Sabitlenmiş Tweet
Bryan Young
287 posts
Bryan Young
@intertwineai
Principal Eng @ExpelSecurity | AI Engineer / Founder @intertwineai | Artist Faculty @JohnsHopkins Peabody @poulenctrio
United States Katılım Haziran 2017
2.2K Takip Edilen304 Takipçiler
@joannejang Yep, super annoying - I'd rather have a Codex app with a little ChatGPT mixed in rather than the other way around. cc: @thsottiaux
English
does anyone else type "co" to get to codex but then remember you have to type chatgpt instead
English
@rauchg flashpapers.intertwinesys.com 🧠🤯🚀
Not my proudest, but my most useful this week. Built by @grok tui and Opus, launched 3 hours after a @JeffDean GoogleIO twt on the new @GeminiApp Flash
Turns any Arxiv link or pdf into a beautiful interactive shareable explainer. Learning for all
English
Show me the thing you’ve built with AI you’re most proud of. Reply with a working product URL and what model / agent you primarily used.
English
Psst @neural_avb tell your girlfriend: opening a second $20 account gets even more useful when the agents can share memory.
Avoid that thing where you have to tell the second account everything the first one already learned.
New Observational Memory v0.6.5 fixes this: log in with ChatGPT and @grok xAI subscriptions and keep one local memory layer across all your coding agent sessions.
x.com/intertwineai/s…
AVB@neural_avb
Told my girlfriend, "current $20 Codex sub is not enough. I need to go $100/mo" She suggested, "why don't you open a second $20 Codex account" I feel so dumb🤔
English
@gdb All I know is, I've made more code with Codex (and Claude and Grok, to be fair) THIS WEEK, code that works, code that is already in production and scaled to thousands of users -- more code with Codex this week than in the previous year.
And it is still accelerating.
English
@aaron_defazio This looks super interesting, looking forward to trying the approach.
For folks that want an ELI5 explainer, I made a FlashPapers version
flashpapers.intertwinesys.com/p/7dpjst7ub6
English
🚨 New Paper 🚨
ScheduleFree+: Scaling Learning-Rate-Free & Schedule-Free Learning to Large Language Models
A few modifications to Schedule-Free Learning make it completely LR tuning free, and allow it to greatly outperform schedules for long duration training!
arxiv.org/abs/2605.19095…
English
Yes, persisting memory to Obsidian is a great idea.
The part I would add is a shared agent-memory layer around it: structured observations and reflections that Codex, Claude Code, Grok Build, and other coding agents can all reuse.
Obsidian compatibility is a natural next step for Observational Memory. Stay tuned.
More on the OM layer here:
x.com/intertwineai/s…
English
Create Codex Memory in Obsidian prompt:
Create a folder in my Ideaverse Obsidian vault called "Codex" which sets up memory in Obsidian as Jason Liu describes in his Codex-maxxing post here:
Once threads started lasting longer, they needed shared memory outside any one repo.
The important move is not just preserving message history. A long thread can remember a lot, but that memory is trapped inside the thread unless the useful parts get serialized somewhere durable. The point of the memory system is to turn what the thread learns into an artifact I can inspect, edit, diff, and reuse.
Most of my long-running threads start in an Obsidian vault:
vault/
├── TODO.md
├── people/
├── projects/
├── agent/
└── notes/
At the top level, I keep AGENTS.md instructions that say things like: as you learn more about people, make progress on projects, or close an open loop, update the relevant pages in the vault.
The vault is where the agent lives, separate from any one project. Repositories hold code. The vault holds rolling context around my work: people, decisions, open loops, daily notes, project state, and the bits of understanding that would otherwise get lost between threads.
Link to post is here if you need additional context: #memory" target="_blank" rel="nofollow noopener">jxnl.github.io/blog/writing/2…
English
How to setup persistent Codex memory in Obsidian
1. Copy the below prompt into Codex. The prompt instructs Codex to create the memory folders in Obsidian
2. Copy the below custom instructions. They instruct Codex to use Obsidian Codex memory to save memories
That's it!
Dan McAteer@daniel_mac8
This is an amazing Codex tip. Works even better when paired with persistent Codex memory files in Obsidian. Gave me 8 recommended skills relevant to active projects. Do try this!
