lmeyerov

I'll just leave this here:

This one is close to my heart: Will be at Berkeley sharing how Louie is speed running security competitions and where that changes how to think about investigation teams & programs. Berkeley is where I began my program synthesis journey, so cool to be back with @dawnsongtweets…

Funny UI shift I'm noticing as folks advance in day-to-day agents for Serious Work. Think 20,000-40,000 lines of code later for vibe coding, not just your first 2-4 weeks, or the same for doing hard data analysis. The IDE hollows out, and the loop is now between: 1. Agent…

One of my bigger lessons in program synthesis is that specifications are proofs only in the limit and under crazy universal functions. The torturous existence that is to be a designer is to forever be trying to bridge the gap between informal intent and precises proof that tools…

For the market shift to AI startups and AI products, one of my favorite lessons is "love the problem, not the solution". Big part of how we build Louie -- Solution me-too-ism: AI startup building agents and UIs for natural language interfaces to automating data Problem mindset:…

AI Automation Fails When Humans Forget the 'Automation' Part. --- Next time an anti-AI person claims, "My experiment proves AI agents can't do X!", it's worth asking: Was their AI loop even set up to achieve X? It sounds obvious, yet otherwise serious people keep overlooking…

"""That place is so crowded, no one goes there anymore""" Nvidia definitely has challengers... but also years of momentum

Useful mantra: An agent is worth 100 workflows 1. 10X Breadth: Python workflows hardcode a specific scenario like "Azure login alert". In contrast, one natural language prompt generalizes over many scenarios, like "investigate for triage most login alerts: ...". 2. 10X Depth: A…

Great AI UX read, a lot of what the Louie team has been doing and WHY Missing initial exchange for 2025 AI: The best UI is no UI. Agent just spends 20min driving you, full problem solved. Agents might get lost or need more instruction, and then we continue into the discussion…

LLM UI/code tooling feel like a transitionary period. Our LLM interface work is increasingly 80% for people and 20% for computers. Previous era chatgpt, langchain, n8n, tines, etc were designed for a dumber 50/50 era when agents could barely do 1 prompt turns, vs 50 2022-24:…

Repeat conversation I'm having with folks on a sign you're doing agents right: Your infra spend is going up every month Two reasons: - Your agents are running - Your team keeps adding more

Now that running agents for 2min-20min is actually useful, been surprises in where @louie_ai dev team time is going We guessed some of it, like that's why we built our AI notebooks with different foundations than normal jupyter/google notebooks, like louie could run several…

Coming back from DC/NYC, the biggest shock folks have is I no longer do agentic flows via code Natural language is surprisingly fine for getting the nuance of a self-managing plan that can run for 20min, same as if I was delegating to a junior colleague Whether I'm coding or…

Unexpected - my 'tool building' nowadays is worm-like prompts. Ex: """ - Upon start, copy this into 'ai_progress/task_<xyz>' and record timestamp + input prompt - Fill out a plan, and mark steps as TODO - Perform until all done, updating plan & statuses as you go """…

Catch us at the AI Agent World Tour – SF this May 28! Alexander morisse @silkspaceships Head of AI at Graphistry, will be in the mix – talking agentic workflows, LLM automation, and how we’re making AI actually do things. We're building a multi-agent workflow engine where…