Lovable shifts from app builder to general agent
Lovable is pivoting away from an app-making focus toward a general-purpose agent.
ben's bites•6h ago
vendor
📰 DailyMe
My personalized AI news feed, curated from newsletters and deduplicated automatically.
Powered by OpenHands + Claude Sonnet 4 • Updated every 30 minutes
Quick links: All stories · Starred · Social Top Stories
99% SOTA memory system for agents
A new memory system for agents reportedly reaches roughly 99% state-of-the-art performance.
ben's bites•6h ago
research
The Minimalist Entrepreneur becomes agent skills
Sahil has converted his book into a set of agent skills aimed at founders.
ben's bites•6h ago
launch
Podcast: Engineering practices that make coding agents work
Simon Willison discusses engineering practices that help coding agents succeed.
ben's bites•6h ago
podcast
The End of Coding: Karpathy on agents and autoresearch
Andrej Karpathy explores agents, autoresearch, and the emerging loopy era of AI in a must-read/listen piece.
ben's bites•6h ago
opinion
How to use Claude Code to design
A Tailwind founder shares a walkthrough on using Claude Code for design workflows.
ben's bites•6h ago
tutorial
Ghost Pepper brings local hold-to-talk STT to macOS
Ghost Pepper offers a fully local, hold-to-talk speech-to-text experience on macOS.
ben's bites•6h ago
launch
Deploy a team of OpenClaw agents securely
A resource explains how to deploy multiple OpenClaw agents with secure controls.
ben's bites•6h ago
tutorial
Agents should interview you
The author argues that having an agent interview you captures preferences and helps overcome blank‑page paralysis, sharing how it shaped his course planning.
ben's bites•6h ago
opinion
Claude Code adds recurring tasks and computer-use automation
Claude Code can schedule recurring cloud tasks and, when connectors are missing, drive apps directly on your computer; Cowork now supports projects.
ben's bites•6h ago
launch
Factory Missions launches long-running agents for end-to-end apps
Factory Missions introduces long-running agents that plan and execute large software projects like full app builds.
ben's bites•6h ago
launch
ChatGPT adds file library as OpenAI plans a superapp
ChatGPT now stores uploaded files in a library for easier reuse, while OpenAI is moving toward a simplified superapp experience.
ben's bites•6h ago
launch
Cursor debuts Composer 2 model and Glass interface
Cursor released Composer 2, revealed to be tuned from Kimi 2.5, and launched the Glass UI; its self-benchmark comparisons sparked criticism.
ben's bites•6h ago
launchbenchmark
SpaceX, Tesla and xAI open TERAFAB chip facility
The companies launched TERAFAB, described as the largest chip manufacturing facility with 1TW/year capacity.
ben's bites•6h ago
launch
Sequoia's Shaun Maguire argues xAI will win
A Sequoia partner claims the market is underestimating xAI and outlines why it will dominate AI.
ben's bites•6h ago
opinion
Codebase to Course makes learning codebases visual
Codebase to Course is a skill that converts codebases into more visual, interactive learning experiences.
ben's bites•6h ago
launch
Luma Labs releases Uni-1 image generation model
Uni-1 adds a canvas workflow and multiple outputs per prompt, though generating many outputs can be slow.
ben's bites•6h ago
launch
Guide claims to improve GPT 5.4 frontend design in Codex
A guide suggests ways to boost GPT 5.4's frontend design quality and adds a frontend skill in Codex.
ben's bites•6h ago
tutorial
Cord lets agents build task trees on the fly
Cord is highlighted as a flexible orchestration tool that allows models to split work into parallel tracks and share context without hardcoded plans. It aims to address rigidity in early orchestration frameworks.
Ben Lorica•7h ago
vendor
Emdash runs multiple coding agents in parallel workspaces
Emdash provides a workspace-oriented orchestration approach that lets developers run multiple coding agents concurrently in isolated environments. The goal is to reduce the friction of juggling terminals and serial runs.
Ben Lorica•7h ago
vendor
Operational skill stores shift agent memory toward procedures
The piece notes a move from chat-history memory toward procedural skill stores and context files that save successful workflows as reusable instructions. This approach aims to improve reliability while reducing compute costs.
Ben Lorica•7h ago
opinion
AGENTBENCH study finds AGENTS.md context files hurt results
A study benchmarking coding agents on standard tests and the new AGENTBENCH shows auto-generated context files lowered success rates and increased inference costs, with only modest gains from developer-written files. The findings suggest guidance files are not a guaranteed improvement.
Ben Lorica•7h ago
benchmark
Leo Meyerovich: teams get what they measure in agent evals
The newsletter cites Meyerovich’s view that teams should keep agent components only if they improve measured outcomes such as task success, speed, safety, or cost. It emphasizes defining clear evals rather than relying on intuition.
Ben Lorica•7h ago
opinion
Dean Wampler outlines the open-source PARK stack for agent experiments
Wampler’s article describes the PARK stack—built on PyTorch, AI models and agents, Ray, and Kubernetes—as a foundation for running computationally intensive agent experiments at scale. The focus is on enabling rigorous evaluation for production readiness.
Ben Lorica•7h ago
vendor
Why smarter agent architecture does not always improve results
The essay argues that building reliable AI agents requires rigorous engineering and evaluation, not just layering on more architectural components. It cautions that complexity can add cost and coordination overhead without improving real-world performance.
Ben Lorica•7h ago
long_formopinion
Cursor admits Kimi K2.5 is best open-source model
A perplexity-based evaluation puts Kimi K2.5 on top while commenters debate methodology and training claims.
AINews•13h ago
benchmark
Honest take on running nine RTX 3090 GPUs for AI
The post outlines PCIe bottlenecks, stability issues, and power constraints, recommending alternatives like Proxmox or PCIe switches.
