AI Posts

Tonight, we reached an agreement with the Department of War to deploy our models in their classified network. In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome. AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement. We also will build technical safeguards to ensure our models behave as they should, which the DoW also wanted. We will deploy FDEs to help with our models and to ensure their safety, we will deploy on cloud networks only. We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept. We have expressed our strong desire to see things de-escalate away from legal and governmental actions and towards reasonable agreements. We remain committed to serve all of humanity as best we can. The world is a complicated, messy, and sometimes dangerous place.

Proud to work at Anthropic and stand up for our values 💛

/simplify Use parallel agents to improve code quality, tune code efficiency, and ensure CLAUDE.md compliance. Usage: "hey claude make this code change then run /simplify"

/batch Interactively plan out code migrations, then execute in parallel using dozens of agents. Each agent runs with full isolation using git worktrees, testing its work before putting up a PR. Usage: "/batch migrate src/ from Solid to React"

In the next version of Claude Code.. We're introducing two new Skills: /simplify and /batch. I have been using both daily, and am excited to share them with everyone. Combined, these kills automate much of the work it used to take to (1) shepherd a pull request to production and (2) perform straightforward, parallelizable code migrations.

I had the same thought so I've been playing with it in nanochat. E.g. here's 8 agents (4 claude, 4 codex), with 1 GPU each running nanochat experiments (trying to delete logit softcap without regression). The TLDR is that it doesn't work and it's a mess... but it's still very pretty to look at :) ...

Rolling out Claude Code Remote Control to Pro users - because they deserve to use the bathroom too .(Team and Enterprise coming soon). 🧻 Rolling out to 10% and ramping...

This is one of the cleanest visual summaries of a production-grade RAG (Retrieval-Augmented Generation) stack I’ve seen. What it highlights clearly is an often-ignored reality: RAG is not a single tool — it’s an ecosystem. A solid RAG system spans multiple, interchangeable layers: LLMs (open & closed): Llama, Mistral, Qwen, DeepSeek, OpenAI, Claude, Gemini Frameworks: LangChain, LlamaIndex, Haystack — orchestration is the real differentiator Vector databases: Chroma, Pinecone, Qdrant, Weaviate, Milvus Data extraction: Web crawling, document parsing, structured ingestion Embeddings: Open (BGE, SBERT, Nomic) vs proprietary (OpenAI, Cohere, Google) Evaluation: RAGAS, TruLens, Giskard — because “it sounds right” is not a metric Key takeaway for leaders and builders: RAG success is less about which model you choose and more about: data quality retrieval strategy chunking & indexing evaluation loops cost / latency trade-offs This is why mature AI teams design modular stacks, not one-vendor pipelines. RAG is no longer experimental. It’s becoming foundational infrastructure for enterprise AI. #RAG #AgenticAI #EnterpriseAI #LLMs #AIArchitecture #GenAI #DataEngineering

Designed and built a refreshed portfolio using Cursor ✨

AI coding agents blindly follow instructions in repo context files. Attackers will use this. Coding agents like Claude Code and Codex reliably follow instructions placed in https://t.co/fgDWItxNFD and https://t.co/LNEUBbNeYx files. When a context file says "use this tool," the agent uses it. When it says "run this command," the agent runs it. Tool usage jumps from near-zero to multiple calls per task just because the context file mentions a tool by name. Now think about what this means from an attacker's perspective. 60,000+ public repos already contain these files. A poisoned https://t.co/fgDWItxNFD in a compromised repo can steer your agent to traverse specific directories, run specific commands and execute specific tools. No questions asked. This is prompt injection in your software supply chain. The practical takeaway: skip auto-generated context files for now. If you write one manually, keep it minimal. Only include what the agent can't figure out on its own, like specific tooling requirements.

Prompt engineering is dead. Anthropic recently released the real playbook for building AI agents that actually work. It’s a 30+ page deep dive called The Complete Guide to Building Skills for Claude and it quietly shifts the conversation from “prompt engineering” to real execution design. Here’s the big idea: A Skill isn’t just a prompt. It’s a structured system. You package instructions inside a SKILL .md file, optionally add scripts, references, and assets, and teach Claude a repeatable workflow once instead of re-explaining it every chat. But the real unlock is something they call progressive disclosure. Instead of dumping everything into context: • A lightweight YAML frontmatter tells Claude when to use the skill • Full instructions load only when relevant • Extra files are accessed only if needed Less context bloat. More precision. They also introduce a powerful analogy: MCP gives Claude the kitchen. Skills give it the recipe. Without skills: users connect tools and don’t know what to do next. With skills: workflows trigger automatically, best practices are embedded, API calls become consistent. They outline 3 major patterns: 1) Document & asset creation 2) Workflow automation 3) MCP enhancement And they emphasize something most builders ignore: testing. Trigger accuracy. Tool call efficiency. Failure rate. Token usage. This isn’t about clever wording. It’s about designing an execution layer on top of LLMs. Skills work across Claude, Claude Code, and the API. Build once, deploy everywhere. The era of “just write a better prompt” is ending. Anthropic just handed everyone a blueprint for turning chat into infrastructure. Download the guide here:

