NEWAgents can now see video via MCP.Try it now →
    Philip Bankier's cover

    Philip Bankier

    Philip’s expertise in RAG, data engineering, and AI agents makes him a natural fit to write about scaling multimodal data warehousing systems.

    Philip Bankier

    Recent Posts

    Multimodal Monday #45: Birds, Whales, and the End of Latency

    Multimodal Monday #45: Birds, Whales, and the End of Latency

    Your Weekly Multimodal AI Roundup (Feb 9 - Feb 16)

    February 17, 20266 min read
    Multimodal Monday #44: Agents Ship Code, Robots Skateboard

    Multimodal Monday #44: Agents Ship Code, Robots Skateboard

    Feb 2 - 9: GPT-5.3-Codex and Claude Opus 4.6 handle full software lifecycles from debug to deploy, MiniCPM-o 4.5 beats GPT-4o on vision tasks at 9B parameters running on-device, HUSKY skateboards using physics-based control, and TinyLoRA fine-tunes models with a single parameter.

    February 10, 20265 min read
    Multimodal Monday #43: Stop Looking, Start Seeing

    Multimodal Monday #43: Stop Looking, Start Seeing

    Week of Jan 26 - Feb 1: World models dominate as EgoWM simulates robot actions from single images, Drive-JEPA predicts what matters for driving, Project Genie halluccinates playable games, and Google's Agentic Vision turns image analysis into active exploration.

    February 3, 20266 min read
    Multimodal Monday #42: See, Click, Speak

    Multimodal Monday #42: See, Click, Speak

    Week of January 19 - January 25, 2026: D4RT turns video into 4D space, PersonaPlex enables full-duplex speech with persona control, EvoCUA hits #1 on OSWorld, and Linum V2 brings 720p video gen to 2B parameters.

    January 26, 20266 min read
    Multimodal Monday #41: Small Models Win, Vision Models Fail

    Multimodal Monday #41: Small Models Win, Vision Models Fail

    Week of January 12-18, 2026: ReinPool cuts retrieval storage by up to 1000x, vision models fail basic tests that 6-year-olds ace, ShowUI-Aloha learns to click and scroll from demonstrations, and agents manage their own memory with frameworks like AgeMem and SimpleMem.

    January 19, 20268 min read
    Multimodal Monday #40: Unified Search, Synthetic Worlds

    Multimodal Monday #40: Unified Search, Synthetic Worlds

    Qwen3-VL-Embedding unifies text, image, and video search in 30+ languages, HY-Video-PRFL trains 1.4x faster using video models as reward signals, PointWorld-1B simulates interactive 3D environments from single images, and Music Flamingo reasons about chord progressions and harmony.

    January 12, 20266 min read
    Multimodal Monday #39: Generative Giants, Retrieval Weaklings

    Multimodal Monday #39: Generative Giants, Retrieval Weaklings

    Weeks of December 29 - January 4, 2026: Research reveals why MLLMs like Qwen2-VL fall down, MegaRAG builds multimodal knowledge graphs without manual construction, HiStream generates 1080p video 107x faster, and Google DeepMind discovers geometric memory in sequence models.

    January 5, 20268 min read
    Multimodal Monday #38: World Models, Real-Time Generation

    Multimodal Monday #38: World Models, Real-Time Generation

    Week of December 16 - December 22, 2025: TurboDiffusion achieves 100-205x video generation speedup, WorldPlay generates interactive 3D worlds with geometric consistency, Step-GUI reaches SOTA on AndroidWorld & OSWorld benchmarks, and LongVie 2 produces 5-minute continuous videos.

    December 22, 20257 min read
    Multimodal Monday AI #37: Less-Is-More, Analogical Vision

    Multimodal Monday AI #37: Less-Is-More, Analogical Vision

    Week of Dec 9-14, 2025: Apple proves one attention layer beats dozens in diffusion models, MokA shows low-rank adaptation outperforms full fine-tuning, Adobe's relsim captures analogical relationships between images, and X-VLA controls different robot types with one transformer.

    December 15, 20256 min read
    Multimodal Monday 36: Factual Recall, Real-Time Video

    Multimodal Monday 36: Factual Recall, Real-Time Video

    Week of Dec 1-7, 2025: Research reveals why 11 of 14 VLMs fail at factual recall, BookRAG adds hierarchical structure to document retrieval, Adobe's RELIC and Alibaba's Reward Forcing enable real-time interactive video, and Microsoft's 0.5B parameter TTS model runs in real time.

    December 8, 20256 min read
    Multimodal Monday 35: Small Models, Modular Vision

    Multimodal Monday 35: Small Models, Modular Vision

    Week of Nov 24-30, 2025: Alibaba's 6B Z-Image impresses, Tencent's 1B HunyuanOCR beats larger models and APIs, VisionRAG uses 6-9x less memory than ColPali, and RynnVLA-002 boosts real-world robot success by 50%.

    December 1, 20256 min read
    Multimodal Monday #34: Visuals Coherence, Semantic Vision

    Multimodal Monday #34: Visuals Coherence, Semantic Vision

    Week of Nov 17-23, 2025: Nano Banana Pro creates coherent visualizations, SAM 3 segments by concept not pixels, HunyuanVideo 1.5 leads open-source video, and Step-Audio-R1 matches Gemini 3 Pro on audio reasoning.

    November 24, 20258 min read
    Multimodal Monday 33: Physical AI, Human Vision

    Multimodal Monday 33: Physical AI, Human Vision

    Week of November 10 - November 16, 2025: Pelican-VL gives humanoid robots spatial intelligence, DeepMind teaches AI to see like humans, Marble creates 3D worlds from single images, and Meta opens speech recognition to 1,600+ languages.

    November 17, 20255 min read
    Multimodal Monday 32: Multi-Query Retrieval, Streaming Video

    Multimodal Monday 32: Multi-Query Retrieval, Streaming Video

    Multimodal Monday 32: AMER shows 4-21% gains on complex queries by generating multiple embeddings, Adobe MotionStream hits 29 fps with interactive motion controls, Step-Audio-EditX edits voice emotion and style through text prompts, and GEN-0 trains robots for general skills.

    November 10, 20256 min read
    Multimodal Monday #31: Visual Thinking, Longer Video

    Multimodal Monday #31: Visual Thinking, Longer Video

    Google Latent Sketchpad lets models sketch thoughts before acting, Amazon Nova MME unifies search, Emu3.5 matches Google's Nano Banana locally, BEAR reveals why AI fails physical tasks.

    November 3, 20256 min read