AINews•13h ago
opinion
Do buyers regret getting an NVIDIA 5090?
Redditors debate whether to buy now or wait, citing price increases, rentals, and performance for gaming and local AI.
AINews•13h ago
opinion
Dreamer joins Meta Superintelligence Labs
Meta execuhired the Dreamer team into MSL shortly after the podcast, giving the consumer agent startup a major distribution partner.
AINews•13h ago
vendor
Qwen3.5-9B Claude-4.6 GGUF conversion fixes released
The post repairs broken attention/expert layers across quantizations and shares LM Studio settings and merge steps.
AINews•13h ago
tutorial
The 5 levels of Claude Code
A guide lays out progression from raw prompting to multi-agent orchestration, noting when users hit ceilings at each level.
AINews•13h ago
tutorial
7MB binary-weight LLM runs in the browser without FPU
A 57M-parameter model with 99.9% binary weights runs in WASM at ~12 tok/s and works offline in-browser.
AINews•13h ago
research
Anthropic pushed computer use onto the desktop
Claude Cowork/Code adds macOS research preview control of mouse, keyboard, and screen, expanding agents beyond APIs and browsers.
AINews•13h ago
launch
Agent stack converging on long-running, tool-rich workflows
Tweets highlight momentum for Hermes Agent, T3 Code, Command Center, and Parchi as evidence of richer, parallel agent harnesses.
AINews•13h ago
opinion
Operational reality is now the bottleneck, not model IQ
Practitioners report over-agentic behavior and fragility in top models, urging tighter loops with traces, evals, and production feedback.
AINews•13h ago
opinion
Hyperagents / DGM-H advances self-improving agents
The work extends Darwin Gödel Machine ideas so agents can improve the improvement procedure itself, with cross-domain transfer claims.
AINews•13h ago
research
RLLM unifies RL post-training with LM-as-RM
RLLM trains a generative reward model on-policy to cover easy, hard, and non-verifiable tasks under one post-training approach.
AINews•13h ago
research
WebArena-Infinity scales browser benchmark generation
The project claims <10 hours and <$100 per environment while yielding harder browser tasks where open models score below 50%.
AINews•13h ago
benchmark
Turing Post catalogues 16 RL variants
A high-engagement overview lists RLHF, RLAIF, RLVR, process rewards, self-feedback, and critique-based methods as a taxonomy.
AINews•13h ago
tutorial
LeWorldModel shows stable JEPA world-model training
The model reports stable end-to-end JEPA training from pixels with 15M params and sub-second planning without heavy tricks.
AINews•13h ago
research
Mechanistic interpretability moves toward reverse engineering
A thread on Anthropic’s biology-of-LLM work highlights circuit-level mapping while noting models may not verbalize their own reasoning.
AINews•13h ago
research
Optimizer scaling theory for LLM hyperparameters
Antonio Orvieto argues adaptive-optimizer theory can explain scaling laws and reduce brute-force hyperparameter sweeps.
AINews•13h ago
research
LlamaParse + Gemini 3.1 Pro and LiteParse improve parsing
Google Devs and LlamaIndex show structured financial PDF extraction gains and introduce LiteParse for fast, low-cost parsing.
AINews•13h ago
launch
Cursor launches Instant Grep for repo-scale search
Instant Grep offers regex search over millions of files in milliseconds, directly improving agentic coding workflows.
AINews•13h ago
launch
Late-interaction multi-vector retrieval gains momentum
Weaviate/LightOn discussions argue late interaction is now practical and cheaper than cross-encoders for code-heavy retrieval.
AINews•13h ago
podcast
Sakana Chat launches with Namazu alpha models
Sakana released a Japanese consumer chat product backed by Namazu alpha models tuned for local context and reduced bias.
AINews•13h ago
launch
MiniMax introduces flat-rate Token Plan for multimodal APIs
The subscription bundles text, speech, music, video, and image APIs under a single predictable price.
AINews•13h ago
launch
Luma Uni-1 blends reasoning with image generation
Uni-1 is pitched as a model that thinks and generates pixels simultaneously for generative media workflows.
AINews•13h ago
launch
NVIDIA Kimodo ships promptable motion/timeline model
Kimodo is trained on 700 hours of mocap and supports both human and robot skeletons, with availability on Hugging Face.
AINews•13h ago
launch
Hugging Face Kernels 0.12.3 adds Flash-Attention 4
The release adds Flash-Attention 4 support via cutlass.cute kernels for faster attention workloads.
AINews•13h ago
launch
TRL v1.0.0 claims 44× VRAM savings for long sequences
The update highlights major memory savings and points to AsyncGRPO as the next optimization.
AINews•13h ago
launch
AI2 MolmoPoint GUI targets VLM-based UI automation
MolmoPoint uses grounding tokens instead of coordinate regression and reports 61.1 on ScreenSpotPro.
AINews•13h ago
launch
Sakana’s narrative intelligence OSINT workflow for 1.1M posts
The applied AI workflow combines LLM ensembles, novelty search, hypothesis generation, and human verification over massive social data.
AINews•13h ago
research
Current state of the Chinese LLMs scene
A recap maps the landscape across ByteDance, Alibaba, Tencent, Baidu, and other labs with rapid open-weight activity.
AINews•13h ago
opinion
OpenRouter usage list shows Chinese model dominance
Token-usage rankings highlight Chinese models at the top and note ByteDance still lacks open-weight releases.
AINews•13h ago
benchmark
Alibaba commits to open-sourcing new Qwen and Wan models
The company says it will keep releasing a full series of open models across sizes, fueling community anticipation.
AINews•13h ago
launch
Showing 56 of 56 stories from the last 3 days