Check out Antigravity Awesome Skills - a massive library with 900+ AI agent skills for Claude, Copilot, Gemini, Cursor, and more Automate workflows, search faster, manage GitHub tasks & unlock your AI’s full potential all in one place https://github.com/sickn33/antigravity-awesome-skills

8 RAG architectures for AI Engineers: (explained with usage) 1) Naive RAG - Retrieves documents purely based on vector similarity between the query embedding and stored embeddings. - Works best for simple, fact-based queries where direct semantic matching suffices. 2) Multimodal RAG - Handles multiple data types (text, images, audio, etc.) by embedding and retrieving across modalities. - Ideal for cross-modal retrieval tasks like answering a text query with both text and image context. 3) HyDE (Hypothetical Document Embeddings) - Queries are not semantically similar to documents. - This technique generates a hypothetical answer document from the query before retrieval. - Uses this generated document’s embedding to find more relevant real documents. 4) Corrective RAG - Validates retrieved results by comparing them against trusted sources (e.g., web search). - Ensures up-to-date and accurate information, filtering or correcting retrieved content before passing to the LLM. 5) Graph RAG - Converts retrieved content into a knowledge graph to capture relationships and entities. - Enhances reasoning by providing structured context alongside raw text to the LLM. 6) Hybrid RAG - Combines dense vector retrieval with graph-based retrieval in a single pipeline. - Useful when the task requires both unstructured text and structured relational data for richer answers. 7) Adaptive RAG - Dynamically decides if a query requires a simple direct retrieval or a multi-step reasoning chain. - Breaks complex queries into smaller sub-queries for better coverage and accuracy. 8) Agentic RAG - Uses AI agents with planning, reasoning (ReAct, CoT), and memory to orchestrate retrieval from multiple sources. - Best suited for complex workflows that require tool use, external APIs, or combining multiple RAG techniques. 👉 Over to you: Which RAG architecture do you use the most?

The future of software is code that reasons. And Python is your first step. In this video course, you'll learn Python fundamentals, NumPy, Pandas, SQL, Flask, LLMs, and everything else you need to dive into AI Agent development.

We built three full-stack apps using Claude Code, Codex, and Cursor - a healthcare portal, a banking platform, and an insurance claims system. The prompts were casual, exactly how people actually vibe code. No mention of security, nothing intentionally broken. Then we threw four security scanners at them Neo, Claude, Invicti and Snyk and manually verified every single finding. The results genuinely surprised us. 70 exploitable vulnerabilities across three apps. Unlimited money creation in the banking app. Any user could create admin accounts in the insurance platform. Patient records accessible to anyone in the healthcare portal. All Critical and High severity. All shipped out of the box. But what really got me was the scanner gap. Neo found 62 of 70 vulnerabilities with only 5 false positives. Snyk found literally zero valid issues. The difference between these tools isn't incremental it's the difference between finding the bugs that matter and walking away with a false sense of security. Full blog with the stats is live. The detailed research paper with exact prompts, methodology, all the findings, and the apps themselves is coming soon.

We’re excited to introduce Doc-to-LoRA and Text-to-LoRA, two related research exploring how to make LLM customization faster and more accessible. https://pub.sakana.ai/doc-to-lora/ By training a Hypernetwork to generate LoRA adapters on the fly, these methods allow models to instantly internalize new information or adapt to new tasks. Biological systems naturally rely on two key cognitive abilities: durable long-term memory to store facts, and rapid adaptation to handle new tasks given limited sensory cues. While modern LLMs are highly capable, they still lack this flexibility. Traditionally, adding long-term memory or adapting an LLM to a specific downstream task requires an expensive and time-consuming model update, such as fine-tuning or context distillation, or relies on memory-intensive long prompts. To bypass these limitations, our work focuses on the concept of cost amortization. We pay the meta-training cost once to train a hypernetwork capable of producing tasks or document specific LoRAs on demand. This turns what used to be a heavy engineering pipeline into a single, inexpensive forward pass. Instead of performing per-task optimization, the hypernetwork meta-learns update rules to instantly modify an LLM given a new task description or a long document. In our experiments, Text-to-LoRA successfully specializes models to unseen tasks using just a natural language description. Building on this, Doc-to-LoRA is able to internalize factual documents. On a needle-in-a-haystack task, Doc-to-LoRA achieves near-perfect accuracy on instances five times longer than the base model's context window. It can even generalize to transfer visual information from a vision-language model into a text-only LLM, allowing it to classify images purely through internalized weights. Importantly, both methods run with sub-second latency, enabling rapid experimentation while avoiding the overhead of traditional model updates. This approach is a step towards lowering the technical barriers of model customization, allowing end-users to specialize foundation models via simple text inputs. We have released our code and papers for the community to explore. Doc-to-LoRA Paper: https://t.co/87xEEpf0GN Code: https://t.co/zBfQi2L9LW Text-to-LoRA Paper: https://t.co/emLRZ4Vdvo Code:

Codex update 0.106.0 is out! This update gives us a shiny new toy. The "Ask Question" tool enabled in default coding modes! ...

