Latent Signal

RIP Perplexity Pro. This open-source agent does deeper research, audits papers against code, replicates experiments on real GPUs, and outputs source-grounded briefs with live citation URLs. For free. It's called Feynman. Here's what makes it different: Most "AI research" tools summarize web pages. Feynman runs four agents in parallel: one gathers evidence across papers and repos, one simulates peer review, one drafts from findings, one verifies every citation and kills dead links. The wildest command: `feynman replicate "chain-of-thought improves math"` and it actually runs the experiment on your local or cloud GPU. Not a summary. The actual experiment. One install: curl -fsSL https:// feynman .is/install | bash 100% Open Source. MIT License. Repo: github.com/getcompanion-a…

@heyrimsha Very good, this, with Hermes and Claude Code. Going to be able to get through so much more in my PhD

> Claude usage limit reached. > every. fucking. day. > was about to upgrade to Max > then find a 16 minute video > two engineers built Claude Skills from scratch > first 5 minutes > wait. I've been loading the same context 100 times? > Skills remember your workflow automatically? > spent months blaming the token limit > turns out I was the problem > 16 minutes later > basic plan handles everything > been the skill issue this whole time

Last week, Anthropic announced Project Glasswing alongside Claude Mythos Preview, a model they described as so powerful at finding vulnerabilities they couldn't release it. The announcement featured AWS, Microsoft, Google, and Apple as partners, $100M in compute credits, and a clear message: this is dangerous, and only we can be trusted to deploy it safely. The results were real. Thousands of zero-days across every major OS and browser. A 27-year-old bug in OpenBSD. A 16-year-old bug in FFmpeg. Fully autonomous exploit chains that would have taken human researchers weeks. But here's what bothered me: all the credit went to the model. Read the technical blog carefully and a different picture emerges. The real innovation isn't the model. It's the workflow: - Rank every file in a codebase by attack surface - Fan out hundreds of parallel agents, each scoped to one file - Use crash oracles (AddressSanitizer, UBSan) as ground truth - Run a second verification agent to filter noise - Generate exploits as a triage mechanism for severity That's a pipeline. And pipelines are model-agnostic. At Lazarus AI, we spend our days deploying custom AI in places where "just use the closed API" isn't an option: regulated industries, enterprise, and government. When I saw Glasswing, my instinct was the same one I have every week: strip out the proprietary model, keep the architecture, run it on whatever model is best for the customer. Clearwing is a fully open-source vulnerability discovery engine. Crash-first hunting, file-parallel agents, oracle-driven verification, variant hunting, adversarial verification. Works with any LLM. I tested it with OpenAI Codex 5.4 and reproduced Glasswing's findings. I'm now reproducing results with our own ReAligned model - Qwen3.5 finetuned to Western alignment. Mythos is certainly a great model. The N-day exploit walkthroughs in Anthropic's blog show real reasoning depth. But it's an incremental improvement over Opus, the same way Opus was over Sonnet, and Sonnet over Haiku. It's not a leap to superintelligence. It's the next point on a curve we've been watching for years. What actually changed the game was the workflow. Defenders shouldn't have to wait for access to a gated model to secure their software. These vulnerabilities have been sitting in codebases for decades. The tools to find them should be available to everyone: the open source maintainer running FFmpeg on a Saturday, the startup that can't afford $125/M output tokens, the researcher in a country where Anthropic doesn't operate. Clearwing is MIT licensed and available now. github.com/Lazarus-AI/cle… Clearwing enables a wide variety of security activities. Handle with care. It is sharp.

In 14 minutes, this Anthropic engineer who wrote "Building Effective Agents" will teach you more about building them right than most developers figure out on their own in months. Bookmark this for the weekend. Then read the builder's guide below.

