Explore what's new in AI

YacineMTB uses an AFK agent to quantize a network for 1k inferences per second

The workflow shows what agentic engineering looks like in practice: write the plan, leave it alone, and let the machine run experiments to hit a perf target.

OpenRouter team spend charts show GPT-5.5 dominating weekly usage

The spend leaderboard is a useful reminder that model adoption is not just about preference, it's about budget. GPT-5.5 is soaking up a lot of real usage.

Reflex shows computer use can cost 45x more than structured APIs

In one task, vision agents burned 551k tokens and still missed work, while API agents finished fast with far fewer tokens. That's a hard ROI lesson for computer-use teams.

Codex app connects to local Gemma 4 through Ollama

The Codex app can point at a local model, not just hosted endpoints. That gives teams a path to private, offline, or cost-controlled workflows.

RepoPrompt cuts workspace sidebar latency with a big performance pass

The post highlights a concrete UX improvement in a tool builders use daily. These are the small speed wins that matter when context windows are crowded.

Claude Code /insights scans sessions for workflow gaps and custom prompts

Run /insights and Claude Code reviews your history for repeated mistakes, workflow gaps, and prompts to paste back. It's a built-in coaching loop for power users.

OpenAI says Codex is gaining traction with scientists across the hard sciences

The question to answer is what Codex still cannot do in math, physics, chemistry, and biology. The reply thread is a useful wishlist for future product work.

Magnific workflow chains one photo into a full AI video ad recipe

The demo shows a repeatable prompt stack for taking one reference image into a full spec ad. Useful if you care about fast creative iteration, not just pretty outputs.

Claude Code powers a content workflow with skills, hooks, and weekly voice updates

The thread breaks down a real production workflow: skills beat giant docs, hooks score hooks, and the voice profile self-updates from performance data. That's a strong playbook for agent-assisted content.

Codex becomes a daily driver for knowledge work at Every

The interview shows Codex moving beyond coding into writing, recruiting, and operations. That's a useful signal that agent value is spreading across the org.

v0 cuts preview load time by 50%

A 50% faster preview load is the kind of incremental speedup that changes daily usage. It removes friction at the exact moment builders want feedback.

Google DeepMind partners with EVE Online for agent memory research

EVE gives DeepMind a safe, messy sandbox for memory and long-horizon planning. That's a better testbed than toy environments for persistent agents.

OBLIQ-Bench pushes retrieval systems on opaque search queries

OBLIQ-Bench targets the first-stage retriever, where current systems miss relevant docs even when a reasoning model would spot them. That's a useful new stress test.

ProgramBench asks models to rebuild real software from binaries

The benchmark is brutal: none of the tested models fully solved a task, and the best system only passed 95% of tests on 3% of programs. Software recreation remains hard.

Gemma 4 lands in Code Arena and shifts the frontier for frontend webdev

The leaderboard shows Gemma 4 on the Pareto frontier at a very low blended price. That's a strong open-model signal for UI-heavy coding tasks.

Google updates AI Mode and AI Overviews with deeper links and source previews

Search now exposes more next-step links, news subscriptions, and creator context. That's a notable move toward citation-aware AI search.

Manus Projects learn from every task and update instructions automatically

Manus now turns reusable context into suggested changes for project files, instructions, and skills. That reduces repeated setup and keeps teams aligned.

OpenAI updates ChatGPT to GPT-5.5 Instant and adds memory sources

The default model shifts, memory now shows its sources, and spreadsheet support expands. That makes ChatGPT more personal and more inspectable.

Cursor 3.3 shows a breakdown of agent context usage

The new view helps teams diagnose where context is going across rules, MCPs, skills, and subagents. That's practical debugging for large agent setups.

Amp rebuilds its CLI as Neo with remote control and server-side execution

Amp's new CLI pushes more of the agent loop to the server, trims data transfer, and keeps native scrollback. That's a concrete UX and latency upgrade.

VS Code adds Agent Mode customization with skills, hooks, and custom agents

Microsoft is exposing more of the agent harness, so teams can tune prompts, hooks, and task-specific agents inside the editor. That's where workflows get real.

OpenAI adds subagents to Codex for parallel specialist work

Codex can split a job across focused subagents and merge the results back. That's a big unlock for reviews, big codebases, and multi-threaded tasks.

Exa MCP brings people, companies, and paper search into ChatGPT

Search becomes a first-class ChatGPT app, so research queries can pull Exa's sources in one place. Useful for deep research and agent workflows.

xAI ships Image Generation Quality Mode on its API

The API exposes a higher-fidelity image path that already powers hundreds of millions of generations. That's a useful production knob for teams shipping visual generation.