Skip to content
Back to archive
AI Intelligence

Mar 2023

5 events · Current month →

Mar 2023

Landmark
Meta

LLaMA-1 Leaks Online — Open-Weight Era Begins

Meta's LLaMA-1 models (7B–65B parameters), initially shared only with select researchers, leaked publicly within days and triggered an immediate wave of community fine-tuning that produced models like Alpaca and Vicuna within weeks. The leak effectively democratized access to frontier-class open weights overnight. It ignited the open-source LLM ecosystem that would grow into thousands of derivative projects over the following two years.

Why it mattersThe leak proved that open weights would spawn a fast-moving derivative ecosystem, forcing closed-model providers to compete on more than raw capability.

Try itIn retrospect, the pattern to notice is that the Alpaca and Vicuna fine-tunes showed instruction-tuning a base model could be done cheaply on consumer hardware.

MetaLLaMAOpen SourceOpen WeightsCommunity
Read source

Mar 2023

Landmark
OpenAI

GPT-4 Released — Professional-Grade Multimodal AI

OpenAI released GPT-4, the first frontier large language model capable of processing both text and images as inputs. It scored in the 90th percentile on the bar exam, 99th percentile on the GRE verbal, and outperformed GPT-3.5 substantially across professional and academic benchmarks. Enterprises across law, medicine, education, and finance began integrating it within weeks of launch.

Why it mattersGPT-4 established that LLMs could reliably pass professional licensing exams, which meant regulated industries could no longer defer AI adoption as a fringe concern.

Try itIn retrospect, the pattern to notice is that GPT-4's system-prompt fidelity made prompt engineering a viable alternative to custom fine-tuning for most enterprise use cases.

OpenAIGPT-4MultimodalBenchmarkProfessional
Read source

Mar 2023

Major
Anthropic

Anthropic Launches Claude — Constitutional AI in Production

Anthropic released Claude, its first publicly available model, trained using Constitutional AI — a technique where a set of written principles guides model behavior rather than relying purely on human feedback labels. Claude was noted for a longer context window than GPT-3.5 and a cautious, thoughtful conversational style. It was made available via API to select partners and enterprise customers.

Why it mattersConstitutional AI introduced a way to specify model behavior with written principles, giving developers a more interpretable alignment lever than pure RLHF reward models.

Try itIn retrospect, the pattern to notice is that explicit written principles in system prompts — not just fine-tuning — remain effective for shaping Claude's output style and refusal behavior.

AnthropicClaudeConstitutional AISafetyRLHF
Read source

Mar 2023

Major
Google

Google Launches Bard — Search Giant Responds to ChatGPT

Google launched Bard, a conversational AI built on its LaMDA language model, as a direct competitive response to ChatGPT's rapid growth. The launch was marred by a factual error in a promotional demo that contributed to Alphabet's stock dropping and erasing roughly $100 billion in market capitalization in a single day. Bard later transitioned to the Gemini brand and underlying Gemini models in February 2024.

Why it mattersThe costly demo error showed that factual accuracy in LLM demos directly affects market capitalization, making grounding and hallucination mitigation a board-level concern.

Try itIn retrospect, the pattern to notice is that grounding every AI-generated claim against a retrieval source before presenting it to users prevents the class of errors that Bard's launch made public.

GoogleBardLaMDAChatbotSearch
Read source

Mar 2023

Major
OpenAI

ChatGPT Plugins — AI Gets Internet Access and Real-World Tools

OpenAI introduced plugins for ChatGPT, allowing the model to browse the web in real time, execute code in a sandbox, and connect to third-party services such as Expedia, Klarna, and Wolfram Alpha. This was the first mainstream demonstration of LLMs operating as tool-using agents capable of taking real-world actions beyond generating text. The plugin architecture was the conceptual precursor to the later GPT Store and OpenAI Assistants API.

Why it mattersPlugins demonstrated that LLMs calling external tools could handle multi-step tasks, shifting developer focus from prompt engineering alone to agent orchestration patterns.

Try itIn retrospect, the pattern to notice is that the tool-calling design — model decides when to call, receives structured output, then reasons further — became the template for today's function-calling APIs.

OpenAIChatGPTPluginsTool UseAgentsWeb Browsing
Read source