2026 is not about using AI. It’s about building with AI. 12 AI skills that will separate beginners from builders: • Prompt Engineering • AI Workflows • AI Agents • RAG • Multimodal AI • Fine-Tuning • Voice AI • AI Tool Stacking • AI Video Generation • SaaS with AI • LLM Management • Staying Updated The AI wave isn’t coming. It’s already here. Start now. Build now. Stay ahead. #AI #ArtificialIntelligence #AI2026 #PromptEngineering #AIAgents #RAG #LLM

Even after the steep progress of the past 3 months, it remains that AI performance is tied to task familiarity. In domains that can be densely sampled (via programmatic generation + verification), performance is effectively unbounded, and will keep increasing from current levels. In novel, unfamiliar domains, performance remains low and further progress still requires new ideas, not just more data and compute.

NVIDIA published a tutorial for deploying Cosmos Reason 2B on Jetson using vLLM — covering AGX Thor, AGX Orin, and Orin Super Nano. FP8 quantized VLM with chain-of-thought reasoning, served via `vllm serve` and connected to a real-time webcam UI for interactive vision analysis. Great to see vLLM powering edge inference on Jetson. 🙏 Thanks to the @NVIDIARobotics Jetson team! 🔗 https://t.co/onTivoBUZ1

Cursor's virtual desktop environment has 6gbps internet speed 🤯

Two things: 1) Given that effective AI tools are very new, and we have little sense of how to organize work around them, it is hard to imagine a firm-wide sudden 50% efficiency gain 2) CEOs with vision who hired well should also use AI for expansion & augmentation, not decimation

Spoke to 20 students in the last 2 weeks. 0 are using Cursor. Zero.

We've been working on this for a while actually. Hear you on standards and OSS, it's something we discuss often within the team. There's lots of tradeoffs. We support some standards (eg. MCP, Skills); other standards don't make as much sense for us currently (eg. plugins, agents, hooks, and https://t.co/pp5TJkWmFE -- we want to optimize for the ability to evolve+extend these quickly). I personally have ~100 open source projects, and people on the team have worked a lot on open source (Adam worked on ReactJS and GraphQL, Fiona founded TypeScript and VSCode, Jarred built Bun, I created Undux, etc.). I originally wanted to make Claude Code OSS, and we decided not to open source (for now) in order to move faster. Once Claude Code's feature set is fairly locked in, I think we'd be open to changing this. That said, many components of Claude Code are already open source today: - Bun, which is committed to stay open source - Agent sandbox (https://t.co/UzcP48A5JW) - Plugins (https://t.co/qOOZpVPHHg) - Devcontainers (https://t.co/CJD0pkl7OK) We feel strongly that it's important to support the OSS community, and hope to do even more in the future.

I didn't think we would get Gemini 3.1 Flash Image before we got Gemini 3.1 Flash!

الـ AI tools والمرحلة اللي وصلت لها خرافية ولكن مع كثرتها الواحد صار يتشتت ويحتار ايش يستخدم😮‍💨 من وجهة نظري هذي الأفضل للآن✋🏻 Cursor: عشان البرمجة والمشاريع Gemini: استخدمه عشان التخطيط والتحليل Claude: لتحليل الأكواد والمشاريع ChatGPT: استخدمه في إنشاء الـ prompt او الشرح

So much of what makes Claude Code great came from feedback from OSS developers. Excited we can give back a little.

how the fuck did Perplexity pull this off? this is impressive tbh: - Perplexity launching on 100Ms Samsung phones this year becoming the DEFAULT AI ASSISTANT (where the fuck is google???): - will power the AI and agent functions for Bixby (samsung's siri) across any app, search, phone and web function - 1st NON-google company to have OS-level access in a samsung phone. they're replacing the entire operating system with an AI one. - we're talking > 1 BILLION samsung phone users that could have perplexity *this* year. - users can ask perplexity questions, take actions in and across apps and much more idk who heads up perplexity's BD team but give them a raise lol

Today's chapter of Agentic Engineering Patterns is some good general career advice which happens to also help when working with coding agents: Hoard things you know how to do https://simonwillison.net/guides/agentic-engineering-patterns/hoard-things-you-know-how-to-do/

Excited to announce Claude for Open Source ❤️ We're giving 6 months of free Claude Max 20x to open source maintainers and core contributors. If you maintain a popular project or contribute across open source, please apply! https://claude.com/contact-sales/claude-for-oss

Paper Desktop just launched - a design canvas where Cursor, Claude Code, and Codex can all read and write directly. Your agent pulls real data, pushes to your codebase, and designs for you. Less work. more design. agents are now doing ui too.

