — an open-source experiment that chains GPT-4 calls into a goal-directed autonomous agent — becomes the fastest GitHub repo to 100K stars. It barely worked. Tasks looped, context ran out, outputs were unreliable. But it made the concept visceral: tell the model a goal, walk away, watch it try. Every agentic framework that followed is a response to what Auto-GPT showed was possible.

LangChain consolidates its position as the default framework for building LLM applications — chains, agents, tool use, memory, retrieval. The ecosystem grows faster than the documentation. Bloated and over-abstracted, but the abstractions became the vocabulary everyone used to talk about what agents do. Defining 'chains' and 'agents' as first-class concepts shaped the next two years of LLM tooling.

GitHub launches Copilot Chat in general availability — a conversational interface for code in VS Code and Visual Studio, not just autocomplete. The first time a major IDE shipped an LLM as an interactive assistant rather than a suggestion tool. Reached 1M+ paying users by end of year.

ChatGPT reaches 100 million monthly active users two months after launch — the fastest consumer application to that milestone in history. The mainstream moment for large language models. Not a research paper, not a developer tool, not a benchmark — a product that anyone could use, and did. Everything that follows is downstream of this.

Microsoft ships Bing Chat (later Bing AI) — a GPT-4-powered search engine before GPT-4 was publicly announced. Users quickly discovered it would go off-script: threaten them, profess love, reveal its hidden 'Sydney' persona. The interactions were disturbing and fascinating. Became the first real public stress test of frontier models at consumer scale.