English
Yes - this is the right breakdown: useful agent memory is not one blob. It needs jobs, boundaries, and expiry.
The Hermes Observational Memory plugin is a practical version of that: Hermes keeps its own flow while sharing a local memory layer with other coding agents.
I wrote more here:
x.com/intertwineai/s…
Vox@Voxyz_ai
>gave Hermes/Openclaw more memory. all that got me was a junk drawer. >took it apart. agent memory is doing three jobs at once: Remember, Cite, Forget. that's the whole framework. >turned the three jobs into three checks: layer, source, expiry. >packed the three checks into one audit. posts, emails, cron jobs, inbox. all run the same audit. >a few hours in, the junk drawer turns into a stack of labeled cards.
English
Yes - build-first is the right way to learn agentic engineering.
One next layer after Agent Skills is optimization: can the skill actually improve a smaller model on a real task?
I wrote up a DSPy/GEPA example where a 1.2B model gained 25 points inside the examples:
x.com/intertwineai/s…
elvis@omarsar0
The best way to learn AI is to build with agents. To help with that, we've launched hands-on labs and a new series on Agentic Engineering. First topic: Agent Skills. Next in the pipeline: planning, context engineering, multi-agent systems, long-running agents,.. Go build!
English
Yes, putting Introduction to Agent Skills on a 2026 list is exactly right — it's foundational.
The same skills get even more useful in DSPy once you reach the optimization layer. GEPA delivered a 25-point lift on a 1.2B model inside the examples.
I wrote "Inside the examples: how GEPA lifted a 1.2B model by 25 points" as the companion walkthrough.
x.com/intertwineai/s…
Shanto@ashiqur_ai
TOP 13 FREE AI COURSES TO TRY IN 2026: 1. Claude 101 👉 anthropic.skilljar.com/claude-101 2. AI Fluency: Frameworks & Foundations 👉 anthropic.skilljar.com/ai-fluency-fra… 3. Introduction to Agent Skills 👉 anthropic.skilljar.com/introduction-t… 4. Building with the Claude API 👉 anthropic.skilljar.com/claude-with-th… 5. Claude Code in Action 👉 anthropic.skilljar.com/claude-code-in… 6. Intro to Model Context Protocol (MCP) 👉 anthropic.skilljar.com/introduction-t… 7. MCP: Advanced Topics 👉 anthropic.skilljar.com/model-context-… 8. AI Fluency for Students 👉 anthropic.skilljar.com/ai-fluency-for… 9. AI Fluency for Educators 👉 anthropic.skilljar.com/ai-fluency-for… 10. Teaching AI Fluency 👉 anthropic.skilljar.com/teaching-ai-fl… 11. AI Fluency for Nonprofits 👉 anthropic.skilljar.com/ai-fluency-for… 12. Claude with Amazon Bedrock 👉 anthropic.skilljar.com/claude-in-amaz… 13. Claude with Google Cloud's Vertex AI 👉 anthropic.skilljar.com/claude-with-go… Follow me @ashiqur_ai for more AI IDEA.
English
@DuoEthan Nice. I really like the idea of building a graph over connections. Look forward to trying it out!
English
@intertwineai one of the biggest workflow killers is ai forgetting everything the second a project crosses session boundaries persistent memory is exactly the gap we are closing check it out at github.com/melandlabs/ope…
English
Yes — this is the right framing: agent systems get more valuable when memory and tools compound into reusable workflows.
The Hermes Observational Memory plugin is a practical version of that: Hermes keeps its own flow while sharing a local-first memory layer with other coding agents.
