Gemma – Google's open model family
Google released Gemma in February 2024 as its open-weight counterpart to Meta's Llama and Mistral's models: a family of models you can download, run on your own hardware,…
An independent analyst portal
The frontier of embodied AI
Foundation models. Voice and speech AI. Avatars and synthetic video. Agentic infrastructure. Edge runtimes. And the trust-and-safety surface around them all – analysed by hand, written to last.
Foundation Models
Gemma – Google's open model family
Google released Gemma in February 2024 as its open-weight counterpart to Meta's Llama and Mistral's models: a family of models you can download, run on your own hardware, fine-tune, and ship inside a product without calling an API. The models descend from the research that…
Continue readingWhat we cover
Seven areas of embodied AI. Each is an ongoing body of work, continuously updated as the field moves.
- Foundation Models Frontier model families and the labs behind them.
- Voice & Speech AI TTS, voice cloning, STT, conversational and call-center AI.
- Avatars & Vision Synthetic video, talking avatars, 3D characters, motion capture.
- Agentic AI Agent frameworks, autonomous coders, RAG and agent infrastructure.
- On-Device & Edge AI Local LLM stack, edge hardware, small open models, on-device privacy.
- Trust & Safety Provenance, deepfake detection, AI governance, safety research.
- Industry Tracker Living trackers – open-source AI weekly, funding monthly.
The model lineage
Eight years of foundation models, in three families. Hover any release to read what it changed.
GPT
GPT-1 Jun 2018 117M params The first decoder-only GPT, trained on BookCorpus. The thesis: pre-train broadly, fine-tune cheaply. Read the analysis → GPT-1 Jun 2018 GPT-2 Feb 2019 1.5B params Coherent long-form generation. Released in stages over misuse concerns. Read the analysis → GPT-2 Feb 2019 GPT-3 May 2020 175B params Few-shot learning from prompt examples. The "scale is enough" bet, vindicated. Read the analysis → GPT-3 May 2020 GPT-3.5 Nov 2022 first RLHF The first RLHF model in the line, and the engine behind the original ChatGPT. Read the analysis → GPT-3.5 Nov 2022 GPT-4 Mar 2023 multimodal Vision input added. OpenAI stops disclosing architecture from here on. Read the analysis → GPT-4 Mar 2023 GPT-4o May 2024 omni Native audio, vision and text through a single shared forward pass. Read the analysis → GPT-4o May 2024 GPT-5 Aug 2025 400K context A unified system that routes between fast inference and deep reasoning. Read the analysis → GPT-5 Aug 2025 GPT-5.5 Apr 2026 MMLU 92.4% Codename Spud. The current flagship as of mid-2026. Read the analysis → GPT-5.5 Apr 2026
Llama
Llama 1 Feb 2023 research-only Leaked publicly within a week, which kick-started the open-weight era. Read the analysis → Llama 1 Feb 2023 Llama 2 Jul 2023 permissive The first genuinely commercial open licence. An accident turned into strategy. Read the analysis → Llama 2 Jul 2023 Llama 3 Apr 2024 8B / 70B Redefined what open models could do at their parameter scale. Read the analysis → Llama 3 Apr 2024 Llama 3.1 Jul 2024 405B · 128K The first open model to genuinely rival the closed frontier on benchmarks. Read the analysis → Llama 3.1 Jul 2024 Llama 4 Apr 2025 MoE · multimodal Scout and Maverick: mixture-of-experts, natively multimodal, 10M-token context. Read the analysis → Llama 4 Apr 2025
Stable Diffusion
SD 1.5 2022 512px The foundational open image model, with a vast derivative ecosystem still alive today. Read the analysis → SD 1.5 2022 SD 2.0 late 2022 OpenCLIP The underrated stumble. Changing the text encoder broke the community contract. Read the analysis → SD 2.0 late 2022 SDXL Jul 2023 1024px The recovery. Larger UNet, base-plus-refiner pipeline, dual text encoders. Read the analysis → SDXL Jul 2023 SD 3 Jun 2024 MMDiT A new transformer backbone, but a rocky initial release on human anatomy. Read the analysis → SD 3 Jun 2024 SD 3.5 Oct 2024 8.1B The return to form, and the current generation as of 2026. Read the analysis → SD 3.5 Oct 2024
Recent essays
Alignment and AI safety from Anthropic's lens
Anthropic sells safety. That is the plainest way to describe a company founded in 2021 by former OpenAI researchers who left, in part, over disagreements about how carefully…
Age verification and digital identity
Practical reference in Avatars & Vision – 3D Avatar Pipeline. The specific question answered, with evidence.
Gemini – Google's model family explained
Gemini is Google DeepMind's frontier model family, launched in December 2023 and folded into the Bard-turned-Gemini consumer app in early 2024. It spans four tiers – Ultra, Pro,…
Hugging Face – the model hub and ecosystem
Hugging Face is the distribution layer for open AI. If you want to find, download, evaluate, fine-tune or serve an open-weight model, the odds are overwhelming that your path runs…
How to use ChatGPT – effective prompting
Most guides to ChatGPT teach you which buttons to press. The harder question, and the one that separates a frustrating tool from a genuinely useful one, is what the system is…
Mistral AI – the European open model lab
Mistral AI is the most credible independent AI lab based in Europe, and it has built its reputation on a hybrid strategy that few of its larger rivals have matched: release…
DeepSeek – the open reasoning model lab
DeepSeek is a Chinese AI lab, founded in 2023 and based in Hangzhou, that released an open-weight reasoning model called DeepSeek-R1 in January 2025 and claimed it was trained for…
Living trackers
Two pieces we republish on a fixed cadence – analyst synthesis, not aggregation.