Amir Elaguizy
6.5K posts
Amir Elaguizy
@amirpc
Builder, Dad - YC S13 - Cofounder @PlayPokerSkill, @Cratejoy, MarketZero - Question the answers
Austin, TX Katılım Aralık 2008
783 Takip Edilen1.3K Takipçiler
Watching anthropic try to figure out how to thread the needle of openclaw like power while still being the whole world's nanny outta be interesting. A case study in product held back by self imposed constraints. They seem to figure it out though.
English
This seems like a formal announcement of what they clearly were already starting to do, the focus on codex + apps + excel integrations, etc. Seems obvious that they cannot surrender enterprise to Anthropic.
Andrew Curran@AndrewCurran_
The WSJ is reporting that OpenAI is about to take a hard turn into enterprise.
English
Training models with the tools in them as compiled deterministic subcomponents seems like a massive step forward in agent capabilities.
Its the difference between someone being really great at using controls to manipulate a grasping device to HAVING HANDS.
Insane!
Christos Tzamos@ChristosTzamos
1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy
English
Autoresearch start at 4 -> 280 by EOD. It's not just optimizing, it's building too. This is more than "research" it's a development paradigm.
English
Paperclip + Openclaw + HZL task + QMD memory + 5.4 xhigh (+ ACP + Codex CLI) is the full company stack that actually works today.
It's real, its replacement quality.
English
@Authy your change phone number process is broken in the app and you denied my manual migration request. Is the solution just to stop using Authy? The approve notification doesn't do anything when I get the push. It's literally broken.
English
Amir Elaguizy retweetledi
1/ Today, I'm excited to announce ✨Mixlayer✨.
Two years ago, I reoriented my life around a singular mission: to build the best inference platform for open source AI.
English
Open sourcing a few skills that are helpful in creating prompts, skills and agents with true commander's intent. E.g. these produce autonomy, executive function and judgement as first class citizens - not narrow heuristics.
github.com/aelaguiz/autho…
English
@karpathy First screenshot is turn 10 or so, second screenshot I think is turn 13. Look at the win rate and CI
English
Holy shit Karpathy's auto-optimization method applied to our domain just cracked a problem I've been stuck on for a couple of months. It got our preflop-RTS to beat baseline by 200 BB/100, previously I could only get RTS to work postflop. Thanks @karpathy !
English
Primitives are emerging. Openclaw agents. Wrap and extend models. The AI economy is starting to show the smallest bit of shape.
English
People are starting to really innovate on the composability of LLM calls sharing the OpenAI REST API format. Very interesting stuff happening. Wrapping and enhancing is going to be a thing.
English
If your SaaS goes down you simply accelerate replacement with a codex-built alternative. Just FYI for those that can't seem to keep their software up @clerk
English
This is the best argument for non frontier and self hosting I've seen. Automated continuous RL of our own agents sounds really compelling.
Bo Wang@BoWang87
This is probably the first RL work on OpenClaw 🔥 MetaClaw: Just talk to your agent and let it evolve automatically. Github: github.com/aiming-lab/Met… Most AI agents are frozen the moment they ship. Every mistake they make, they'll make again tomorrow. MetaClaw fixes that. It's an online RL layer built on top of OpenClaw that lets agents learn from their own interactions — no GPU cluster, no offline dataset, no engineering team required. The loop is simple: every conversation is logged as a training trajectory. When the agent fails, it analyzes what went wrong and proposes a new reusable skill. LoRA updates train asynchronously in the background. The next time a similar situation comes up, the relevant skill gets retrieved into the prompt automatically. The agent doesn't just accumulate conversations. It accumulates capability. What makes this different from fine-tuning: there's no human labeling pipeline, no batch training runs, no deployment cycle. The improvement happens continuously, invisibly, in production. Interaction → learning → improvement, on a loop. No offline dataset. No coding required. No GPU cluster. The part worth paying attention to: this turns every user interaction into a training signal. The agent you deploy on day one is not the agent you have on day thirty. It's been shaped by everything it got wrong and fixed. Great work by @HuaxiuYaoML !
English
Amir Elaguizy retweetledi
🦞 OpenClaw 3.7 JUST DROPPED and it's HUGE.
✅ Context Engine Plugin API - Say goodbye to context loss. Plugins like lossless-claw can now preserve full memory without touching core logic.
✅ 30% Faster with LongCat - New official integration solves the Google/Anthropic ban problem. Faster tasks, stable API, zero account risk.
✅ Threads Survive Restarts - Discord/Telegram bindings now persist. Your conversations stay connected even after reboot.
✅ Docker Slim + i18n - Multi-stage builds, Spanish locale support, iOS App Store prep. OpenClaw is going mainstream.
Your AI agent just got a serious upgrade. The AI OS war has begun.
English