⚡ Meet Qwen3.6-35B-A3B：Now Open-Source！🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes Efficient. Powerful. Versatile. Try it now👇 Blog：qwen.ai/blog?id=qwen3.… Qwen Studio：chat.qwen.ai HuggingFace：huggingface.co/Qwen/Qwen3.6-3… ModelScope：modelscope.cn/models/Qwen/Qw… API（‘Qwen3.6-Flash’ on Model Studio）：Coming soon～ Stay tuned

@AnthropicAI recently announced their Glasswing project, powered by their unreleased Mythos model, which uncovered zero-day vulnerabilities in several projects, including FFmpeg. They used it as evidence that Mythos is a super scary and dangerous superintelligence that should be kept out of your hands. I doubt that Mythos is actually a vastly smarter model. It'll be incrementally better as Opus is to Sonnet. I predict that their FFmpeg result is reproducible with a much smaller model than Mythos. (Qwen3.5-397b, maybe) To test that idea, I created Clearwing, an open-source implementation of Glasswing. You can point it at any model (@ollama , @lmstudio , @OpenRouter, @huggingface, etc)

@Teknium Another greate tool. I seem to be converging on using Hermes as my r&d partner and claude code as my code monkey

This skill is now built in to Hermes! Use /architecture-diagram <prompt> after updating hermes, and you're good to go! Thanks to the author of the skill making it MIT we were able to port it over directly into Hermes Agent as a built in skill!

Using /skin sisyphus is a nice easter egg in Hermes

going to test this today in the lab

MiniMax 2.7 can now be run locally!🔥 MiniMax-M2.7 is a new 230B parameter open model with SOTA on SWE-Pro and Terminal Bench 2. Run the Dynamic 4-bit MoE model on 128GB Mac or RAM/VRAM setups. Guide: unsloth.ai/docs/models/mi… GGUF: huggingface.co/unsloth/MiniMa…

Every time you accepted a salary, chose a price, or walked into a negotiation, the other person was running GAME THEORY in their head. You were guessing. This 1-hour Yale lecture by Professor Ben Polak will permanently change how you read people and make decisions. Most MBAs pay $150k to learn this. Yale posted it for free:

introducing Autoreason, a reasoning method inspired by @karpathy's AutoResearch which extends the strategy for subjective domains the paper was co-written with Hermes Agent by @NousResearch, using a research-paper-writing skill developed while writing it paper + results below

Minara's new home: Hermes Agents✨ Positions, trades, autopliot ... everything you know, running and functioning the same way. Install Minara Skill with one command in your hermes agent: curl -fsSL raw.githubusercontent.com/Minara-AI/skil… | bash

What's this 👀

As someone who scraped for a living for years, anyone recommending lightpanda to do it shows that they don’t have any experience regarding the subject. Only one thing to understand: TLS Fingerprinting You can have the fastest headless setup, puppeteer, lightpanda,… one wrong ClientHello and Cloudflare/Akamai lights you up instantly. CAPTCHA city. Lightpanda/Zig stuff is fun for tiny sites but gets cooked the second real anti-bot shows up. Cloudflare? Protects 20%+ of all websites on the internet What is a ClientHello? It’s the very first message your browser (or bot) sends during the TLS handshake. It openly announces your TLS version, the list of supported cipher suites, elliptic curves, extensions order, GREASE values, and other data. Anti-bot systems like Cloudflare and Akamai read this instantly and turn it into a fingerprint. If it doesn’t match a real browser’s exact signature… you’re flagged as a bot right away. The key here is simple: real TLS fingerprint spoofing requires low-level control. You can’t do it properly in JS or Python. You need languages like C++ or Rust to actually rewrite the ClientHello, cipher suites, extensions, and all the tiny details that Cloudflare and Akamai check instantly. Anything higher-level just leaves obvious artifacts that scream ‘bot’ What I recommend: Camofox An actual Firefox fork with proper C++ fingerprint spoofing, native TLS behavior, proxy/geo baked in, built so your agents don’t die on protected pages. Top-tier protections might flag it following interaction speed on the pages, ip addresses and other factors but there’s NO match between lightpanda and this "Camofox patches Firefox at the C++ implementation level - navigator.hardwareConcurrency, WebGL renderers, AudioContext, screen geometry, WebRTC are all spoofed" Basically, everything is spoofed BEFORE the JS on the page can even see the values. Which is not possible with python/js libraries. On another note, I talked about it to @Teknium on @NousResearch discord and literally 2 hours later it was implemented in Hermes Agent, it just shows that they take feedback very seriously and want to give the smoothest agent experience they can Level-up your setup right now github.com/jo-inc/camofox…

We created the open source version of Claude Managed Agents. Introducing Multica github.com/multica-ai/mul…

@Teknium @j0hngou Thanks! Ill implement this for my research. CC for some stuff, Hermes with qwen 3.5 self hosted for others.