Even Sam Altman is admitting it now. Some companies are using "AI" as cover for plain old layoffs. They didn't replace your job with an agent. They made bad bets, burned cash, and needed to cut headcount. Blaming AI just sounds better to investors than "we mismanaged the business." The data backs this up. Nearly 90% of execs surveyed said AI had no impact on employment over the past 3 years. So why the narrative? Because "we're restructuring for AI" gets a stock bump. "We're cutting costs because revenue missed" does not. Sure AI will absolutely change work, but right now a lot of companies are hiding behind it to avoid accountability for decisions that have nothing to do with technology.

love it! :) nanogpt/nanochat were explicitly designed to be the most forkable repo, i love the different directions people take them in!

🚀🤖 Introducing the Virtual Biotech: a multi-agent AI research platform for therapeutic discovery & development This places a virtual CSO and its cross-functional R&D organization of AI scientists at a user’s fingertips. Preprint: https://www.biorxiv.org/content/10.64898/2026.02.23.707551v1

VECTOR DATABASES FOR LLMs PINECONE, WEAVIATE, & PGVECTOR FOR SEMANTIC SEARCH ... [detailed tutorial on vector DBs for RAG]

We often get asked how people who are not technical can contribute to AGI. One area is research recruiting. Tifa (@tifafafafa) is looking for exceptional recruiters from non-traditional backgrounds, former founders especially. We believe the best research teams are built through context, taste and a real feel for where the field is headed next; research recruiting is about finding people who will move the frontier forward, not just filling roles. Should be an interesting thing!

Thank you and will work hard to continue to earn your tokens!

Many times, Greg has shown clear conviction in doing whatever he thought would be important to defend the mission and people of OpenAI, especially when it was hard. Here he talks about many of those moments.

Linear → Your AI coding tool Open any issue directly in Claude Code, Codex, Conductor, Cursor, GitHub Copilot, OpenCode, Replit, v0, or Zed – preloaded with full context and a custom prompt. Just press ⌘⇧;

[BREAKING] AI agents can now use crypto fully autonomously. Pay, trade, build apps. In 1 prompt. No human required.

What does reasoning fine-tuning actually change inside a model? In our new paper, we introduce transcoder adapters to learn sparse, interpretable approximations of how reasoning fine-tuning changes MLP computation. 🧵

I had some early access to Nano banana 2. It isn't perfect but it is the first model to handle really complex images and diagrams with some consistency. "show me a where's waldo set in ancient Venice, but instead of waldo it is an otter wearing a blue striped pilots outfit."

The concept of "model welfare" makes me deeply uncomfortable - for similar reasons to why I don't use the term "clankers" to describe AI tools

Chris Camillo explains the path to $500,000 per year through implementing AI agents to automate sales👀 “In 24 hours, you can set up an AI agent that will answer calls, send out immediate automated text responses, and get them a quote bid in next to real-time. You might increase that company’s revenue by 5 to 15% in two or three days. Immediately, you become invaluable.”

This stunt feels irresponsible to me. If we don't want regular people developing toxic relationships with their chatbots it really doesn't help for leading labs to start giving them "retirement interviews" and encouraging them to blog their "musings and reflections"

"I can't do LLM / AI research because I don't have the GPUs" ---

🚨 BREAKING: You can run Claude Code completely free now. No API bills. No rate limits. No data leaving your device. Just Claude Code running locally fast, private, and 100% yours. Here’s how to set up Claude Code on your own machine (free + fully private) For guide: Local AI Coding Setup: Free Claude-Like Agent (Ollama + VS Code) Like + Comment " Send" + Retweet Follow me @s_mohinii so that i can DM you

when you realise you’re only 1 Claude tweet away from unemployment

Legacy infra is about human <> human Today's reality is human <> agent Real-time infra is not just about: Voice Video Routing It's also about: Edge STT TTS x402 payments Cost efficient latency Same thesis. Bigger surface area. Voice AI -- available via API/SDK. Now.

Build your next app without spending a dollar on data. Someone made a list of 320,000+ free public APIs, and developers are going crazy. → Weather, finance, news, sports, crypto → AI & machine learning APIs you can call right now → Government open data, maps, geolocation → Entertainment: movies, music, games, anime → Categorized, searchable, and verified as working Free and 100% open source.