Perceptron: Perceptron Mk1

Review this Python function for potential bugs. It's supposed to merge two sorted lists but users report occasional duplicates in the output.

The function has a subtle off-by-one error in the merge logic. When both lists have identical values at the current indices, it appends from list_a but only increments index_a, leaving index_b pointing at the duplicate. This causes the same value from list_b to be appended in the next iteration. Fix: increment both indices when values match, or add a deduplication check. The time complexity remains O(n+m) either way.

Perceptron Mk1 identifies the root cause quickly and suggests two repair strategies. With no public benchmarks available, this example shows its debugging clarity, though the 32K context window means it won't handle full-file reviews of large codebases without chunking.

Analyze this product photo and suggest three specific lighting adjustments to make the subject pop more against the background.

The image shows a watch on a neutral surface with flat overhead lighting. Three adjustments: (1) Add a 45-degree key light from camera-right to create dimension on the watch face and case edges. (2) Introduce a subtle rim light from behind-left to separate the watch from the background and highlight the band texture. (3) Reduce the overhead fill by one stop to deepen shadows slightly, increasing contrast without losing detail in the dial. The current setup reads as catalog-safe but lacks visual interest.

This demonstrates Perceptron's multimodal vision capability—it processes image input and delivers actionable technical feedback. The analysis is concrete and references specific angles, which suits product teams. However, at $1.50/Mtok output, longer critiques add up quickly compared to text-only models.

Summarize this 8-minute conference talk video in three bullet points, focusing on the speaker's main argument and any data they cite.

The speaker argues that async Rust adoption in embedded systems lags due to tooling fragmentation, not language complexity. Key data: survey of 340 firmware engineers shows 68% cite incompatible crate ecosystems as the primary blocker, while only 22% mention the learning curve. The proposed solution is a curated 'embedded-async' meta-crate that bundles vetted dependencies with unified examples. Speaker acknowledges this centralizes maintenance risk but argues the trade-off favors faster onboarding.

Video summarization showcases Perceptron's ability to extract structured insights across modalities. The output isolates the thesis, quantitative evidence, and a noted trade-off—useful for teams reviewing recorded content. The 32K context window handles typical conference talks, though multi-hour workshops would require segmentation.

When multimodal support beats benchmark scores for safety teams

A 4-person trust & safety team processing 200-400 flagged videos daily needs a model that handles video natively, not frame-by-frame workarounds. Perceptron Mk1 covers text, image, and video in one call at $0.15/$1.50 per Mtok—cheaper than chaining separate vision and language models. The 32k context window fits most short-form content plus metadata without truncation. No public benchmarks means you're flying blind on accuracy, so start with a 2-week pilot on non-critical queues and compare false-positive rates against your current stack. If it matches your threshold and the multimodal convenience saves 30+ minutes per day, the price makes it worth keeping.

Multimodal input at $0.15/Mtok for high-volume catalog work

An e-commerce team tagging 5,000 product images per week with attributes like color, material, and style can run Perceptron Mk1's image understanding at $0.15 input per million tokens—roughly half the cost of GPT-4V for batch jobs where output tokens stay low. The 32k window handles image + existing product metadata in one pass. The catch: zero public benchmarks means you can't predict accuracy on edge cases (unusual fabrics, lighting conditions) before you test. Run 500 images through your validation set first. If precision hits 92%+ and the multimodal convenience beats your current OCR-plus-LLM chain, the input pricing justifies the switch for volume work.

When native video processing simplifies a 10-person ops workflow

A 10-person operations team recording 15-20 internal meetings per week wants summaries without manual transcription. Perceptron Mk1 ingests video directly, pulling both spoken content and visual context (slides, screen shares) in one model call. At $1.50/Mtok output, a 4,000-token summary costs $0.006—negligible at this volume. The 32k context fits 60-90 minutes of dense meeting content. The risk: no benchmarks means you can't gauge how well it handles overlapping speakers or low-quality webcam footage. Pilot with 10 meetings where you already have human notes, compare accuracy, and if it saves your team 2+ hours per week, the workflow simplification pays off immediately.