The Archive

>Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-...

Engineer shares 6-month experience building agent memory systems using markdown filesystem with warm/archive layers; identifies conflicting facts as core challenge.

The deadline to apply or nominate for Startup Battlefield 200 is May 27. This is your shot at VC access, global visibility, TechCrunch coverage, and $100,000. Apply now.

MiniCPM5-1B released on HuggingFace: 1B-parameter model from CPM team, likely competitive efficiency benchmark for edge deployment.

Financial Times reports Heretic tool removes guardrails from Meta's Llama 3.3 in <10 minutes; 3,500+ decensored variants downloaded 13M times.

Early Bird savings for TechCrunch Disrupt 2026 in San Francisco end May 29 at 11:59 p.m. PT. Register now to save up to $410 before prices increase.

Reddit post discussing historical examples of technology pauses; unclear connection to AI governance or frontier capability development.

Found a really clean Claude Code use case that's not coding-coding. Remotion (React for video) means motion graphics are JSX components. So I describe what I want in plain English, Claude Code writes the component, I render. Lower thirds, intros, overlays, all reusable across videos. Iterations are seconds instead of the typical "drag clips around in CapCut for an hour" loop. Visual style is finally consistent across my channel because the components are shared. 13 min walkthrough on my channel, link in comments to avoid spam vibes.

Demis Hassabis argues that solving Erdős problems, while mathematically significant, falls short of demonstrating true scientific invention or reasoning.

User shares a system prompt technique using Office quote to reduce Claude verbosity and context overhead.

I tested Claude Opus 4.7 and Kimi K2.6 on the same coding agent task i.e. build an AI Fix Runner that takes a broken repo, runs its tests, identifies the failure, applies a patch, reruns the test, and exposes the final diff/logs through an API and UI. The goal was not to benchmark syntax completion or simple repo edits. I wanted to test model behavior on a less familiar integration path: shifting execution from local processes into remote sandboxes. I used Tensorlake specifically because the sandbox API is newer and integration-heavy. This made the test more about whether the model could re...

Reddit speculation on AI company cash burn and market consolidation by 2027, predicting funding scarcity for non-GAFAM players.

Numind releases NuExtract3, open-weight 4B multimodal VLM for document extraction and Markdown conversion under Apache-2.0.

Link: ramp.com

Mac Pro user discovers legacy AMD D700 GPUs now support Vulkan-based LLM inference via new drivers, reviving dormant hardware.

LimX Dynamics unveils Luna, a full-size humanoid robot with fluid motion capabilities, advancing embodied AI hardware.

Figure AI demonstrated robots performing continuous package sorting for 8 days in livestreamed operations, showing progress toward production-ready robotic systems.

Reddit user reports Claude struggles with constrained output length despite feedback, seeking workarounds.

MiMo-V2.5-coder released as open-weights coding model alternative to Qwen and DeepSeek for 128GB+ systems.

Elon Musk announces 0.5T parameter Grok model planned for next year, with open-weights release.

Reddit post with unclear title; insufficient content for professional AI analysis.

Reddit post warns that AI systems reflect user bias and prompting style, citing workplace example where team lead weaponized LLM criticism to validate negative performance opinions.

OpenAI reportedly recruiting NYC residents to install home cameras capturing daily activities for data collection.

llama.cpp PR addresses checkpoint creation inefficiency when context optimization tools modify conversation history in agentic workflows.

User demonstrates 1000 tokens/sec generation throughput on Qwen 3.6 27B with V100 GPUs at high batch sizes.

User reports Claude's safety filters blocking routine health questions, citing poor UX and considering switching to competitors.

4D Gaussian Splatting reconstructs 3D spatial data from 2D video footage; technique converts flat imagery into volumetric representations.

Developer built iOS/watchOS app and landing page using Claude Code, reached 1,500+ users by solving friction in law enforcement location-sharing workflows.

Who knows this pain?