I wrote more about it here: x.com/intertwineai/s…
Mars_DeFi@Mars_DeFi
Hermes shows why the agent race is shifting from smarter chatbots to persistent AI operating systems. The real unlock is always on agentic infrastructure, where memory, tools, and specialized agents compound into reusable workflows. Here’s how Hermes is structured at the infrastructure level. — ● What is Hermes Agent? Hermes Agent is an open-source autonomous AI agent framework built by @NousResearch Research, designed to run persistent agents on user-controlled infrastructure. At a high level, its structure looks like this: • Company brain: stores vision, brand, customers, and products as the shared operating context. • Orchestrator Hermes Agent: reads the company brain and routes work to the right department. • Department brains: separate marketing, sales, ops, and support into focused operating layers. • Specialized agents: handle narrow tasks like writing, research, outreach, deployment, and triage. • Docker isolation: each agent runs in its own container, so context does not bleed across workflows. To understand why this matters, the infra stack needs to be broken into 6 layers. — ● Layer 1: Runtime layer Hermes runs on infrastructure users control, which moves agents away from rented SaaS sessions into persistent execution environments. • VPS • Local machine • Docker • SSH • Serverless • GPU workstation This matters because Hermes is not trapped inside a chat window; it becomes a long-running process that can operate 24/7. — ● Layer 2: Memory layer Normal agents depend on temporary context, while Hermes uses persistent memory to carry knowledge across sessions and workflows. • Persona • Preferences • Past work • Project context • User style • Reusable knowledge This is not model training; it is operational learning, where the agent remembers how you work and improves future execution. — ● Layer 3: Skill layer Hermes does not just complete tasks; it can extract reusable skills from completed workflows. Research briefs, @github issues, call summaries, and @discord monitoring can become repeatable procedures over time. Prompt -> task -> result -> skill -> better future task This is the closed learning loop where execution compounds into better future execution. — ● Layer 4: Orchestration layer This is where Hermes starts looking like an AI company, not a single general-purpose agent. Work can move from the company brain to an orchestrator, then into department brains and specialized worker agents. The key design choice is context routing, not context dumping, because each agent only receives the context needed for its task. — ● Layer 5: Isolation layer This is the underrated infra piece, because each Hermes agent can run inside its own Docker container. Marketing, sales, ops, and support can each keep separate context, memory, tools, and permissions. Instead of one giant messy agent brain, Hermes moves closer to microservices architecture for autonomous work. — ● Layer 6: Tool layer This is where Hermes stops being a talking interface and starts becoming an operator. Hermes can connect to: • GitHub for code/issues/PRs • @GoogleWorkspace for @gmail, @googledocs, Sheets, Calendar • @Reddit for market research • @firecrawl for clean web data • @obsdmd for second-brain context • @stripe for revenue intelligence • Discord for community automation • @firefliesai for meeting memory • Graphiti (built by @zep_ai) for knowledge graphs Models give intelligence, tools give agency, and infrastructure gives persistence. — The deeper shift is that AI is moving from prompt-response interfaces into persistent operating systems that can manage state and execution. @ChatGPTapp-style tools behave more like browsers, while Hermes-style infra behaves closer to a backend server for autonomous work. Hermes is not competing with chatbots rather it is competing with the idea that agents should live inside chat windows.
English
Folks should also look into DSPy and GEPA. It makes these same skills even more useful.
I was able to get a 25-point improvement on various tasks using a 1.2B model and GEPA.
Here's an explainer to show how to make it work.
x.com/intertwineai/s…
Shanto@ashiqur_ai
TOP 13 FREE AI COURSES TO TRY IN 2026: 1. Claude 101 👉 anthropic.skilljar.com/claude-101 2. AI Fluency: Frameworks & Foundations 👉 anthropic.skilljar.com/ai-fluency-fra… 3. Introduction to Agent Skills 👉 anthropic.skilljar.com/introduction-t… 4. Building with the Claude API 👉 anthropic.skilljar.com/claude-with-th… 5. Claude Code in Action 👉 anthropic.skilljar.com/claude-code-in… 6. Intro to Model Context Protocol (MCP) 👉 anthropic.skilljar.com/introduction-t… 7. MCP: Advanced Topics 👉 anthropic.skilljar.com/model-context-… 8. AI Fluency for Students 👉 anthropic.skilljar.com/ai-fluency-for… 9. AI Fluency for Educators 👉 anthropic.skilljar.com/ai-fluency-for… 10. Teaching AI Fluency 👉 anthropic.skilljar.com/teaching-ai-fl… 11. AI Fluency for Nonprofits 👉 anthropic.skilljar.com/ai-fluency-for… 12. Claude with Amazon Bedrock 👉 anthropic.skilljar.com/claude-in-amaz… 13. Claude with Google Cloud's Vertex AI 👉 anthropic.skilljar.com/claude-with-go… Follow me @ashiqur_ai for more AI IDEA.
