feed
last scrape: 3h ago · 961 in archive · 9 today
this week
YC Founder Firesides chat with Conductor co-founders on trying 12 ideas before building their Mac app for parallel coding agents, 10x growth, $22M raise, and Cloud launch.
Conductor Founders on Pioneering Multi-Agent Coding
Tutorial on Vercel DeepSec CLI workflow with Claude Code: deepsec init (info.md prompt), scan (regex filter), process (batched agent review), report (JSON/MD), optional revalidation/export for codebase security scans. Found 3/10 OWASP issues in tests, misses runtime vulns.
DeepSec CLI Secures AI-Generated Codebases
Speaker argues Mozilla's use of Anthropic's Mythos—which found 271 vulnerabilities in Firefox—erodes trust in human-written code as the security baseline, urging a "golden refactor window" to prioritize code comprehensibility. Full writeup here.
Mythos' 271 Firefox Vulns Flip Human Code Trust
Demo of Talkie, a 13B model trained solely on pre-1931 texts (newspapers, patents, books) to test AI reasoning vs. memorization without modern web contamination — shows predictions, slang, basic Python learning, and training hurdles like OCR and temporal leakage.
Talkie: 13B LLM with Pre-1931 Knowledge Only
Creator's subjective tier list of ~30 AI tools: S-tier daily drivers (Claude Code in VS Code, Glido), A-tier weekly (Codex, Perplexity, Grok), down to C-tier experiments and "graduated" ones he's dropped like ChatGPT and Cursor, plus a framework for sticking to a lean stack.
Nate Herk's Lean AI Tool Tier List
Advocates running Claude Code's terminal inside the Codex desktop app for a dual-tool workflow, with a walkthrough of Codex's UI/settings/plugins and pitches for the creator's Skool courses.
Run Claude Code in Codex for Dual AI Coding Power
Podcast recaps Anthropic's DevDay agent features—Dreaming for scheduled memory review and pattern extraction, Outcomes for rubric-based output grading (+8-10% quality gains), and multi-agent orchestration— overshadowed by a brief SpaceX partnership mention.
Anthropic's Agent Harness Leap and SpaceX Compute Lifeline
Live stream recaps Sam Altman firing memes ("directionally bad" texts with Mira Murati), teases OpenAI's GPT Realtime 2, then chats with Oracle's Richmond De on Anthropic's Claude "Dreaming" for agent memory consolidation—like MEMGPT's sleep-time compute from a year ago.
Agent Memory: Dreaming, Feedback, and Continual Learning Path
Host recaps Anthropic's SpaceX deal for full Colossus 1 capacity (220k+ GPUs), doubling Claude Code quotas and hiking Opus API limits up to 10M input tokens/min for top tiers.
Anthropic Leases xAI's Colossus 1 to End Compute Crunch
Step-by-step guide to building an Obsidian vault with five folders (audience, community/Buildroom, coaching transcripts, content/hooks, AI/SOPs), index files linking your offer/ICP/brand voice/proof bank, and plugins/skills to feed it into Claude for personalized social posts. Sells a template via the Buildroom community.
Obsidian Vault Indexes Brand Data for Claude Content Generation
Demo of HubSpot's free AEO tool: input brand/competitors/products/ICPs, auto-generate persona-specific prompts, track daily LLM visibility/share-of-voice/sentiment vs rivals, get optimization recs.
HubSpot's Free AEO Tool Tracks AI Visibility
Commentary on OpenClaw's April updates maturing it into an agent runtime with task flows, retries, and model-swapping via durable workflows—memory via OpenBrain is the real strategic shift, not just escaping provider lock-in.
OpenClaw Matures into Model-Swappable Agent Runtime
Live demo of Gemini File Search on the "Attention Is All You Need" paper: upload to a File Search Store (auto-chunks/embeds text+images via Embedding 2), then query diagrams directly—4 API calls replace the full multimodal RAG stack. Simplifies engineering, doesn't kill it.
Gemini File Search 2.0 Simplifies Multimodal RAG to API Calls
Overview of IBM's Granite Speech 4.1 2B ASR models: base leads Open ASR Leaderboard (WER 5.33, RTFx 231x), Plus adds diarization/word timestamps (5 langs), NAR uses NLE transcript editing for batched RTFx 1820x on H100. Local code demo.
IBM Granite Speech 4.1: 2B ASR models for speed vs features
Creator reacts to Anthropic's higher limits announcement leasing Nvidia compute from SpaceX's Colossus to ease Claude shortages, despite past xAI bans and Elon's jabs.
Anthropic Buys xAI Compute to Fix Claude Crunch
Anthropic doubled Claude Code's 5-hour session limits, removed peak-hour throttling, and boosted API rates (e.g., output from 8k to 80k tokens/min) via SpaceX compute deal—speaker recaps changes and lists five builder tweaks.
Anthropic Doubles Claude Code Limits via SpaceX Deal
Walks through DeepSeek V4's paper, explaining DSA (sparse attention via lightning indexer), CSA (4x compressed sparse), and HCA (128x compressed attention) interleaved to slash KV cache to 10% and flops to 27% vs. prior models.
DeepSeek V4 Slashes Inference via DSA, CSA, HCA
Walkthrough of official n8n MCP server setup as Claude remote connector vs Czlonkowski's unofficial: official rebuilds full workflows on updates (token-heavy, no executions list/delete) while unofficial enables partial edits; they complement for building/iteration.
n8n Official MCP vs Czlonkowski Unofficial
Daily AI news podcast skeptical of Coinbase blaming layoffs on AI (crypto market tanked), recaps Anthropic's $200B Google Cloud deal, Palantir's 85% revenue surge, and BlackRock's CEO calling compute a futures-tradable commodity.
AI Shifts Hard to Enterprise, Dumps Consumer Hype
Meng To demos Google's Design.md as a prompt-attached blueprint for typography/colors/spacing/effects to keep AI outputs consistent across landing pages, slides, motion, and mobile—stacked with "skills" (skeuomorphic/3D/lasers) and live-builds one using HTML refs in Aura.
Design.md: Portable DNA for AI Design Consistency
Live reaction to Anthropic renting xAI's full Colossus 1 data center (300MW, 220k GPUs) to double Claude Code's 5-hour rate limits, drop peak-hour cuts, and raise Opus API tokens/min from 30k to 500k (Tier 1).
Anthropic Secures SpaceX Colossus for Claude Compute Boost
Reaction to an OpenAI insider's tweet calling Anthropic a "Claude-worshipping cult" that might let the model shape its own team and ethics—contrasted with OpenAI's tool-first view. Mostly vibes and quotes, no new info.
Anthropic's AI Worship Scares Berman vs OpenAI Tools
Intro to Codex desktop app: AI accesses your local files/folders directly (vs. uploading to ChatGPT), with setup via three questions—work location (chat/folder), reasoning (light/extra high), permissions (default/auto/full)—plus ChatGPT feature mappings and five use cases. Prompts in slides.
Codex: AI Accesses Your Files Directly, Beats ChatGPT Context Limits
Interview with Every's head of growth on switching from Claude Code's CLI to OpenAI's Codex desktop app for 80% of his knowledge work like transcript-to-GTM plans and Notion KPI dashboards.
Codex Replaces Claude Code as Knowledge Work OS
Argues AI agents win via semantic "work primitives" (meaning layer) over raw computer access, via three layers—access, meaning, authority—with examples from coding agents, Perplexity's evolution, and Salesforce vs. SAP. Full post w/ prompt kit: here.
Semantic Work Primitives Over Computer Access