Sabitlenmiş Tweet
Peter Pang
172 posts
Peter Pang
@intuitiveml
🚀 Co-founder, CreaoAI. Previously GenAI @ Meta (LLaMA). xApple. Join the journey ↓
Cupertino Katılım Ocak 2020
100 Takip Edilen6.1K Takipçiler
We rebuilt our entire company development and operation around AI. AI builds the tools, runs them, and improves them.
Now we're building a platform so any team can build the same self-improvement loop.
I did an interview talking through some of the details
• “AI-first” means redesigning the workflow, not adding AI to the old one.
• Engineers are moving from writing code to designing the systems that govern how agents write code.
• Self-healing software is closer than most people think.
• The real product is not the agent. It is the harness around the agent.
here's the link:
unite.ai/peter-pang-co-…
English
I kind of agree with this, hire super junior or super senior engineer
Brian Halligan@bhalligan
The talent shortage of 2030 will be caused by companies not hiring and training juniors in 2026. @ivanhzhao shifted to a barbell 🏋️♂️approach for hiring on the Engineering team at @NotionHQ: junior ICs paired with very senior architects. -The senior architects provide direction the model can't -The juniors are AI-pilled by default. -A senior managing 2–3 juniors managing 2–3 agents each compounds way better than a senior managing 4–6 agents directly. Links below👇 for the full episode
English
Peter Pang retweetledi
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
English
@yyyiiillluuu Claude Code remembering failures is useful. Learning reusable execution patterns from them is where compounding productivity starts.
Love the focus on local-first + open source too.
English
Claude Code can now self-improve with this plugin.
Introducing claude-smart — an open-source plugin that helps Claude Code learn from every session.
Memory helps Claude Code remember what happened.
claude-smart helps Claude Code improve what it does next.
Example:
Claude Code runs `npm test` without `--run`, and the command hangs in your repo.
Memory stores:
“npm test kept hanging.”
claude-smart learns:
“When running tests in this repo, use `npm test -- --run` because default watch mode hangs.”
claude-smart’s learnings are reusable and actionable, even across different projects.
It can also reduce unnecessary planning iterations and token use by 70%+ on similar future tasks.
Runs locally. 100% open source. No data is shared.
Install:
npx claude-smart install
With Codex:
npx claude-smart install --host codex
GitHub:
github.com/ReflexioAI/cla…
English
This is the missing layer for AI coding workflows: not just memory, but adaptation.
Yi Lu@yyyiiillluuu
Claude Code can now self-improve with this plugin. Introducing claude-smart — an open-source plugin that helps Claude Code learn from every session. Memory helps Claude Code remember what happened. claude-smart helps Claude Code improve what it does next. Example: Claude Code runs `npm test` without `--run`, and the command hangs in your repo. Memory stores: “npm test kept hanging.” claude-smart learns: “When running tests in this repo, use `npm test -- --run` because default watch mode hangs.” claude-smart’s learnings are reusable and actionable, even across different projects. It can also reduce unnecessary planning iterations and token use by 70%+ on similar future tasks. Runs locally. 100% open source. No data is shared. Install: npx claude-smart install With Codex: npx claude-smart install --host codex GitHub: github.com/ReflexioAI/cla…
English
small fix: had the colorbar legend reversed on the rubric chart, green now correctly maps to low poor%, red to high. same data, just the scale reading the right way around.
English
Who said agents are only for work?
@aritdeveloper built one that tracks football schedules across Top 5 leagues, UCL, UEL, Euro and World Cup.
Another great perk? It drops a Telegram match preview every morning before anyone's awake. No code. Fully automated.
Sports journalists and fan accounts: the part where someone manually checks seven league calendars every morning — that's the agent's job now.
ARIT@aritdeveloper
I built an AI agent that automatically tracks football schedules across the Top 5 leagues + UCL, UEL, Euro & World Cup And posts a daily match preview to Telegram every morning 🤖⚽ Built with @CreaoAI - zero code, fully automated 🧵👇
English
Yeah, you are right about human in the middle, we always have engineer to look through the ticket from triage system. Currently, only a small fraction of ticket is solved with no human in the loop. Majority of them still need engineer guidance. But it already saved majority of time to identify and triage the issue
English
I'm intrigued. Seems similar to the foundations of what Zouroboros is built on. I'm glad I'm on the right track.
Does yours allow a human in the middle to guide direction?
I kept a human in the middle because I wanted a self-healing and self-evolving AI system that grows with the user's interactions.
English
You’re absolutely right, and I should clarify further. Our grader does have full visibility into the entire tool-use history. However, we have a separate agent dedicated to optimizing that aspect, I’ll share more details on that in a separate post. This system, on the other hand, is primarily designed to detect system issues and bugs, which is where the healing process is focused.
English
@intuitiveml Evaluating final output alone is not correct. The path is important for long term accuracy. Eh in code if path is not correct agents create unwanted code and gradually degrade the agnet accuracy
English
A note on intent.
We care a lot about accuracy and fairness, but we’re not building a leaderboard or ranking models against each other. The Grader exists to surface issues in our agent system: bad prompts, broken tool contracts, drifted integrations, infra flakes, regressions from our own deployments. Per-model scores are just a debugging signal, not a benchmark. If two judges score a response "poor" on the same messageId, we don’t learn that one model is better than another. We learn that something in our pipeline produced a bad answer, and we need to fix it.
English
@ethankongee Agreed. Software engineering will not disappear, but it will continue to evolve, as will the role of the software engineer
English
The future of software engineering lies in self-evolving agentic systems: software that can self-maintain, self-improve, self-secure, and self-heal.
So we need to redefine “harness engineering.”
It’s not the discipline of manually tweaking the harness. It’s the discipline of building agentic systems that can iterate on the harness themselves.
And that might be where software engineering gets interesting again.
English
The chart below shows the sampled evaluations and average scores collected on the CREAO platform over the past 7 days.
English
Tried ChatGPT Image 2.0 to turn my article into a single image, impressed by how intelligent it is and how accurately it captured the text.
Peter Pang@intuitiveml
English
@fchollet this is the core thesis behind why agent benchmarks keep misleading us, we design them from first principles instead of iterating from real-world failure modes. The field needs more empirical humility and fewer thought experiments.
English
You cannot think your way to a perfect design. Only building and testing, over many iterations, can reveal the flaws in your mental model and provide the feedback you need to create the best design possible.
English
@fchollet This principle is underused in agent design. The best-performing agents we've studied aren't the most capable, they're the most constrained to a well-scoped task. Unbounded agency is a research problem, not a product feature
English
Constraints are the catalyst of invention. An infinite search space leads to paralysis. The most creative inventions happen when you are forced to solve a problem within appropriately narrow constraints.
English