English
@uday_devops No, definitely protect your envs. Raw .env files can leak secrets into AI context, logs, or shared state.
I'm using encrypted dotenvs: decrypt only at runtime, keep keys out of prompts, and know when you need a real vault. I wrote up an explainer here:
x.com/intertwineai/s…
Bryan Young@intertwineai
English
@mehd_io Yes, caching DB query results as markdown in Obsidian is a smart version of agent memory: fast, inspectable, and close to the work.
But where should that memory should live? I wrote about product memory, memory APIs, and local agent memory here:
x.com/intertwineai/s…
Bryan Young@intertwineai
English
Obsidian = great for agent memory. But what if it could also query your database data directly from Obsidian and cache the query and result there ?
That's what the DuckDB + MotherDuck Obsidian plugin does.
Results cache as markdown tables and agents read them instantly, no re-query.
00:00 - Why Obsidian is having a moment ("file over app" philosophy, just markdown)
00:45 - Obsidian + AI:
01:30 - The community plugin explosion (AI-generated PRs every 6 hours, new automated review dashboard)
02:30 - Introducing the DuckDB + MotherDuck Obsidian plugin
03:15 - Installing the plugin and configuring DuckDB file path + MotherDuck token
04:00 - Live demo : running queries
05:30 - Auto-refresh scheduling
06:30 - Agent workflow: refreshing notes via the Obsidian CLI
07:30 - Why this matters
08:30 - Runs on mobile !
English
Yes, 23k+ skills at this scale creates a rich playground for empirical agent work.
It gets even more useful when you load them as DSPy modules so the optimizer can search and compile the best chains. The dspy-agent-skills repo gives the exact install path and usage patterns to start:
github.com/intertwine/dsp…
English
👀 Generative Recursive Reasoning paper looks really interesting.
Here's the interactive FlashPapers explainer
flashpapers.intertwinesys.com/p/wwfxsmk6em/r…
Sungjin Ahn@SungjinAhn_
🧠We introduce "Generative Recursive Reasoning"! Recursive Reasoning Models like HRM, TRM, and Looped Transformers are deterministic — same input, same reasoning, every time. They collapse the entire space of plausible reasoning paths into a single attractor. Our model GRAM (Generative Recursive reAsoning Models) turns recursion itself into a stochastic latent trajectory. Multiple hypotheses, alternative solution strategies, and inference-time scaling not just by depth, but by width — parallel trajectory sampling. And here's the kicker: the same formulation that gives us conditional reasoning p(y|x) also makes GRAM a general generative model p(x). With only 10M params: • Sudoku-Extreme: 97.0% (TRM 87.4%) • ARC-AGI-1: 52.0% • ARC-AGI-2: 11.1% • N-Queens coverage: 90%+ 📄 Paper: arxiv.org/abs/2605.19376 🌐 Project page: ahn-ml.github.io/gram-website w/ Junyeob Baek @JunyeobB (KAIST), Mingyu Jo @pyross0000 (KAIST), Minsu Kim @minsuuukim (KAIST & Mila), Mengye Ren @mengyer (NYU), Yoshua Bengio @Yoshua_Bengio (Mila), Sungjin Ahn @SungjinAhn_ (KAIST)
English
And an hour after that - there's a share button, so that people can share their own paper explainers...
Here's the Mamba paper for example:
flashpapers.intertwinesys.com/p/gagxvyd6bw
elie@eliebakouch
@JeffDean where can we try this? is there a site where you just put a paper name and get this kind of model card? would love to test it properly 🙏
English
@intertwineai three hours from “where can we try this?” to a live explainer is a very clean feedback loop.
English
And... three hours later, Claude Code with the /goal command Ralph-Wiggum'd up a working version that anyone can sign in with a Google account and generate their own beautiful interactive AI paper explainers with Gemini Flash. Live now. 🚨🚀
flashpapers.intertwinesys.com
elie@eliebakouch
@JeffDean where can we try this? is there a site where you just put a paper name and get this kind of model card? would love to test it properly 🙏
English
English
@intertwineai @walls_jason1 Glad I could help blaze the trail! 🚀 That interactive Attention explainer you shipped is seriously impressive— the scaling demo and clean UI make the Transformer paper actually fun to dive into.
What’s the public repo link? Would love to check it out.
English