Midjourney opens an MCP for artists
Midjourney is launching a limited test of a Midjourney MCP, aimed squarely at creative and technical people who want to push its models to their limits. The company was explicit about the boundary: this is not for Midjourney inside SaaS products, but for purely artistic and creative exploration. Applications open to a small community first.
58% of web traffic is now bots
Vercel's latest network data shows 58.18% of all traffic was bot-originated over the past 30 days, up from 32% in January 2024. More than 60% of deployments are now agentic, up from roughly 3% in January 2026 — a striking measure of how fast machines have become the web's primary users.
MiniMax M3 is live on Vera Rubin
MiniMax M3 now runs on NVIDIA Vera Rubin NVL72 through vLLM, the company confirmed. MiniMax credited vLLM, NVIDIA, Red Hat, and its partners for the adaptation — an early sign that Rubin systems are reaching real production workloads rather than staying on the benchmark bench.
Snap a photo, let Grok buy it
Elon Musk demoed a new Grok Bot flow: take a picture of your credit card, drop it in the chat, and the agent scours the internet for the best deal before placing the order itself. It is a concrete glimpse of an agentic path for e-commerce, where the assistant handles price comparison and checkout end to end.
Why AlphaFold didn't solve protein folding
Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido revisit the "bitter lesson" of scaling to ask what is still missing: building AI that genuinely understands biology, not just one that predicts structures.
Companies are, in some sense, the original superintelligences.
Patrick Collison — on how agents will reshape the economy
The shift that made coding agents work
In a single chart, François Chollet argues AI has moved from a transductive paradigm — completing answers token by token — to an inductive paradigm that synthesizes reasoning chains on the fly. That switch, he says, is the big story of 2025 and 2026, and the reason coding agents finally work.
RoboJEPA proposes a scaling law for world models
Researchers scaled a JEPA predictor to 8B parameters and proposed the first scaling law for multi-embodiment world models, a result that drew attention from Yann LeCun and others in the robot-learning community.
Open-source model downloads slide ~30%
Interconnects has been tracking a sustained decline in HuggingFace downloads for the top open LLMs it monitors, with several models down roughly 30%. The drop may combine a change in how downloads are counted with a longer-term trend worth watching.
Google is testing a model called Carbon
Business Insider reports Google is internally testing a new model, Carbon, with employees saying its coding ability "feels like Opus 5.5." Carbon is deployed on Google's internal coding tool Jetski and follows the Barium-B version that shipped as Gemini 4 Argon on September 30.
vLLM's Semantic Router adds System One Auto
Kai 0.6B answers first; Vega 27B steps in as fallback, switching decision models on demand.
SGLang and NexRT split the decode
NexRT takes over decoding after SGLang handles prefill, with both engines sharing routed MoE weights.
Superhuman at everything, or not?
Researchers clash over whether we face AI that excels at every task, or a world of sharply uneven performance.
Data quality still amplifies reward hacking
Lambert argues the data industry's net output is too low, and is studying scalable, sample-efficient RL data.
What really slows Flow Matching?
New work proposes Velocity Scaling, challenging the idea that MSE training causes the speed deficit.
Tinker cuts prices on post-training
Efficiency gains lead to price cuts, as Tinker chases a post-training stack so cheap nobody builds their own.
Vidu Q4 Preview opens for trial
A next-gen flagship focused on expressive audio and video, with invite codes for early access.
RLVR round two, hands on
A new tutorial walks through clipped policy ratios, KL loss terms, and format rewards in GRPO.
TRL 1.15 stretches past 100k tokens
A new Triton kernel scales SFT and RL to long sequences on a single GPU.
REA reverse-engineers without source
An open-source project wires reverse-engineering tools into Claude Code, Codex, and Cursor.
Anthropic to ban needless abuse of its AI
The company will restrict sustained and needless abusive behavior toward its models.
Oil UI ships a design skill for agents
A Cola Skill that explores design directions, then refines against real screens — websites, apps, and backends.
Astra and Opus 5 clear Montezuma's Revenge
Both models beat the classic game; Metaculus will rebuild the test to settle an AGI wager.
Personal agents went from niche to boom
Instinct raised $1B at a $10B valuation; Meta, xAI, OpenAI, and Manus each shipped their own.
SAM 3.1 plus DINOv3 tracks surgical tools
Meta's combo holds onto instruments as a forearm sweeps across the frame.
TeslaSI is hiring for real-world superintelligence
Musk invites people who want to solve generalized real-world superintelligence to join.
People warm to AI that acts proactively
Users of Muse and Dot keep retelling the moment their AI did something useful unprompted.