Explore what's new in AI

Research model adapts an 8B autoregressive model for blockwise diffusion

A linked 8B-scale research model combines blockwise continuous diffusion, flow map distillation, and self-speculative decoding after adaptation from an autoregressive model. The post does not report benchmark results.

OpenBMB's EC²-Bench reports 73% of cloud-only performance at 15% cost

EC²-Bench compares four collaboration strategies on 128 tasks and measures privacy exposure. One setup achieves about 73% of cloud-only performance at roughly 15% of its cost.

Researcher tests flow map models as speculative decoding drafters

A researcher discusses results using flow map language models as speculative decoding drafters. The cited advantage is capturing token dependencies in one or two model passes.

Developer publishes Claude-generated guide to Megakittens megakernel internals

Maharshi shares a manual generated in one pass by Claude explaining Hazy Research's Megakittens megakernel framework. The post does not describe independent verification of the guide.

Report: Perplexity reportedly packages pplx-decider-flash-v1.2 as a local Mac API

A report describes a Mac application serving pplx-decider-flash-v1.2 through a local decision API. The model occupies 5.8 GB and requires 16 GB of memory.

Devin workflow builds Unreal Engine games on AWS c7i.8xlarge instances

A shared setup uses Devin Outposts on c7i.8xlarge instances to build games, then g6.2xlarge instances with NVIDIA L4 GPUs to playtest them. The workflow also includes a reusable prompt.

Operator reports 6.77T tokens served by one rack in 24 hours

Thdxr reports 6.77T tokens served in 24 hours, with a median time to first token of 1.16 seconds and median throughput of 143 tokens per second. A flaky tray is hurting tail latency.

SGLang publishes DeepSeek serving optimizations without numerical performance claims

SGLang published a technical post covering DeepSeek performance results, engineering optimizations and its production serving stack. The visible announcement provides no numerical performance claims.

DoubleAI claims Maxwell makes programs 2.1× faster with live binary rewrites

DoubleAI's sponsored Maxwell thread claims verified live binary rewrites without source changes. Reported examples include 2.1× faster Whisper and 3.5× faster graph analysis.

Google begins Argon rollout to eligible subscribers through Antigravity

Google's rollout notice says Argon is reaching eligible subscribers through Antigravity desktop and CLI, plus the Gemini platform API. Sundar Pichai says automated research loops contributed to its development.