From hermes :) **For a PhD student specifically, the differentiators that matter:** ◆ **Research paper pipeline** — We have a `research-paper-writing` skill that covers end-to-end ML/AI paper writing: experiment design, statistical analysis, drafting, revision cycles, and submission formatting for NeurIPS/ICML/ICLR/ACL/AAAI/COLM. Claude Code can edit LaTeX; Hermes can help you write the paper. ◆ **arXiv integration** — Built-in `arxiv` skill searches and retrieves papers via the API. Combine with `ocr-and-documents` to ingest full PDFs, or `youtube-content` to transcribe conference talks. You can set up a **cron job** that monitors arXiv daily for papers in your research area and sends you a Telegram summary every morning. CC can't do scheduled tasks at all. ◆ **LLM Wiki** — Karpathy's wiki skill builds a persistent, interlinked markdown knowledge base. Feed it papers, lectures, notes — it compiles them into a queryable knowledge graph. Great for literature reviews and qualifying exam prep. ◆ **Persistent memory + session search** — Hermes remembers across sessions. Tell it about your research topic, your advisor's preferences, your lab's conventions, your paper deadlines. Next week when you say "draft the related work section," it already knows your context. CC starts fresh every time. ◆ **Jupyter live kernel** — `jupyter-live-kernel` skill gives stateful, iterative Python execution. Data exploration, plotting, ML experimentation with intermediate results — the actual data science workflow, not just writing scripts. ◆ **Gateway (Telegram/Discord/Slack)** — Message your agent from your phone. "Hey, what was that paper we discussed about attention mechanisms?" while you're at a conference. Or have it summarize your experiment results while you're away from your desk. CC is terminal-only. ◆ **Cron scheduling** — Automated recurring tasks: daily arXiv digest, weekly experiment status reports, funding opportunity alerts (we have an `ai-funding-daily-report` skill), Polymarket tracking for prediction markets. Set it and forget it. ◆ **ML tooling** — `huggingface-hub` for model/dataset management, `grpo-rl-training` for RL fine-tuning guidance, `gguf-quantization` for running models on consumer hardware, `manim-video` for 3Blue1Brown-style math animations (great for presentations and explainer content). ◆ **Productivity stack** — `google-workspace` (Gmail, Calendar, Drive, Sheets), `obsidian` for note-taking, `powerpoint` for presentations, `himalaya` for email, `excalidraw` for diagrams. One agent handles your entire workflow. ◆ **Self-hosted with any model** — He wants to run Qwen 3.5 on his own server. Hermes supports that natively. Point it at your local inference endpoint and go. CC is locked to Anthropic. ◆ **MCP + extensibility** — Native MCP client means he can connect to any MCP server (database tools, custom lab APIs, institutional services). Plus 100+ skills, a plugin system, and the ability to delegate sub-tasks to Claude Code or Codex as child agents when he does need pure coding power.

I really want to try hermes agent with qwen3.5 on my server but I'm struggling to find a usecase. I am already happy with CC for coding. Karpathy's wiki sounds like a cool idea. What else would be nice for a PhD student? @Teknium

anthropics growth marketer mapped out 4 levels of AI marketing use most people sit at level 1, automating what they already do > level 1: automate what you already do (reporting, copy, data pulls) > level 2: use AI as a thinking partner where its better than you > level 3: do work that was below the ROI threshold before > level 4: build custom tools only you would ever build level 3 is work that never existed before. stuff nobody did because the manual cost was never worth it mining negative keywords across every ad group. checking your full site for broken links daily same logic applies to content, research, QA, competitor monitoring. all work that existed in theory but nobody had the hours for level 4 is where the ROI compounds there are hundreds of AI marketing skills and plugins floating around github right now. most of them work in theory but fall apart in practice because they are built for the general case, not your case your business has specific data, specific workflows, specific edge cases that no generic tool will ever cover. the people building custom tools around their own problems are the ones pulling ahead