Coolest part: the actuators that control the 10,000 PSI nitrogen cold-gas thrusters are from the Optimus robot project, which I found really surprising! It just shows how much technology development had to happen before we could get to this place.

Biggest question: how long do the tires last? With that much downforce and torque, this car looks to be a rubber shredder.

Small VLMs are quite capable at tool use. I am working on a problem where we want to use tool-use agentic VLMs to study videos generated by T2V models, and identify errors in them. I provide the model with tools for viewing frames from videos at higher fps, showing frames side-by-side for comparisons, zooming into different frame regions, and even optical flow based frame motion comparison.

Here’s something interesting I found: models like Qwen 3.5 VL 9B are absolutely capable of tool-use, but they are hampered by their inability to reason about the problem well, come up with hypotheses to test, and strategically use tools.

Research Question

If a bigger model, say Opus 5.5 or GPT-6 Sol, could provide the Qwen model with a strategy, can this strategy be injected into Qwen’s multi-turn conversation to improve tool use?

The short answer is YES! In fact, it produced some rather interesting outcomes… not only did Qwen become more accurate at tool-use, but its reasoning traces on the video were shorter, more succinct and more confident. This side effect is interesting: by modifying the expectation of how the Qwen model should think about the video, it started adapting that thinking setting to all future turns.

I find the photographic styles paired with the texture control to be really capable at making iPhone photos look unique and have richer, more contrasty colours. I find the grain to be too heavy-handed, and I’d like to see an update where we can control the amount of grain we want to add to the images.

The aperture control on iPhone 18 Pro is not useful in all situations, but I find myself reaching for it a lot when I am shooting pictures with a purpose… the f/4 gives you the sun stars, and the f/1.48 provides really nice background separation, especially when shooting at 2x zoom.

The benchmark contains 600 videos from state-of-the-art T2V models including Veo3, Seedance, and LTX-2, with over 1,600 expert annotations spanning physics, semantics, and anatomy.

Our experiments show that VLMs still lag far behind humans, with human performance nearly doubling the strongest baselines, highlighting the need for more robust perception and hallucination mitigation in automated video evaluation.

arXiv · Paper, code and dataset: spotlight-video.github.io

SPIKE is an inference-time framework that measures Bayesian Surprise as the belief update triggered by new visual evidence, identifying when incoming frames conflict with prior beliefs. SPIKE-RL further improves these belief hypotheses with GRPO using rewards from video captions.

Together, SPIKE and SPIKE-RL enable query-agnostic, surprise-weighted frame sampling that allocates more visual context to important moments, improving performance across five downstream video benchmarks.

arXiv

Explaining and understanding such out-of-distribution events requires models to extend beyond basic pattern recognition and regurgitation of their prior knowledge. To this end, we introduce BlackSwanSuite, a benchmark for evaluating VLMs’ ability to reason about unexpected events through abductive and defeasible tasks. Our tasks artificially limit the amount of visual information provided to models while questioning them about hidden unexpected events, or provide new visual information that could change an existing hypothesis about the event.

We curate a comprehensive benchmark suite comprising over 3,800 MCQ, 4,900 generative and 6,700 yes/no tasks, spanning 1,655 videos. After extensively evaluating various state-of-the-art VLMs, including GPT-4o and Gemini 1.5 Pro, as well as open-source VLMs such as LLaVA-Video, we find significant performance gaps of up to 32% from humans on these tasks. Our findings reveal key limitations in current VLMs, emphasizing the need for enhanced model architectures and training strategies.

arXiv · Website