Explore what's new in AI

Tavus Image-to-Replica turns a single image into a video avatar

Image-to-Replica drops the requirement for source footage by turning a single image into a video-ready avatar. That opens avatars for mascots, sketches, and synthetic characters.

Devin onboarding uncovers a bug worth $3M a year

Two days of Devin onboarding in Texas reportedly surfaced a bug worth $3M a year. That is the best argument yet for agent-assisted maintenance work.

SWE-ZERO-12M becomes the largest open agent trace dataset

The open trace corpus is now 5.7x larger than the previous record, which should matter for training and evals on agentic workflows.

Hugging Face crosses 1M datasets and pushes coding traces

Hugging Face crossed 1M datasets, and the next push is more open coding-session traces. That's a useful reminder that dataset scale is still compounding fast.

GitHub Copilot app enters technical preview

GitHub's Copilot app is in technical preview, bringing mobile access to chats and tasks. The app makes Copilot feel more like a workspace than a sidebar.

VS Code model picker groups models by provider

VS Code now groups the model picker by provider and shows provider names on recent models. That sounds small, but it makes multi-model switching much less error-prone.

Claude for Small Business bundles connectors and approval-gated workflows

Claude for Small Business packages connectors and workflows for ops, sales, HR, and finance, then waits for approval before sending money or email. That is a real business-agent shape.

Raycast publishes a v2 stack writeup for its multi-platform app

Raycast's v2 stack splits cleanly across Swift, React, Node, and Rust, with per-platform hosts and isolated extension workers. It's a strong architecture writeup for devtools teams.

ComfyUI lands Gemma 4, VOID, and BiRefNet in native workflows

ComfyUI now loads Gemma 4, VOID, and BiRefNet natively. That puts multimodal reasoning, video object removal, and segmentation into one workflow.

Pika MCP exposes creative models to agents

Installing Pika MCP gives agents access to creative models behind one interface. The pitch is fewer prompts and more reusable workflows.

Devin users automate daily standups with scheduled agent reports

Scheduled agents can auto-send standup summaries, ticket status, PR links, and test recordings. It's a nice low-friction pattern for keeping teams in sync.

Turbopuffer says longer agent queries improve FTS tradeoffs

The thread argues agents write longer queries than humans, and FTS engines can take advantage of that. It's a small but important shift in search tradeoffs.

Junwei Liao paper says agentic AI systems are the path to AGI

The paper argues agentic systems, not bigger base models alone, are the most foreseeable route to AGI. It breaks the problem into separate memory, tool-use, and self-improvement bottlenecks.

NandoDF publishes RL tutorial with notebooks and TeX source

The tutorial bundles policy-gradient notebook code and TeX source, so it can be translated and reused easily. It covers RL basics through tool use and self-improvement.

Vercel Firewall lets teams write WAF rules in natural language

You can now write WAF rules in plain language, like blocking /wp-admin or rate limiting /login. It lowers the friction for quick edge security changes.

fal launches FLUX.2 Pro Outpaint with per-side canvas expansion

FLUX.2 Pro Outpaint lets you extend any side of an image while keeping the original anchored. That should cut seams for reframing and product shots.

OpenRouter adds Recraft 4.1 image-generation variants

OpenRouter now exposes Recraft 4.1 variants for photorealism, SVG, and restraint-first product imagery. That split should make image model selection less guessy.

LangSmith Context Hub centralizes AGENTS.md, skills, and mcp.json

LangSmith Context Hub gives teams a place to store and version AGENTS.md, skills, and mcp.json. The point is to treat agent context like a shared artifact.

AEvo turns agentic evolution logs into a meta-edit loop

AEvo uses a meta-agent to edit the search procedure itself, not just the next candidate. The paper reports a 26% gain over the strongest evolution baseline.

mcporter 0.11.0 turns browser automation into steadier harness

mcporter is pitched as a more stable browser-automation CLI for agent tests, with fewer restarts and better MCP support.

Higgsfield Supercomputer routes tasks across models, skills, and memory

One prompt fans out to skills and the best model per subtask, then stitches outputs back together with memory and evals. That is a real orchestration stack.

ClaudeDevs show how to pre-warm prompt cache before real traffic

Send the system prompt first so Claude fills cache without answering. It can shave time-to-first-token on long prompts and repeated sessions.

Ollama adds Codex app support for local model selection

Ollama 0.24 can launch the Codex app against an open model, so local workflows no longer require a closed provider. Good for offline agent testing.

Figure hits 30 hours of autonomous package sorting on Helix-02

Figure pushed past its 8-hour goal and is now over 30 hours of continuous sorting, with no teleop. That is a real durability test for humanoids.